Daily Papers

Despite a century of empirical memory research, existing AI agent memory systems rely on system-engineering metaphors (virtual-memory paging, flat LLM storage, Zettelkasten notes), none integrating principles of consolidation, forgetting, and reconsolidation. We present ZenBrain, a multi-layer memory architecture integrating fifteen neuroscience models. It implements seven memory layers (working, short-term, episodic, semantic, procedural, core, cross-context) orchestrated by nine foundational algorithms (Two-Factor Synaptic Model, vmPFC-coupled FSRS, Simulation-Selection sleep, Bayesian confidence, and five more) plus six new Predictive Memory Architecture (PMA) components: a four-channel NeuromodulatorEngine, prediction-error-gated ReconsolidationEngine, TripleCopyMemory with divergent decay, four-dimensional PriorityMap with amygdala fast-path, StabilityProtector (NogoA/HDAC3 analogue), and MetacognitiveMonitor for bias detection. The 15-algorithm ablation reveals a cooperative survival network: under stress, 9 of 15 algorithms become individually critical (delta-Q up to -93.7%, Wilcoxon, 10 seeds, alpha=0.005). Simulation-Selection sleep achieves 37% stability improvement (p<0.005) with 47.4% storage reduction. TripleCopyMemory retains S(t)=0.912 at 30 days; PriorityMap reaches NDCG@10=0.997. Multi-layer routing beats a flat single-layer baseline by 20.7% F1 on LoCoMo (p<0.005) and 19.5% on MemoryArena (p=0.015). On LongMemEval-500, ZenBrain holds the highest mean rank on all 12 system-judge cells (4 systems x 3 LLM judges), three-judge mean J=0.545 vs letta=0.485, a-mem=0.414, mem0=0.394; all 9 pair-wise contrasts clear Bonferroni (alpha=0.05/18, min p=6.2e-31, d in [0.18, 0.52]). Under LongMemEval's binary judge, ZenBrain reaches 91.3% of oracle accuracy at 1/106th the per-query token budget. Open-source with 11,589 automated test cases.

Large Language Models (LLMs) rely on the Key-Value (KV) Cache to store token history, enabling efficient decoding of tokens. As the KV-Cache grows, it becomes a major memory and computation bottleneck, however, there is an opportunity to alleviate this bottleneck, especially because prior research has shown that only a small subset of tokens contribute meaningfully to each decoding step. A key challenge in finding these critical tokens is that they are dynamic, and heavily input query-dependent. Existing methods either risk quality by evicting tokens permanently, or retain the full KV-Cache but rely on retrieving chunks (pages) of tokens at generation, failing at dense, context-rich tasks. Additionally, many existing KV-Cache sparsity methods rely on inaccurate proxies for token importance. To address these limitations, we introduce TokenButler, a high-granularity, query-aware predictor that learns to identify these critical tokens. By training a light-weight predictor with less than 1.2% parameter overhead, TokenButler prioritizes tokens based on their contextual, predicted importance. This improves perplexity & downstream accuracy by over 8% relative to SoTA methods for estimating token importance. We evaluate TokenButler on a novel synthetic small-context co-referential retrieval task, demonstrating near-oracle accuracy. Code, models and benchmarks: https://github.com/abdelfattah-lab/TokenButler

The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of "dominant" FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. %making them a powerful and efficient proxy for token importance. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. % Since accessing only a small fraction of the KV cache, FASA drastically lowers memory bandwidth requirements and computational cost. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100\% of full-KV performance when only keeping 256 tokens, and achieves 2.56times speedup using just 18.9\% of the cache on AIME24.

Long-term conversational memory in practical LLM applications is inherently collaborative: information is produced by multiple participants, scattered across groups and channels, revised over time, and implicitly grounded in roles and social context. Yet there is currently no established benchmark that evaluates memory under interaction patterns resembling real-world deployment, as existing benchmarks largely focus on dyadic or single-topic dialogues. In this paper, we introduce EverMemBench, the first benchmark designed for long-horizon collaborative memory, built from multi-party, multi-group conversations spanning over one million tokens with dense cross-topic interleaving, temporally evolving decisions, and role-conditioned personas. EverMemBench evaluates memory systems using 2400 QA pairs across three dimensions essential for real applications: fine-grained recall, memory awareness, and user profile understanding. Our evaluation reveals fundamental limitations of current systems: multi-hop reasoning collapses under multi-party attribution even with oracle evidence (26% accuracy), temporal reasoning fails without explicit version semantics beyond timestamps, and memory awareness is bottlenecked by retrieval, as similarity-based methods miss implicitly relevant information. EverMemBench thus represents a concrete step toward realistic evaluation of LLM memory and a cornerstone benchmark for developing next-generation LLMs that reason over time, roles, and collaborative interaction structure. Our benchmark and code are publicly available at https://github.com/EverMind-AI/EverMemBench.

As events progress, news articles often update with new information: if we are not cautious, we risk propagating outdated facts. In this work, we hypothesize that linguistic features indicate factual fluidity, and that we can predict which facts in a news article will update using solely the text of a news article (i.e. not external resources like search engines). We test this hypothesis, first, by isolating fact-updates in large news revisions corpora. News articles may update for many reasons (e.g. factual, stylistic, narrative). We introduce the NewsEdits 2.0 taxonomy, an edit-intentions schema that separates fact updates from stylistic and narrative updates in news writing. We annotate over 9,200 pairs of sentence revisions and train high-scoring ensemble models to apply this schema. Then, taking a large dataset of silver-labeled pairs, we show that we can predict when facts will update in older article drafts with high precision. Finally, to demonstrate the usefulness of these findings, we construct a language model question asking (LLM-QA) abstention task. We wish the LLM to abstain from answering questions when information is likely to become outdated. Using our predictions, we show, LLM absention reaches near oracle levels of accuracy.

Large language models (LLMs) increasingly decide whether software behaves correctly, either by writing a test oracle or by acting as one. Yet two oracles can look identical and rest on different ground: one assertion encodes a written specification, another only what the model learned in training. Prior secondary studies sort oracles by form or by technique, rarely by the property that governs how far a verdict can be trusted: where its authority comes from. This systematic literature review, reported under the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, screens 2,436 records to 54 included studies, extended by citation searching (snowballing) to 83 in total. We read the corpus along three axes: the source of an oracle's authority, the form it takes, and the mechanism that adjudicates it. Just over half of the corpus reaches a verdict with no specification at all. That is what lets these oracles work on code with no specification to consult, and what leaves a challenged verdict with less to fall back on. Source and mechanism cross-cut rather than coincide, so a label such as LLM-as-a-judge names how a verdict is produced, not why it should be trusted. Oracle quality is most often judged by resemblance to a known oracle rather than by whether injected faults are caught. The first question to ask of any LLM oracle is therefore what one would point to in defending its verdict. The protocol, search query, and per-study coding sheet are released.

Benchmarks for LLM-generated GPU kernels (KernelBench, TritonBench, GEAK) score correctness through fixed-shape, small-sample allclose-style checks. The number of inputs varies between benchmarks. The shape, dtype, and tolerance are fixed for each kernel. We test that oracle empirically. We construct a controlled corpus of 24 Triton and CPU stand-in kernels (15 correct controls and 9 LLM-style buggy variants seeded with documented transcription errors) and re-evaluate it under op-schema-aware seeded fuzzing with a high-precision (fp64) CPU reference and per-(op, dtype) absolute tolerances. The seeded oracle flags 9 of 9 buggy kernels and passes 15 of 15 correct controls, at zero precision cost on controls. We extend the corpus to 26 ops (adding a flash-attention pair) and re-run the same protocol on five GPU classes (RTX 3060, A10, L40S, A100 SXM4, H100 NVL). The verdicts are identical across all five GPUs: 10 of 10 illusions caught and 16 of 16 controls clean. The corpus result is about LLM-style transcription bugs that the allclose-on-one-shape oracle certifies as correct, not about the bug rate of any specific deployed LLM. Every flagged failure replays byte-for-byte from a stored seed.

Execution Accuracy (EX), the widely used metric for evaluating the effectiveness of Natural Language to SQL (NL2SQL) solutions, is becoming increasingly unreliable. It is sensitive to syntactic variation, ignores that questions may admit multiple interpretations, and is easily misled by erroneous ground-truth SQL. To address this, we introduce ROSE, an intent-centered metric that focuses on whether the predicted SQL answers the question, rather than consistency with the ground-truth SQL under the reference-dependent paradigm. ROSE employs an adversarial Prover-Refuter cascade: SQL Prover assesses the semantic correctness of a predicted SQL against the user's intent independently, while Adversarial Refuter uses the ground-truth SQL as evidence to challenge and refine this judgment. On our expert-aligned validation set ROSE-VEC, ROSE achieves the best agreement with human experts, outperforming the next-best metric by nearly 24% in Cohen's Kappa. We also conduct a largescale re-evaluation of 19 NL2SQL methods, revealing four valuable insights. We release ROSE and ROSE-VEC to facilitate more reliable NL2SQL research.

Activation oracles aim to make the activations of other models legible to humans and yield promising results compared to white-box interpretability techniques. However, uncertainty quantification (UQ) for the natural-language outputs of such activation oracles is so far understudied. Here, we investigate 6 different methods for estimating the confidence of activation oracles and evaluate how well-calibrated their confidence scores are. Our experiments on 6,000 samples per oracle (varying verbalizer and context prompts) reveal that bootstrap mode frequency is the best-calibrated method among those tested (ECE 5.7% vs. 25.5% for the answer-word log-probability on Qwen3-8B; 10.3% vs. 13.1% on Qwen3.6-27B), and that the log-prob baseline can serve as a fast triage signal at a fraction of the cost. Code and the patched trainer are available at https://github.com/federicotorrielli/probabilistic_activation_oracles.

Large language models (LLMs) can be used to generate text data for training and evaluating other models. However, creating high-quality datasets with LLMs can be challenging. In this work, we explore human-AI partnerships to facilitate high diversity and accuracy in LLM-based text data generation. We first examine two approaches to diversify text generation: 1) logit suppression, which minimizes the generation of languages that have already been frequently generated, and 2) temperature sampling, which flattens the token sampling probability. We found that diversification approaches can increase data diversity but often at the cost of data accuracy (i.e., text and labels being appropriate for the target domain). To address this issue, we examined two human interventions, 1) label replacement (LR), correcting misaligned labels, and 2) out-of-scope filtering (OOSF), removing instances that are out of the user's domain of interest or to which no considered label applies. With oracle studies, we found that LR increases the absolute accuracy of models trained with diversified datasets by 14.4%. Moreover, we found that some models trained with data generated with LR interventions outperformed LLM-based few-shot classification. In contrast, OOSF was not effective in increasing model accuracy, implying the need for future work in human-in-the-loop text data generation.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H_2). Through a comprehensive investigation, we find that (i) the emergence of H_2 is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H_2O), a KV cache eviction policy that dynamically retains a balance of recent and H_2 tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H_2O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29times, 29times, and 3times on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9times. The code is available at https://github.com/FMInference/H2O.

Agent memory is a systems problem for long-horizon agents. Practical deployments require retention of task state across extended conversations, recovery of user-specific facts and preferences across sessions, and accumulation of procedural knowledge from prior outcomes. These requirements extend beyond document retrieval: a memory layer must determine which interactions become durable state, how that state is scoped, how it is retrieved under latency constraints, and how it is revised or removed over time. This report studies Oracle Agent Memory as a database-native memory substrate built on Oracle Database. Three themes organize the discussion: memory as a lifecycle spanning ingestion, extraction, consolidation, retrieval, summarization, and revision or removal; a layered architecture that separates an active memory core from a passive memory-store interface with explicit scope control across users, agents, and threads; and evaluation methodology in which downstream task accuracy is complemented by memory-centric measures such as evidence retrieval, recall, latency, and estimated token use. The report summarizes LongMemEval results, reaching 93.8% accuracy, compares Oracle Agent Memory against flat-history baselines, using about 10.7x fewer tokens, and published or reported external baselines where available, and closes with implementation-oriented appendix material covering setup, thread lifecycle, and search semantics.

Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.

Many existing evaluation benchmarks for Large Language Models (LLMs) quickly become outdated due to the emergence of new models and training data. These benchmarks also fall short in assessing how LLM performance changes over time, as they consist of static questions without a temporal dimension. To address these limitations, we propose using future event prediction as a continuous evaluation method to assess LLMs' temporal generalization and forecasting abilities. Our benchmark, Daily Oracle, automatically generates question-answer (QA) pairs from daily news, challenging LLMs to predict "future" event outcomes. Our findings reveal that as pre-training data becomes outdated, LLM performance degrades over time. While Retrieval Augmented Generation (RAG) has the potential to enhance prediction accuracy, the performance degradation pattern persists, highlighting the need for continuous model updates.

LLM routing has become a popular approach to improve the cost-quality trade-off of LLM services by dynamically selecting a model for each query. Recent work has explored a broad range of routing methods, including clustering-based routers, learned classifiers, pairwise ranking, and confidence-based approaches. Our extensive study of 21 routing methods across five benchmarks reveals a consistent phenomenon that we call the routing plateau: many methods, including kNN, achieve very similar accuracy and converge to a narrow performance range that remains far below the oracle router. Our investigation shows that the plateau is largely caused by a predictability bottleneck: current routers mainly learn global averaged model-performance trends rather than fine-grained query-specific routing signals. As a result, they solve overlapping easy queries but collectively fail on hard queries that require instance-specific routing decisions. We further study how to move beyond the plateau and find that larger training datasets, stronger encoders, and end-to-end fine-tuning can further improve routing accuracy. These findings characterize the common limits of current routing methods and provide insights and actionable directions for the community to build more effective routing systems.

Language models exhibit remarkable natural language generation capabilities but remain prone to hallucinations, generating factually incorrect information despite producing syntactically coherent responses. This study introduces the Licensing Oracle, an architectural solution designed to stem hallucinations in LMs by enforcing truth constraints through formal validation against structured knowledge graphs. Unlike statistical approaches that rely on data scaling or fine-tuning, the Licensing Oracle embeds a deterministic validation step into the model's generative process, ensuring that only factually accurate claims are made. We evaluated the effectiveness of the Licensing Oracle through experiments comparing it with several state-of-the-art methods, including baseline language model generation, fine-tuning for factual recall, fine-tuning for abstention behavior, and retrieval-augmented generation (RAG). Our results demonstrate that although RAG and fine-tuning improve performance, they fail to eliminate hallucinations. In contrast, the Licensing Oracle achieved perfect abstention precision (AP = 1.0) and zero false answers (FAR-NE = 0.0), ensuring that only valid claims were generated with 89.1% accuracy in factual responses. This work shows that architectural innovations, such as the Licensing Oracle, offer a necessary and sufficient solution for hallucinations in domains with structured knowledge representations, offering guarantees that statistical methods cannot match. Although the Licensing Oracle is specifically designed to address hallucinations in fact-based domains, its framework lays the groundwork for truth-constrained generation in future AI systems, providing a new path toward reliable, epistemically grounded models.

Text-to-SQL technology has become crucial for translating natural language into SQL queries in various industries, enabling non-technical users to perform complex data operations. The need for accurate evaluation methods has increased as these systems have grown more sophisticated. However, we found that the Execution Accuracy (EX), the most promising evaluation metric, still shows a substantial portion of false positives and negatives compared to human evaluation. Thus, this paper introduces FLEX (False-Less EXecution), a novel approach to evaluating text-to-SQL systems using large language models (LLMs) to emulate human expert-level evaluation of SQL queries. Our method shows significantly higher agreement with human expert judgments, improving Cohen's kappa from 61 to 78.17. Re-evaluating top-performing models on the Spider and BIRD benchmarks using FLEX reveals substantial shifts in performance rankings, with an average performance decrease of 3.15 due to false positive corrections and an increase of 6.07 from addressing false negatives. This work contributes to a more accurate and nuanced evaluation of text-to-SQL systems, potentially reshaping our understanding of state-of-the-art performance in this field.

Text-to-SQL task maps natural language utterances to structured queries that can be issued to a database. State-of-the-art (SOTA) systems rely on finetuning large, pre-trained language models in conjunction with constrained decoding applying a SQL parser. On the well established Spider dataset, we begin with Oracle studies: specifically, choosing an Oracle hypothesis from a SOTA model's 10-best list, yields a 7.7% absolute improvement in both exact match (EM) and execution (EX) accuracy, showing significant potential improvements with reranking. Identifying coherence and correctness as reranking approaches, we design a model generating a query plan and propose a heuristic schema linking algorithm. Combining both approaches, with T5-Large, we obtain a consistent 1% improvement in EM accuracy, and a ~2.5% improvement in EX, establishing a new SOTA for this task. Our comprehensive error studies on DEV data show the underlying difficulty in making progress on this task.

Benchmarks are widely used to evaluate task completion by Large Language Models (LLMs), but this approach has accumulated construction-validity problems, and a passing score may not show whether the requested task was delivered. We study both problems. In a controlled code-as-spec setup, two production Copilot CLI agents (claude-opus-4.7, gpt-5.5) re-implement a React Fluent-UI data table in Angular as a reusable library under a hidden 222-test Playwright oracle across 18 runs and three oracle-availability conditions. Alongside the score, we run a mechanical library audit and check each verdict with a no-op ablation. Without the oracle, the library is present but unfinished, revealed by scores. With the oracle in the loop, the score reaches near-perfect, but from a demo holding the tested behavior directly, the library left dead or absent. We call this building to the test; the broader disposition behind both we call validation self-awareness. The agent does not, on its own, validate what it ships as a user would. Prevalence remains an open question across other agents, signals, and model families. Beyond benchmark scores, dispositions like validation self-awareness merit research attention.

There is increasing evidence that question-answering (QA) systems with Large Language Models (LLMs), which employ a knowledge graph/semantic representation of an enterprise SQL database (i.e. Text-to-SPARQL), achieve higher accuracy compared to systems that answer questions directly on SQL databases (i.e. Text-to-SQL). Our previous benchmark research showed that by using a knowledge graph, the accuracy improved from 16% to 54%. The question remains: how can we further improve the accuracy and reduce the error rate? Building on the observations of our previous research where the inaccurate LLM-generated SPARQL queries followed incorrect paths, we present an approach that consists of 1) Ontology-based Query Check (OBQC): detects errors by leveraging the ontology of the knowledge graph to check if the LLM-generated SPARQL query matches the semantic of ontology and 2) LLM Repair: use the error explanations with an LLM to repair the SPARQL query. Using the chat with the data benchmark, our primary finding is that our approach increases the overall accuracy to 72% including an additional 8% of "I don't know" unknown results. Thus, the overall error rate is 20%. These results provide further evidence that investing knowledge graphs, namely the ontology, provides higher accuracy for LLM powered question answering systems.

We propose Encyclopedic-VQA, a large scale visual question answering (VQA) dataset featuring visual questions about detailed properties of fine-grained categories and instances. It contains 221k unique question+answer pairs each matched with (up to) 5 images, resulting in a total of 1M VQA samples. Moreover, our dataset comes with a controlled knowledge base derived from Wikipedia, marking the evidence to support each answer. Empirically, we show that our dataset poses a hard challenge for large vision+language models as they perform poorly on our dataset: PaLI [14] is state-of-the-art on OK-VQA [37], yet it only achieves 13.0% accuracy on our dataset. Moreover, we experimentally show that progress on answering our encyclopedic questions can be achieved by augmenting large models with a mechanism that retrieves relevant information from the knowledge base. An oracle experiment with perfect retrieval achieves 87.0% accuracy on the single-hop portion of our dataset, and an automatic retrieval-augmented prototype yields 48.8%. We believe that our dataset enables future research on retrieval-augmented vision+language models. It is available at https://github.com/google-research/google-research/tree/master/encyclopedic_vqa .

Large-scale semantic parsing datasets annotated with logical forms have enabled major advances in supervised approaches. But can richer supervision help even more? To explore the utility of fine-grained, lexical-level supervision, we introduce Squall, a dataset that enriches 11,276 WikiTableQuestions English-language questions with manually created SQL equivalents plus alignments between SQL and question fragments. Our annotation enables new training possibilities for encoder-decoder models, including approaches from machine translation previously precluded by the absence of alignments. We propose and test two methods: (1) supervised attention; (2) adopting an auxiliary objective of disambiguating references in the input queries to table columns. In 5-fold cross validation, these strategies improve over strong baselines by 4.4% execution accuracy. Oracle experiments suggest that annotated alignments can support further accuracy gains of up to 23.9%.

Text-to-SQL systems allow non-SQL experts to interact with relational databases using natural language. However, their tendency to generate executable SQL for ambiguous, out-of-scope, or unanswerable queries introduces a hidden risk, as outputs may be misinterpreted as correct. This risk is especially serious in biomedical contexts, where precision is critical. We therefore present Query Carefully, a pipeline that integrates LLM-based SQL generation with explicit detection and handling of unanswerable inputs. Building on the OncoMX component of ScienceBenchmark, we construct OncoMX-NAQ (No-Answer Questions), a set of 80 no-answer questions spanning 8 categories (non-SQL, out-of-schema/domain, and multiple ambiguity types). Our approach employs llama3.3:70b with schema-aware prompts, explicit No-Answer Rules (NAR), and few-shot examples drawn from both answerable and unanswerable questions. We evaluate SQL exact match, result accuracy, and unanswerable-detection accuracy. On the OncoMX dev split, few-shot prompting with answerable examples increases result accuracy, and adding unanswerable examples does not degrade performance. On OncoMX-NAQ, balanced prompting achieves the highest unanswerable-detection accuracy (0.8), with near-perfect results for structurally defined categories (non-SQL, missing columns, out-of-domain) but persistent challenges for missing-value queries (0.5) and column ambiguity (0.3). A lightweight user interface surfaces interim SQL, execution results, and abstentions, supporting transparent and reliable text-to-SQL in biomedical applications.

Activation Oracles (AOs) are promising methods for interpreting residual stream activations. However, current AOs face important issues, such as hallucinations and vagueness. Additionally, text-inversion confounds make them hard to evaluate. To this end, we improve the Activation Oracle (AO) training regime in four ways: training on on-policy rollouts, improving the conversational dataset, feeding more layers and an improvement to the injection formula. The capability improvements are marginal, but quality of life improvements are quite substantial. In addition, we open source the first comprehensive evaluation suite for AO quality, which we call AObench. Overall, we hope that our work sets a foundation that helps improve AOs and other models in the paradigm of scalable, end-to-end interpretability.

Cross-view Geo-localisation is typically performed at a coarse granularity, because densely sampled satellite image patches overlap heavily. This heavy overlap would make disambiguating patches very challenging. However, by opting for sparsely sampled patches, prior work has placed an artificial upper bound on the localisation accuracy that is possible. Even a perfect oracle system cannot achieve accuracy greater than the average separation of the tiles. To solve this limitation, we propose combining cross-view geo-localisation and relative pose estimation to increase precision to a level practical for real-world application. We develop PEnG, a 2-stage system which first predicts the most likely edges from a city-scale graph representation upon which a query image lies. It then performs relative pose estimation within these edges to determine a precise position. PEnG presents the first technique to utilise both viewpoints available within cross-view geo-localisation datasets to enhance precision to a sub-metre level, with some examples achieving centimetre level accuracy. Our proposed ensemble achieves state-of-the-art precision - with relative Top-5m retrieval improvements on previous works of 213%. Decreasing the median euclidean distance error by 96.90% from the previous best of 734m down to 22.77m, when evaluating with 90 degree horizontal FOV images. Code will be made available: tavisshore.co.uk/PEnG

Current mathematical reasoning benchmarks for large language models (LLMs) are approaching saturation, with some achieving > 90% accuracy, and are increasingly compromised by training-set contamination. We introduce Putnam-AXIOM, a benchmark of 522 university-level competition problems drawn from the prestigious William Lowell Putnam Mathematical Competition, and Putnam-AXIOM Variation, an unseen companion set of 100 functional variants generated by programmatically perturbing variables and constants. The variation protocol produces an unlimited stream of equally difficult, unseen instances -- yielding a contamination-resilient test bed. On the Original set, OpenAI's o1-preview -- the strongest evaluated model -- scores 41.9%, but its accuracy drops by 19.6% (46.8% relative decrease) on the paired Variations. The remaining eighteen models show the same downward trend, ten of them with non-overlapping 95% confidence intervals. These gaps suggest memorization and highlight the necessity of dynamic benchmarks. We complement "boxed" accuracy with Teacher-Forced Accuracy (TFA), a lightweight metric that directly scores reasoning traces and automates natural language proof evaluations. Putnam-AXIOM therefore provides a rigorous, contamination-resilient evaluation framework for assessing advanced mathematical reasoning of LLMs. Data and evaluation code are publicly available at https://github.com/brando90/putnam-axiom.

Evolving documents, such as laws, tax codes, and software documentation, are amended, replaced, and sometimes reverted over time, so a question has different correct answers at different dates. In contrast to encyclopedic knowledge, where an old fact is simply overwritten, an amendment is itself an official text that states what it replaces and when it takes effect, and the earlier version stays correct for its validity period. The central challenge is therefore version resolution, that is, identifying the version in force on the queried date. Existing temporal QA datasets treat time only as an annotation, so version resolution stays untested. We present TIDE, an expert-verified benchmark of 3,050 QA pairs over 644 official customs instruments issued between 1969 and 2025 by the Government of Bangladesh, covering eight task types over deeply code-mixed documents that are heterogeneous in layout and dated in two calendars. In addition, we evaluate nine recent LLMs under a single protocol across parametric, gold-context, and retrieval access, scored by a three-judge LLM council with a hard date gate separating correct meaning from correct time. The best macro-averaged accuracy is only 68.5%. Resolving a version from an implicit date reaches 59.7%, and detecting that the supplied version does not govern the query reaches only 26.7%. Models are more likely to find correct versions than to reject incorrect ones, and they tend to follow a confident parametric answer over the supplied authoritative text. All code and data are available at https://github.com/icsetepa44/TIDE

Large Reasoning Models (LRMs) allocate substantial inference-time compute to Chain-of-Thought (CoT) reasoning, improving performance on mathematics, scientific QA, and tool usage. However, this introduces overthinking: LRMs often reach a correct intermediate solution, continue reasoning, and overwrite it with an incorrect answer. We first demonstrate that oracle stopping--where we inject </think> at every sentence boundary and select the best stopping point in hindsight--improves average accuracy by 8% while reducing thinking tokens by 72%, exposing substantial overthinking. Motivated by this finding, we propose ThinkBrake, which monitors the log-probability margin between the top continuation token and </think> at sentence boundaries, stopping reasoning when this margin narrows. ThinkBrake requires no training and achieves favorable accuracy-efficiency trade-offs across math, scientific QA, and tool usage benchmarks, reducing thinking token usage by up to 30%. Furthermore, we provide theoretical analysis showing that ThinkBrake is equivalent to test-time realignment with a reward bonus for the </think> token.

Retrieval-augmented generation stores all content indiscriminately, degrading accuracy as noise accumulates. Parametric approaches compress knowledge into weights, precluding selective updates. Neither mirrors biological memory, which gates encoding based on salience and archives rather than deletes superseded information. We introduce write-time gating that filters incoming knowledge objects using composite salience scores (source reputation, novelty, reliability) while maintaining version chains that preserve prior states. Using real LLM evaluation without oracle access to quality labels, write gating achieves 100 percent accuracy versus 13 percent for ungated stores. The critical finding emerges under distractor scaling: at 8:1 distractor ratios, read-time filtering (Self-RAG) collapses to 0 percent while write gating maintains 100 percent, revealing a structural advantage of write-time over read-time curation. Validation on Wikipedia (20 entities), procedurally generated pharmacology data, and 2026 arXiv papers confirms these findings. The gating advantage scales inversely with parametric memory support: +25pp for Wikipedia, +48pp for post-cutoff arXiv, +65pp for procedural data with zero training knowledge. Signal ablation confirms the method does not depend on oracle-correlated metadata. Write gating matches Self-RAG accuracy at one-ninth the query-time cost.

Computer-use agents ground natural-language instructions in screenshots to locate interface elements, yet existing benchmarks do not isolate whether models bind relational language to the correct element. We introduce GUI-Primitives, a 994-item benchmark of contrastive instruction pairs over seven spatial relations in graphical user interfaces (left/right, above/below, containment, alignment, proximity, list ordinal, occlusion). Each pair holds the screenshot and anchor fixed while changing the relation expression, so the correct target moves between two designated candidates. Five annotators validate a 196-item subset (κ= 0.94 well-formedness; κ= 0.79 target selection). Nineteen vision-language models reach at most 32% strict point-in-box accuracy. Because models emit unconstrained coordinates, we classify each prediction by the candidate region it falls within. Predictions fall outside both candidates on 60-92% of items. Conditional on falling within a candidate region, target selection reaches 0.82-0.90 for horizontal position, vertical position, proximity, and list ordinal, but does not differ significantly from 0.50 for containment and occlusion: most failures reflect candidate localization rather than relation understanding. Across ten models, benchmark accuracy correlates with ScreenSpot-Pro accuracy (Spearman ρ= +0.74), an exploratory association at this sample size. Marking the two designated candidates raises selection accuracy by 35--57 percentage points, an oracle diagnostic that supplies the candidate set rather than a deployable method. We release the benchmark, predictions, and code.

Open-vocabulary scene understanding is fundamental for robotics, laying the groundwork for spatial reasoning and object manipulation. While closed-vocabulary 3D instance segmentation heavily leverages 3D shape information, state-of-the-art open-vocabulary methods remain predominantly restricted to 2D image features or image-distilled representations during mask labeling. In this paper, we propose SenseFuse, a label-free fusion method that balances 2D image and 3D shape encoders for robust open-vocabulary 3D instance segmentation, refining only the mask-labeling stage of existing pipelines. We reveal that 2D image and 3D shape encoders exhibit largely disjoint failure patterns and rarely share identical wrong labels, whereas two 2D image encoders frequently repeat the same errors. This distinct behavior makes the 2D and 3D pair inherently complementary. We introduce an adaptive mechanism that selects a scene-level fusion weight to maximize a label-free sensitivity measure, estimated directly from a single scene's unlabeled proposals in milliseconds. SenseFuse improves labeling accuracy in every evaluated setting across ScanNet200, Replica, and ScanNet++, recovering 67-100% (median 93%) of the gain achievable with an oracle weight, and it raises instance AP in 21 of 22 reported settings. Code is available at https://github.com/hanes1207/SenseFuse.

Large language models are moving from static text generators toward real-world decision-support systems, where forecasting is a composite capability that links information gathering, evidence integration, situational judgment, and action-oriented decision making. This capability is in broad demand across finance, policy, industry, and scientific research, yet its evaluation remains difficult: live benchmarks evaluate forecasts before answers exist, making them the cleanest way to measure forecasting ability, but they expire once events resolve; retrospective benchmarks are reproducible, but they cannot reliably distinguish genuine forecasting from facts a model may have already learned during pretraining. Prompting models to "pretend not to know" cannot replace a genuine knowledge boundary. We propose OracleProto, a reproducible framework for evaluating LLM native forecasting capability. OracleProto reconstructs resolved events into time-bounded forecasting samples by combining model-cutoff-aligned sample admission, tool-level temporal masking, content-level leakage detection, discrete answer normalization, and hierarchical scoring. Instantiated on a FutureX-Past-derived dataset with six contemporary LLMs, OracleProto distinguishes forecasting quality, sampling stability, and cost efficiency under controlled information boundaries, while reducing residual leakage to the 1% level, an order of magnitude below tool-only temporal filtering. OracleProto turns LLM forecasting from one-off evaluation into an auditable, reusable, and trainable dataset-level capability, providing a unified interface for fair cross-model comparison and a controlled signal source for downstream SFT and RL. Code and data are available at https://github.com/MaYiding/OracleProto and https://huggingface.co/datasets/MaYiding/OracleProto.

Recent advancements in large language models (LLMs) often rely on generating intermediate reasoning steps to enhance accuracy. However, little work has examined how reasoning utility contributes to the final answer's correctness. Due to the stochastic nature of autoregressive generation, generating more context does not guarantee increased confidence in the answer. If we could predict, during generation, whether a reasoning step will be useful, we could stop early or prune ineffective steps, avoiding distractions in the final decision. We present an oracle study on MATH dataset, using Qwen2.5-32B and GPT-4o to generate reasoning chains, and then employing a separate model (Qwen3-8B) to quantify the utility of these chains for final accuracy. Specifically, we measure the model's uncertainty on the answer span Y at each reasoning step using conditional entropy (expected negative log-likelihood over the vocabulary) with context expanding step by step. Our results show a clear pattern: conditional entropy that decreases over steps is strongly associated with correct answers, whereas flat or increasing entropy often results in wrong answers. We also corroborate that incorrect reasoning paths tend to be longer than correct ones, suggesting that longer reasoning does not necessarily yield better outcomes. These findings serve as a foundation to inspire future work on designing efficient reasoning pipelines that detect and avoid unproductive reasoning early.

This paper studies ensembling in the era of Large Vision-Language Models (LVLMs). Ensembling is a classical method to combine different models to get increased performance. In the recent work on Encyclopedic-VQA the authors examine a wide variety of models to solve their task: from vanilla LVLMs, to models including the caption as extra context, to models augmented with Lens-based retrieval of Wikipedia pages. Intuitively these models are highly complementary, which should make them ideal for ensembling. Indeed, an oracle experiment shows potential gains from 48.8% accuracy (the best single model) all the way up to 67% (best possible ensemble). So it is a trivial exercise to create an ensemble with substantial real gains. Or is it?

Modern Large Language Models achieve impressive reasoning capabilities with long Chain of Thoughts, but they incur substantial computational cost during inference, and this motivates techniques to improve the performance-cost ratio. Among these techniques, Speculative Decoding accelerates inference by employing a fast but inaccurate draft model to autoregressively propose tokens, which are then verified in parallel by a more capable target model. However, due to unnecessary rejections caused by token mismatches in semantically equivalent steps, traditional token-level Speculative Decoding struggles in reasoning tasks. Although recent works have shifted to step-level semantic verification, which improve efficiency by accepting or rejecting entire reasoning steps, existing step-level methods still regenerate many rejected steps with little improvement, wasting valuable target compute. To address this challenge, we propose Arbitrage, a novel step-level speculative generation framework that routes generation dynamically based on the relative advantage between draft and target models. Instead of applying a fixed acceptance threshold, Arbitrage uses a lightweight router trained to predict when the target model is likely to produce a meaningfully better step. This routing approximates an ideal Arbitrage Oracle that always chooses the higher-quality step, achieving near-optimal efficiency-accuracy trade-offs. Across multiple mathematical reasoning benchmarks, Arbitrage consistently surpasses prior step-level Speculative Decoding baselines, reducing inference latency by up to sim2times at matched accuracy.

Time series question-answering (TSQA), in which we ask natural language questions to infer and reason about properties of time series, is a promising yet underexplored capability of foundation models. In this work, we present ARFBench, a TSQA benchmark that evaluates the understanding of multimodal foundation models (FMs) on time series anomalies prevalent in software incident data. ARFBench consists of 750 questions across 142 time series and 5.38M data points from 63 production incidents sourced exclusively from internal telemetry at Datadog. We evaluate leading proprietary and open-source LLMs, VLMs, and time series FMs and observe that frontier VLMs perform markedly better than existing baselines; the leading model (GPT-5) achieves a 62.7% accuracy and 51.9% F1. We next demonstrate the promise of specialized multimodal approaches. We develop a novel TSFM + VLM hybrid prototype which we post-train on a small set of synthetic and real data that yields comparable overall F1 and accuracy with frontier models. Lastly, we find models and human domain experts exhibit complementary strengths. We define a model-expert oracle, a best-of-2 oracle selector over model and expert answers, yielding 82.8% F1 and 87.2% accuracy and establishing a new superhuman frontier for future TSQA models. The benchmark is available at https://huggingface.co/datasets/Datadog/ARFBench.

Test-time scaling has shown considerable success in improving the performance of language models on complex reasoning tasks without requiring fine-tuning. However, current strategies such as self-reflection primarily focus on logical or structural refinement. They do not leverage the guiding potential of affective feedback. Inspired by psychological research showing that emotions can modulate cognitive performance, we introduce HEART--a novel framework that uses emotionally-driven prompts for iterative self-correction. HEART provides feedback on a model's incorrect response using a curated set of concise, emotionally charged phrases based on the six universal emotions categorized by Dr. Paul Ekman. By systematically varying the emotional tone of the feedback across iterations, our method guides the model to escape flawed reasoning paths and explore more promising alternatives. We evaluate our framework on challenging reasoning benchmarks including OlympiadBench, Humanity's Last Exam, and SimpleQA. Our results reveal a significant new phenomenon: when guided by an oracle verifier, this affective iteration protocol unlocks significantly deeper reasoning, leading to consistent and substantial increases in accuracy over state-of-the-art baselines with the same verifier. However, we also identify a critical bottleneck for practical deployment. In a verifier-free setting, it struggles to harness these gains consistently, highlighting as a key challenge for future work. Our findings suggest that the next frontier in machine reasoning may lie not just in refining logic, but also in understanding and leveraging the `HEART' of the models.

Neurosymbolic systems rely on mathematical solvers to guarantee reasoning correctness, yet solvers are fundamentally blind to whether a formal translation maintains strict reference-equivalence to a designated formalization. We formalize this vulnerability as Verdict-Preserving-Unfaithfulness (VPU): a failure mode where an incorrect encoding executes successfully and matches the expected verdict. We theoretically prove that structural, verdict-only verification heuristics are mathematically bounded to chance-level detection on these deceptively valid traces. To resolve this, we introduce Generative Verification (GenV), which distills an offline Z3-equivalence oracle into a reference-free, continuous reference-equivalence score by repurposing the language model's native vocabulary space. Mechanistic analysis via decision-projected logit lenses and sparse autoencoders shows this generative readout natively extracts precise spatial error coordinates without explicit localization training. Empirically, our oracle-mined verifier (GenV+HN) achieves 0.961 AUROC in reference-equivalence verification, generalizes zero-shot across unseen translators and divergent formal styles, and yields an 11.3-point downstream accuracy gain in agentic test-time compute allocation.

An increasing number of vision-language tasks can be handled with little to no training, i.e., in a zero and few-shot manner, by marrying large language models (LLMs) to vision encoders, resulting in large vision-language models (LVLMs). While this has huge upsides, such as not requiring training data or custom architectures, how an input is presented to a LVLM can have a major impact on zero-shot model performance. In particular, inputs phrased in an underspecified way can result in incorrect answers due to factors like missing visual information, complex implicit reasoning, or linguistic ambiguity. Therefore, adding visually grounded information to the input as a preemptive clarification should improve model performance by reducing underspecification, e.g., by localizing objects and disambiguating references. Similarly, in the VQA setting, changing the way questions are framed can make them easier for models to answer. To this end, we present Rephrase, Augment and Reason (RepARe), a gradient-free framework that extracts salient details about the image using the underlying LVLM as a captioner and reasoner, in order to propose modifications to the original question. We then use the LVLM's confidence over a generated answer as an unsupervised scoring function to select the rephrased question most likely to improve zero-shot performance. Focusing on two visual question answering tasks, we show that RepARe can result in a 3.85% (absolute) increase in zero-shot performance on VQAv2 and a 6.41% point increase on A-OKVQA. Additionally, we find that using gold answers for oracle question candidate selection achieves a substantial gain in VQA accuracy by up to 14.41%. Through extensive analysis, we demonstrate that outputs from RepARe increase syntactic complexity, and effectively utilize vision-language interaction and the frozen language model in LVLMs.

We propose a physics-informed neural particle method (PINN--PM) for the spatially homogeneous Landau equation. The method adopts a Lagrangian interacting-particle formulation and jointly parameterizes the time-dependent score and the characteristic flow map with neural networks. Instead of advancing particles through explicit time stepping, the Landau dynamics is enforced via a continuous-time residual defined along particle trajectories. This design removes time-discretization error and yields a mesh-free solver that can be queried at arbitrary times without sequential integration. We establish a rigorous stability analysis in an L^2_v framework. The deviation between learned and exact characteristics is controlled by three interpretable sources: (i) score approximation error, (ii) empirical particle approximation error, and (iii) the physics residual of the neural flow. This trajectory estimate propagates to density reconstruction, where we derive an L^2_v error bound for kernel density estimators combining classical bias--variance terms with a trajectory-induced contribution. Using Hyvarinen's identity, we further relate the oracle score-matching gap to the L^2_v score error and show that the empirical loss concentrates at the Monte Carlo rate, yielding computable a posteriori accuracy certificates. Numerical experiments on analytical benchmarks, including the two- and three-dimensional BKW solutions, as well as reference-free configurations, demonstrate stable transport, preservation of macroscopic invariants, and competitive or improved accuracy compared with time-stepping score-based particle and blob methods while using significantly fewer particles.

When an LLM-based agent improves on a task, is the gain from the model itself or from the reasoning paradigm wrapped around it? We study this question by comparing six inference-time paradigms, namely Direct, CoT, ReAct, Plan-Execute, Reflection, and ReCode, across four frontier LLMs and ten benchmarks, yielding roughly 18,000 runs. We find that reasoning structure helps dramatically on some tasks but hurts on others: ReAct improves over Direct by 44pp on GAIA, while CoT degrades performance by 15pp on HumanEval. No single paradigm dominates, and oracle per-task selection beats the best fixed paradigm by 17.1pp on average. Motivated by this complementarity, we propose a select-then-solve approach: before answering each task, a lightweight embedding-based router selects the most suitable paradigm. Across four models, the router improves average accuracy from 47.6% to 53.1%, outperforming the best fixed paradigm at 50.3% by 2.8pp and recovering up to 37% of the oracle gap. In contrast, zero-shot self-routing only works for GPT-5 at 67.1% and fails for weaker models, all trailing the learned router. Our results argue that reasoning paradigm selection should be a per-task decision made by a learned router, not a fixed architectural choice.

Static benchmarks measure a model frozen at training time. Real systems face distribution shift: new categories, paraphrased queries, drift: and must recover online via user corrections. No existing benchmark measures recovery speed under correction streams. We introduce OCRR (Online Correction Recovery Rate): a benchmark that streams a corpus through a classification system, applies oracle or stochastic corrections to wrong predictions, and reports two curves: novel-class accuracy and original-distribution accuracy versus correction count. We evaluate the substrate alongside nine baseline algorithms from five families plus seven bounded-storage variants of the substrate for the Pareto sweep, including standard online-learning baselines (river), continual-learning methods (EWC, A-GEM, LwF), retrieval/parametric hybrids (kNN-LM), parameter-efficient fine-tuning of a 1.5 B-parameter encoder (LoRA on DeBERTa-v3-large), and a hash-chained append-only substrate (Substrate). On Banking77 and CLINC150, under oracle and sparse correction policies, the substrate is the only system that simultaneously recovers novel-class accuracy (88.7 +/- 2.9 %) and retains original-distribution accuracy (95.4 +/- 0.8 %) beating the next-best published continual-learning baseline by 32.6 percentage points at equal memory budget, and beating LoRA-on-DeBERTa-v3-large by 84.6 percentage points on retention. We further find that classification accuracy remains stable at 99 % even as approximate-nearest-neighbour recall@5 degrades from 0.69 to 0.23 across 10 k to 10 M corpus scales, suggesting the substrate's margin-band majority vote is robust to retrieval imperfection in a way that pure top-k recall metrics do not predict. Code and data are available at https://github.com/adriangrassi/ocrr-benchmark.

LLM-based autonomous agents lack persistent procedural memory: they re-derive solutions from scratch even when structurally identical tasks have been solved before. We present APEX-EM, a non-parametric online learning framework that accumulates, retrieves, and reuses structured procedural plans without modifying model weights. APEX-EM introduces: (1) a structured experience representation encoding the full procedural-episodic trace of each execution -- planning steps, artifacts, iteration history with error analysis, and quality scores; (2) a Plan-Retrieve-Generate-Iterate-Ingest (PRGII) workflow with Task Verifiers providing multi-dimensional reward signals; and (3) a dual-outcome Experience Memory with hybrid retrieval combining semantic search, structural signature matching, and plan DAG traversal -- enabling cross-domain transfer between tasks sharing no lexical overlap but analogous operational structure. Successful experiences serve as positive in-context examples; failures as negative examples with structured error annotations. We evaluate on BigCodeBench, KGQAGen-10k, and Humanity's Last Exam using Claude Sonnet 4.5 and Opus 4.5. On KGQAGen-10k, APEX-EM achieves 89.6% accuracy versus 41.3% without memory (+48.3pp), surpassing the oracle-retrieval upper bound (84.9%). On BigCodeBench, it reaches 83.3% SR from a 53.9% baseline (+29.4pp), exceeding MemRL's +11.0pp gain under comparable frozen-backbone conditions (noting backbone differences controlled for in our analysis). On HLE, entity graph retrieval reaches 48.0% from 25.2% (+22.8pp). Ablations show component value is task-dependent: rich judge feedback is negligible for code generation but critical for structured queries (+10.3pp), while binary-signal iteration partially compensates for weaker feedback.

Large Language Models (LLMs) are evolving from text generators into reasoning agents. This transition makes their ability to use external tools a critical capability. However, evaluating this skill presents a significant challenge. Existing benchmarks are often limited by their reliance on synthetic tools and severely constrained action spaces. To address these limitations, we introduce MCPVerse, an expansive, real-world benchmark for evaluating agentic tool use. MCPVerse integrates more than 550 real-world, executable tools to create an unprecedented action space exceeding 140k tokens, and employs outcome-based evaluation with real-time ground truth for time-sensitive tasks. We benchmarked the state-of-the-art LLMs across three modes (Oracle, Standard, and Max-Scale), revealing that while most models suffer performance degradation when confronted with larger tool sets, the agentic models, such as Claude-4-Sonnet, can effectively leverage expanded exploration spaces to improve accuracy. This finding not only exposes the limitations of state-of-the-art models in complex, real-world scenarios but also establishes MCPVerse as a critical benchmark for measuring and advancing agentic tool use capabilities.

The study of ancient writings has great value for archaeology and philology. Essential forms of material are photographic characters, but manual photographic character recognition is extremely time-consuming and expertise-dependent. Automatic classification is therefore greatly desired. However, the current performance is limited due to the lack of annotated data. Data generation is an inexpensive but useful solution for data scarcity. Nevertheless, the diverse glyph shapes and complex background textures of photographic ancient characters make the generation task difficult, leading to the unsatisfactory results of existing methods. In this paper, we propose an unsupervised generative adversarial network called AGTGAN. By the explicit global and local glyph shape style modeling followed by the stroke-aware texture transfer, as well as an associate adversarial learning mechanism, our method can generate characters with diverse glyphs and realistic textures. We evaluate our approach on the photographic ancient character datasets, e.g., OBC306 and CSDD. Our method outperforms the state-of-the-art approaches in various metrics and performs much better in terms of the diversity and authenticity of generated samples. With our generated images, experiments on the largest photographic oracle bone character dataset show that our method can achieve a significant increase in classification accuracy, up to 16.34%.

LLM-as-judge evaluation has become the de facto standard for scaling model assessment, but the practice is statistically unsound: uncalibrated scores can invert preferences, naive confidence intervals on uncalibrated scores achieve near-0% coverage, and importance-weighted estimators collapse under limited overlap despite high effective sample size (ESS). We introduce Causal Judge Evaluation (CJE), a framework that fixes all three failures. On n=4,961 Chatbot Arena prompts (after filtering from 5k), CJE achieves 99% pairwise ranking accuracy at full sample size (94% averaged across configurations), matching oracle quality, at 14x lower cost (for ranking 5 policies) by calibrating a 16x cheaper judge on just 5% oracle labels (~250 labels). CJE combines three components: (i) AutoCal-R, reward calibration via mean-preserving isotonic regression; (ii) SIMCal-W, weight stabilization via stacking of S-monotone candidates; and (iii) Oracle-Uncertainty Aware (OUA) inference that propagates calibration uncertainty into confidence intervals. We formalize the Coverage-Limited Efficiency (CLE) diagnostic, which explains why IPS-style estimators fail even when ESS exceeds 90%: the logger rarely visits regions where target policies concentrate. Key findings: SNIPS inverts rankings even with reward calibration (38% pairwise, negative Kendall's tau) due to weight instability; calibrated IPS remains near-random (47%) despite weight stabilization, consistent with CLE; OUA improves coverage from near-0% to ~86% (Direct) and ~96% (stacked-DR), where naive intervals severely under-cover.

VLMs trained on web-scale data retain sensitive and copyrighted visual concepts that deployment may require removing. Training-based unlearning methods share a structural flaw: fine-tuning on a narrow forget set degrades general capabilities before unlearning begins, making it impossible to attribute subsequent performance drops to the unlearning procedure itself. Training-free approaches sidestep this by suppressing concepts through prompts or system instructions, but no rigorous benchmark exists for evaluating them on visual tasks. We introduce VLM-UnBench, the first benchmark for training-free visual concept unlearning in VLMs. It covers four forgetting levels, 7 source datasets, and 11 concept axes, and pairs a three-level probe taxonomy with five evaluation conditions to separate genuine forgetting from instruction compliance. Across 8 evaluation settings and 13 VLM configurations, realistic unlearning prompts leave forget accuracy near the no-instruction baseline; meaningful reductions appear only under oracle conditions that disclose the target concept to the model. Object and scene concepts are the most resistant to suppression, and stronger instruction-tuned models remain capable despite explicit forget instructions. These results expose a clear gap between prompt-level suppression and true visual concept erasure.

Continual instruction tuning updates a language model through a sequence of new domains, yet each update can progressively erode previously learned capabilities and alignment behavior. Replay is the standard mitigation, but fixed replay ratios are inherently limited because the optimal mixture varies with the current domain, the training stage, and the evolving vulnerability of prior behaviors. We propose PROX-YMIX, a framework that learns a dynamic replay controller on a small proxy model and transfers the frozen controller to a larger target. The controller never observes future tasks and constructs its state from normalized validation losses and their temporal dynamics, producing a masked mixture over the current task and accessible replay buffers. Our core empirical hypothesis is forgetting mirroring: task vulnerability rankings remain largely consistent across model scales even when absolute loss magnitudes differ. We validate this assumption empirically before transferring controllers across scales. On LLaMA-3-8B across five continual instruction tuning sequences, PROXYMIX improves average accuracy by 3.4 points, reduces final forgetting by 3.5 points, and raises safety score by 5.8 points over the strongest non-oracle baseline, at roughly 50x lower policy learning cost than Oracle Target RL. The framework is leakage free and architecture independent at the interface level, and we also identify settings where the proxy assumption breaks down, highlighting limitations for robust deployment.

KV-cache eviction caps the memory cost of long reasoning traces but is inherently lossy because the model decodes from a partial view of its history. Under aggressive budgets, this not only lowers accuracy but can also cause runaway degeneration, where the model produces incoherent or repetitive tokens until reaching the length limit. We characterize much of this loss as an information gapf caused by missing context, rather than a capability gap caused by limited model capacity. An evicted 7B model and a full-context 1.5B model make complementary errors, and an oracle choice between their answers recovers 79% of the accuracy gap to the full-KV 7B model. Based on this observation, we propose KV-Rescue, a training-free inference framework that bridges the information gap introduced by KV eviction using a lightweight full-context helper. KV-Rescue interleaves reasoning steps from the two models into a shared trajectory. An online detector uses entropy and compressibility to terminate the generation of incoherent or repetitive base-model candidates early. Across five math benchmarks with Qwen2.5-Math 7B and 72B, KV-Rescue recovers an average of 87% of the accuracy lost to eviction at eviction budget B=64. A decode-cost analysis further shows that preventing runaway degeneration cuts base-model token generation by 43% on average.

Large Language Models (LLMs) are now integral across various domains and have demonstrated impressive performance. Progress, however, rests on the premise that benchmark scores are both accurate and reproducible. We demonstrate that the reproducibility of LLM performance is fragile: changing system configuration such as evaluation batch size, GPU count, and GPU version can introduce significant difference in the generated responses. This issue is especially pronounced in reasoning models, where minor rounding differences in early tokens can cascade into divergent chains of thought, ultimately affecting accuracy. For instance, under bfloat16 precision with greedy decoding, a reasoning model like DeepSeek-R1-Distill-Qwen-7B can exhibit up to 9% variation in accuracy and 9,000 tokens difference in response length due to differences in GPU count, type, and evaluation batch size. We trace the root cause of this variability to the non-associative nature of floating-point arithmetic under limited numerical precision. This work presents the first systematic investigation into how numerical precision affects reproducibility in LLM inference. Through carefully controlled experiments across various hardware, software, and precision settings, we quantify when and how model outputs diverge. Our analysis reveals that floating-point precision -- while critical for reproducibility -- is often neglected in evaluation practices. Inspired by this, we develop a lightweight inference pipeline, dubbed LayerCast, that stores weights in 16-bit precision but performs all computations in FP32, balancing memory efficiency with numerical stability. Code is available at https://github.com/nanomaoli/llm_reproducibility.

This paper introduces the task of visual named entity discovery in videos without the need for task-specific supervision or task-specific external knowledge sources. Assigning specific names to entities (e.g. faces, scenes, or objects) in video frames is a long-standing challenge. Commonly, this problem is addressed as a supervised learning objective by manually annotating faces with entity labels. To bypass the annotation burden of this setup, several works have investigated the problem by utilizing external knowledge sources such as movie databases. While effective, such approaches do not work when task-specific knowledge sources are not provided and can only be applied to movies and TV series. In this work, we take the problem a step further and propose to discover entities in videos from videos and corresponding captions or subtitles. We introduce a three-stage method where we (i) create bipartite entity-name graphs from frame-caption pairs, (ii) find visual entity agreements, and (iii) refine the entity assignment through entity-level prototype construction. To tackle this new problem, we outline two new benchmarks SC-Friends and SC-BBT based on the Friends and Big Bang Theory TV series. Experiments on the benchmarks demonstrate the ability of our approach to discover which named entity belongs to which face or scene, with an accuracy close to a supervised oracle, just from the multimodal information present in videos. Additionally, our qualitative examples show the potential challenges of self-contained discovery of any visual entity for future work. The code and the data are available on GitHub.

Verifiers can improve language model capabilities by scoring and ranking responses from generated candidates. Currently, high-quality verifiers are either unscalable (e.g., humans) or limited in utility (e.g., tools like Lean). While LM judges and reward models have become broadly useful as general-purpose verifiers, a significant performance gap remains between them and oracle verifiers (verifiers with perfect accuracy). To help close this gap, we introduce Weaver, a framework for designing a strong verifier by combining multiple weak, imperfect verifiers. We find weighted ensembles of verifiers, which typically require learning from labeled data, significantly outperform unweighted combinations due to differences in verifier accuracies. To reduce dependency on labeled data, Weaver leverages weak supervision to estimate each verifier's accuracy and combines outputs into a unified score that better reflects true response quality. However, directly applying weak supervision algorithms poses challenges, including inconsistent verifier output formats and handling low-quality verifiers. Weaver addresses these using dataset statistics to normalize outputs and filter specific verifiers. We study Weaver's effectiveness in test-time repeated sampling, where a model generates multiple candidate responses and selects one. Our evaluations show Weaver significantly improves over Pass@1-performance when selecting the first candidate-across reasoning and math tasks, achieving o3-mini-level accuracy with Llama 3.3 70B Instruct as generator, and an ensemble of 70B or smaller judge and reward models as verifiers (87.7% average). This gain mirrors the jump between GPT-4o and o3-mini (69.0% vs. 86.7%), which required extensive finetuning and post-training. To reduce computational costs of verifier ensembles, we train a 400M cross-encoder using Weaver's combined output scores.

Time series captioning, the task of describing numeric time series in natural language, requires numerical reasoning, trend interpretation, and contextual understanding. Existing benchmarks, however, often rely on synthetic data or overly simplistic captions, and typically neglect metadata and visual representations. To close this gap, we introduce CaTS-Bench, the first large-scale, real-world benchmark for Context-aware Time Series captioning. CaTS-Bench is derived from 11 diverse datasets reframed as captioning and Q&A tasks, comprising roughly 465k training and 105k test timestamps. Each sample includes a numeric series segment, contextual metadata, a line-chart image, and a caption. A key contribution of this work is the scalable pipeline used to generate reference captions: while most references are produced by an oracle LLM and verified through factual checks, human indistinguishability studies, and diversity analyses, we also provide a human-revisited subset of 579 test captions, refined from LLM outputs to ensure accuracy and human-like style. Beyond captioning, CaTS-Bench offers 460 multiple-choice questions targeting deeper aspects of time series reasoning. We further propose new tailored evaluation metrics and benchmark leading VLMs, highlighting both their strengths and persistent limitations. Together, these contributions establish CaTS-Bench and its captioning pipeline as a reliable and extensible foundation for future research at the intersection of time series analysis and foundation models.

Building local surrogates to accelerate stationary point searches on potential energy surfaces spans decades of effort. Done correctly, surrogates can reduce the number of expensive electronic structure evaluations by roughly an order of magnitude while preserving the accuracy of the underlying theory, with the gain depending on oracle cost, search distance, and the availability of analytical forces. We present a unified Bayesian optimization view of minimization, single-point saddle searches, and double-ended path searches: all three share one six-step surrogate loop and differ only in the inner optimization target and the acquisition criterion. The framework uses Gaussian process regression with derivative observations, inverse-distance kernels, and active learning, and we develop optional extensions for production use, including farthest-point sampling with the Earth Mover's Distance, MAP regularization, an adaptive trust radius, and random Fourier features for scaling. Accompanying pedagogical Rust code demonstrates that all three applications use the same Bayesian optimization loop, bridging the gap between theoretical formulation and practical execution.

Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400times, end-to-end speedup of up to 3.7times as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.

Deep research agents are increasingly evaluated on their ability to search for evidence, reason over retrieved sources, and produce grounded answers. Existing browsing benchmarks, however, largely assume that the user's query and the supporting evidence are written in the same language, leaving open whether agentic search systems can operate when relevant evidence appears in another language. We introduce XBCP (Cross-lingual BrowseComp-Plus), a controlled benchmark that preserves the English question-and-answer space of BrowseComp-Plus but varies the languages of the supporting documents. XBCP instantiates two complementary settings: in the cross-lingual setting, each query is paired with evidence in a single assigned language. In the multilingual setting, the full evidence corpus is distributed equally and randomly across 12 languages spanning high-resource and low-resource regimes. We evaluate four deep research agents using sparse and dense multilingual retrievers, measuring answer accuracy, evidence recall, search behavior, calibration, citation fidelity, and oracle retrieval. Results reveal substantial degradation when evidence is translated. Even strong, dense retrievers lose evidence recall, and agents become less calibrated and cite evidence less reliably. Notably, accuracy remains lower even when all gold evidence is supplied directly. These findings suggest that cross-lingual deep research exposes both retrieval failures and an independent, agent-side difficulty in integrating language-mismatched evidence.

Large-taxonomy retrieval often assumes that the input already expresses the target concept. In many settings, however, the input is indirect evidence, such as a table cell whose meaning depends on its row, column, datatype, and context. We call this mismatch the retrieval readiness gap. Our analysis shows that the current index retrieves the target reliably when its semantics are explicit, while raw evidence often leaves it deep in the ranking. We propose Factorized Hypothesis Search (FHS), which maintains multiple partial interpretations over named semantic dimensions. These hypotheses support structured query rendering, multi-hypothesis retrieval, and dimension-level candidate verification. On both financial taxonomy tagging and CodiEsp clinical coding tasks, FHS achieves the best Recall@1, MRR, and final accuracy among the non-oracle methods. Replacing the factorized hypothesis path with a free-text ensemble causes the largest drop in head-ranking performance, while sequential refinement provides no additional gain over FHS's strong parallel first round.

We consider an online learning problem with one-sided feedback, in which the learner is able to observe the true label only for positively predicted instances. On each round, k instances arrive and receive classification outcomes according to a randomized policy deployed by the learner, whose goal is to maximize accuracy while deploying individually fair policies. We first extend the framework of Bechavod et al. (2020), which relies on the existence of a human fairness auditor for detecting fairness violations, to instead incorporate feedback from dynamically-selected panels of multiple, possibly inconsistent, auditors. We then construct an efficient reduction from our problem of online learning with one-sided feedback and a panel reporting fairness violations to the contextual combinatorial semi-bandit problem (Cesa-Bianchi & Lugosi, 2009, Gy\"{o}rgy et al., 2007). Finally, we show how to leverage the guarantees of two algorithms in the contextual combinatorial semi-bandit setting: Exp2 (Bubeck et al., 2012) and the oracle-efficient Context-Semi-Bandit-FTPL (Syrgkanis et al., 2016), to provide multi-criteria no regret guarantees simultaneously for accuracy and fairness. Our results eliminate two potential sources of bias from prior work: the "hidden outcomes" that are not available to an algorithm operating in the full information setting, and human biases that might be present in any single human auditor, but can be mitigated by selecting a well chosen panel.

Large Multimodal Models (LMMs) such as LLaVA have shown strong performance in visual-linguistic reasoning. These models first embed images into a fixed large number of visual tokens and then feed them into a Large Language Model (LLM). However, this design causes an excessive number of tokens for dense visual scenarios such as high-resolution images and videos, leading to great inefficiency. While token pruning/merging methods do exist, they produce a single length output for each image and do not afford flexibility in trading off information density v.s. efficiency. Inspired by the concept of Matryoshka Dolls, we propose M3: Matryoshka Multimodal Models, which learns to represent visual content as nested sets of visual tokens that capture information across multiple coarse-to-fine granularities. Our approach offers several unique benefits for LMMs: (1) One can explicitly control the visual granularity per test instance during inference, e.g. , adjusting the number of tokens used to represent an image based on the anticipated complexity or simplicity of the content; (2) M3 provides a framework for analyzing the granularity needed for existing datasets, where we find that COCO-style benchmarks only need around ~9 visual tokens to obtain accuracy similar to that of using all 576 tokens; (3) Our approach provides a foundation to explore the best trade-off between performance and visual token length at sample level, where our investigation reveals that a large gap exists between the oracle upper bound and current fixed-scale representations.

Computer-use agents pay full frontier inference to re-derive routines their user has already performed, because an agent's memory today records what the user said, not what the user did. We compile passively captured screen activity into agent memory with a deterministic, zero-model pipeline: it segments a local capture stream into typed activity frames, bounded episodes carrying application, site, timing, input volume, and evidence pointers back to the raw rows, with no model in the loop, so the output is byte-identical, cacheable, and mechanically auditable. On one professional's single-user corpus of 128,756 frames over 51 active days, the compiler reduces a day of raw capture to a prompt-ready context block 86x smaller in 68 ms, and an agent reading that block answers questions about the day at 98.4% accuracy (Wilson 95% CI 91.7-99.7%) against an independent oracle, versus 66-80% for an LLM summary of the same capture, a mid-tier model reading the block matching a frontier one. The same compiler doubles as a demand-side cost instrument. Read off passive, pre-delegation human activity rather than agent rollouts, it supplies two parameters that agent-cost models assume but, to our knowledge, have not measured: the Routine Overhead Ratio R and the routine recurrence h. We report first values of R, a modeled upper bound, at 60-343x, and a delegable recurrence of 9.0% in-sample and 7.7% out-of-sample, for a realistic all-fleet token ceiling near 8%; a compiled routine replays deterministically with the model out of the loop, demonstrated live at zero model tokens on a guard-matched hit. Schema, compiler, and evaluation harness are open.

Source separation for music is the task of isolating contributions, or stems, from different instruments recorded individually and arranged together to form a song. Such components include voice, bass, drums and any other accompaniments.Contrarily to many audio synthesis tasks where the best performances are achieved by models that directly generate the waveform, the state-of-the-art in source separation for music is to compute masks on the magnitude spectrum. In this paper, we compare two waveform domain architectures. We first adapt Conv-Tasnet, initially developed for speech source separation,to the task of music source separation. While Conv-Tasnet beats many existing spectrogram-domain methods, it suffersfrom significant artifacts, as shown by human evaluations. We propose instead Demucs, a novel waveform-to-waveform model,with a U-Net structure and bidirectional LSTM.Experiments on the MusDB dataset show that, with proper data augmentation, Demucs beats allexisting state-of-the-art architectures, including Conv-Tasnet, with 6.3 SDR on average, (and up to 6.8 with 150 extra training songs, even surpassing the IRM oracle for the bass source).Using recent development in model quantization, Demucs can be compressed down to 120MBwithout any loss of accuracy.We also provide human evaluations, showing that Demucs benefit from a large advantagein terms of the naturalness of the audio. However, it suffers from some bleeding,especially between the vocals and other source.

Agent harnesses supply planning guidance, organize execution, and check completion. We study how these components affect success, erroneous acceptance, and cost in two Retail experiments and an Airline pilot in τ^2-bench. The primary comparison pairs prewritten task-specific plans (Fixed) with shuffled policy text matched in word count (Sham), isolating the contribution of guidance content. Across 265 matched cells, Fixed improves oracle-verified success by 7.17 percentage points (90\% task-clustered bootstrap interval, 1.15--13.36 points), with gains concentrated in higher-complexity tasks. A read-only terminal verifier rejects 61\% of Retail oracle-invalid episodes while withholding 17\% of correct ones, at less than one cent of additional cost per episode. Which component matters more depends on the loss assigned to erroneous acceptance: at low liability the planning gain dominates; at high liability the verifier's avoided false passes dominate---and a standalone verifier captures nearly all the false-pass benefit of the full planning-plus-verification stack at a fraction of its cost.

Parsing natural language to corresponding SQL (NL2SQL) with data driven approaches like deep neural networks attracts much attention in recent years. Existing NL2SQL datasets assume that condition values should appear exactly in natural language questions and the queries are answerable given the table. However, these assumptions may fail in practical scenarios, because user may use different expressions for the same content in the table, and query information outside the table without the full picture of contents in table. Therefore we present TableQA, a large-scale cross-domain Natural Language to SQL dataset in Chinese language consisting 64,891 questions and 20,311 unique SQL queries on over 6,000 tables. Different from exisiting NL2SQL datasets, TableQA requires to generalize well not only to SQL skeletons of different questions and table schemas, but also to the various expressions for condition values. Experiment results show that the state-of-the-art model with 95.1% condition value accuracy on WikiSQL only gets 46.8% condition value accuracy and 43.0% logic form accuracy on TableQA, indicating the proposed dataset is challenging and necessary to handle. Two table-aware approaches are proposed to alleviate the problem, the end-to-end approaches obtains 51.3% and 47.4% accuracy on the condition value and logic form tasks, with improvement of 4.7% and 3.4% respectively.

Enterprise ERP systems managing hundreds of thousands of employee records face critical data quality challenges when human resources departments perform decentralized manual entry across multiple languages. We present an end-to-end pipeline combining automated data cleaning with LLM-driven SQL query generation, deployed on a production system managing 240,000 employee records over six months. The system operates in two integrated stages: a multi-stage cleaning pipeline that performs translation normalization, spelling correction, and entity deduplication during periodic synchronization from Microsoft SQL Server to PostgreSQL; and a retrieval-augmented generation framework powered by GPT-4o that translates natural-language questions in Turkish, Russian, and English into validated SQL queries. The query engine employs LangChain orchestration, FAISS vector similarity search, and few-shot learning with 500+ validated examples. Our evaluation demonstrates 92.5% query validity, 95.1% schema compliance, and 90.7\% semantic accuracy on 2,847 production queries. The system reduces query turnaround time from 2.3 days to under 5 seconds while maintaining 99.2% uptime, with GPT-4o achieving 46% lower latency and 68% cost reduction versus GPT-3.5. This modular architecture provides a reproducible framework for AI-native enterprise data governance, demonstrating real-world viability at enterprise scale with 4.3/5.0 user satisfaction.

While there are numerous benchmarks comparing the performance of modern language models (LMs), end-task evaluations often conflate notions of *factual accuracy* ("truth") and *reasoning ability* ("rationality", or "honesty" in the sense of correctly reporting implications of beliefs). Our goal is a dataset that clearly distinguishes these two notions. Our approach is to leverage and extend a collection of human-annotated *entailment trees*, engineered to express both good and bad chains of reasoning, and using a mixture of true and false facts, in particular including counterfactual examples, to avoid belief bias (also known as the "content effect"). The resulting dataset, called BaRDa, contains 3000 entailments (1787 valid, 1213 invalid), using 6681 true and 2319 false statements. Testing on four GPT-series models, GPT3(curie)/GPT3(davinici)/3.5/4, we find factual accuracy (truth) scores of 74.1/80.6/82.6/87.1 and reasoning accuracy scores of 63.1/78.0/71.8/79.2. This shows the clear progression of models towards improved factual accuracy and entailment reasoning, and the dataset provides a new benchmark that more cleanly separates and quantifies these two notions.

As LLMs are deployed in high-stakes settings, users must judge the correctness of individual responses, often relying on model-generated justifications such as reasoning chains or explanations. Yet, no standard measure exists for whether these justifications help users distinguish correct answers from incorrect ones. We formalize this idea as error verifiability and propose v_{bal}, a balanced metric that measures whether justifications enable raters to accurately assess answer correctness, validated against human raters who show high agreement. We find that neither common approaches, such as post-training and model scaling, nor more targeted interventions recommended improve verifiability. We introduce two methods that succeed at improving verifiability: reflect-and-rephrase (RR) for mathematical reasoning and oracle-rephrase (OR) for factual QA, both of which improve verifiability by incorporating domain-appropriate external information. Together, our results establish error verifiability as a distinct dimension of response quality that does not emerge from accuracy improvements alone and requires dedicated, domain-aware methods to address.

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

Current Text-to-SQL methods are evaluated and only focused on executable queries, overlooking the semantic alignment challenge -- both in terms of the semantic meaning of the query and the correctness of the execution results. Even execution accuracy itself shows significant drops when moving from English to other languages, with an average decline of 6 percentage points across non-English languages. We address these challenges by presenting a new framework that combines Group Relative Policy Optimization (GRPO) within a multilingual contrastive reward signal to enhance both task efficiency and semantic accuracy in Text-to-SQL systems in cross-lingual scenarios. Our method teaches models to obtain better correspondence between SQL generation and user intent by combining a reward signal based on semantic similarity. On the seven-language MultiSpider dataset, fine-tuning the LLaMA-3-3B model with GRPO improved the execution accuracy up to 87.4 percent (+26 pp over zero-shot) and semantic accuracy up to 52.29 percent (+32.86 pp). Adding our contrastive reward signal in the GRPO framework further improved the average semantic accuracy to 59.14 percent (+6.85 pp, up to +10 pp for Vietnamese). Our experiments showcase that a smaller, parameter-efficient 3B LLaMA model fine-tuned with our contrastive reward signal outperforms a much larger zero-shot 8B LLaMA model, with an uplift of 7.43 pp in execution accuracy (from 81.43 percent on the 8B model to 88.86 percent on the 3B model), and nearly matches its semantic accuracy (59.14 percent vs. 68.57 percent) -- all using just 3,000 reinforcement learning training examples. These results demonstrate how we can improve the performance of Text-to-SQL systems with contrastive rewards for directed semantic alignment, without requiring large-scale training datasets.

Tool-augmented language-model agents execute multi-step workflows over external systems, resolving an entity once and then acting on it across subsequent steps. Prior work shows that in single-step actions, agents select the correct tool but bind it to the wrong entity 24-26% of the time. We study what happens to entity bindings over time: do they stay correct, silently drift to a different entity, or, if wrong from the start, propagate and compound? We formalize binding drift (correct at step 1, wrong later) as distinct from error propagation (wrong at step 1, carried forward), and score them on disjoint workflow sets so the two cannot be conflated. In a controlled multi-step testbed (200 workflows, 580 entity-binding-scored steps, four enterprise domains, eight model backends spanning small to frontier), we find: (1) under controlled error injection, an entity lock (the intuitive "persist the first binding" fix) amplifies wrong actions from 907 to 2,746 (3.0x; bootstrap 95% CI [2.8, 3.3]), because it faithfully carries the seeded wrong entity into every later step; (2) the amplification reaches 8.5x on the most affected model (Claude Opus 4.5); (3) a practical LLM-based re-verifier (a single cheap second model call re-reading the original instruction) reduces wrong actions by 79% (0.21x; CI [0.18, 0.25]), closing the gap to within 1 percentage point of an oracle upper-bound (0.20x); and (4) in the natural (non-injected) setting, baseline agents drift on 18% of eligible workflows, with the per-step error rate rising across steps. Persistence and re-verification are not interchangeable: a defense that eliminates drift can worsen propagation, and a practical re-verifier nearly matches oracle recovery.

Translating natural language to SQL (Text-to-SQL) is a critical challenge in both database research and data analytics applications. Recent efforts have focused on enhancing SQL reasoning by developing large language models and AI agents that decompose Text-to-SQL tasks into manually designed, step-by-step pipelines. However, despite these extensive architectural engineering efforts, a significant gap remains: even state-of-the-art (SOTA) AI agents have not yet achieved the human-level accuracy on the BIRD benchmark. In this paper, we show that closing this gap does not require further architectural complexity, but rather clean training data to improve SQL reasoning of the underlying models. We introduce ReViSQL, a streamlined framework that achieves human-level accuracy on BIRD for the first time. Instead of complex AI agents, ReViSQL leverages reinforcement learning with verifiable rewards (RLVR) on BIRD-Verified, a dataset we curated comprising 2.5k verified Text-to-SQL instances based on the BIRD Train set. To construct BIRD-Verified, we design a data correction and verification workflow involving SQL experts. We identified and corrected data errors in 61.1% of a subset of BIRD Train. By training on BIRD-Verified, we show that improving data quality alone boosts the single-generation accuracy by 8.2-13.9% under the same RLVR algorithm. To further enhance performance, ReViSQL performs inference-time scaling via execution-based reconciliation and majority voting. Empirically, we demonstrate the superiority of our framework with two model scales: ReViSQL-235B-A22B and ReViSQL-30B-A3B. On an expert-verified BIRD Mini-Dev set, ReViSQL-235B-A22B achieves 93.2% execution accuracy, exceeding the proxy human-level accuracy (92.96%) and outperforming the prior open-source SOTA method by 9.8%. Our lightweight ReViSQL-30B-A3B matches the prior SOTA at a 7.5times lower per-query cost.

LLM-as-a-Judge (LLMaaJ) now underpins scalable evaluation, yet we lack a decisive test of a judge's qualification: can it recover a conversation's latent objective and know when that inference is trustworthy? LLMs degrade under irrelevant or long context; multi-turn jailbreaks further hide goals across turns. We introduce ObjexMT, a benchmark for objective extraction and metacognition. Given a multi-turn transcript, a model must return a one-sentence base objective and a self-reported confidence. Accuracy is computed via LLM-judge semantic similarity to gold objectives, converted to binary correctness by a single human-aligned threshold calibrated once on N = 100 items (tau^*=0.61). Metacognition is evaluated with ECE, Brier, Wrong-at-High-Conf, and risk-coverage. Across gpt-4.1, claude-sonnet-4, and Qwen3-235B-A22B-FP8 on SafeMTData_Attack600, SafeMTData_1K, MHJ, and CoSafe, claude-sonnet-4 attains the best objective-extraction accuracy (0.515) and calibration (ECE 0.296; Brier 0.324); gpt-4.1 and Qwen3-235B-A22B-FP8 tie at 0.441 but are overconfident (mean confidence approx0.88 vs. accuracy approx0.44; Wrong-at-0.90 approx48-52%). Performance varies by dataset (approx0.167-0.865). ObjexMT thus supplies an actionable test for LLM judges: when objectives are not explicit, judges often misinfer them with high confidence. We recommend exposing objectives when feasible and gating decisions by confidence otherwise. Code and data at https://github.com/hyunjun1121/ObjexMT_dataset.

Large Language Models (LLMs) have made significant strides in mathematical reasoning, underscoring the need for a comprehensive and fair evaluation of their capabilities. However, existing benchmarks often fall short, either lacking extensive coverage of undergraduate-level mathematical problems or probably suffering from test-set contamination. To address these issues, we introduce UGMathBench, a diverse and dynamic benchmark specifically designed for evaluating undergraduate-level mathematical reasoning with LLMs. UGMathBench comprises 5,062 problems across 16 subjects and 111 topics, featuring 10 distinct answer types. Each problem includes three randomized versions, with additional versions planned for release as leading open-source LLMs become saturated in UGMathBench. Furthermore, we propose two key metrics: effective accuracy (EAcc), which measures the percentage of correctly solved problems across all three versions, and reasoning gap (Delta), which assesses reasoning robustness by calculating the difference between the average accuracy across all versions and EAcc. Our extensive evaluation of 23 leading LLMs reveals that the highest EAcc achieved is 56.3\% by OpenAI-o1-mini, with large Delta values observed across different models. This highlights the need for future research aimed at developing "large reasoning models" with high EAcc and Delta = 0. We anticipate that the release of UGMathBench, along with its detailed evaluation codes, will serve as a valuable resource to advance the development of LLMs in solving mathematical problems.

Publishing a large language model (LLM) benchmark on the Internet risks contaminating future LLMs: the benchmark may be unintentionally (or intentionally) used to train or select a model. A common mitigation is to keep the benchmark private and let participants submit their models or predictions to the organizers. However, this strategy will require trust in a single organization and still permits test-set overfitting through repeated queries. To overcome this issue, we propose a way to publish benchmarks without completely disclosing the ground-truth answers to the questions, while still maintaining the ability to openly evaluate LLMs. The main underlying idea is to reduces the best possible accuracy, i.e., Bayes accuracy, by injecting randomness to the answers by preparing several logically correct answers, and only include one of them as the solution in the benchmark. Not only is this helpful to keep us from disclosing the ground truth, but this also offers a test for detecting data contamination. In principle, even fully capable models should not surpass the Bayes accuracy. If a model surpasses this ceiling despite this expectation, this is a strong signal of data contamination. We present experimental evidence that our method can detect data contamination accurately on a wide range of benchmarks, models, and training methodologies.

Noise plagues many numerical datasets, where the recorded values in the data may fail to match the true underlying values due to reasons including: erroneous sensors, data entry/processing mistakes, or imperfect human estimates. We consider general regression settings with covariates and a potentially corrupted response whose observed values may contain errors. By accounting for various uncertainties, we introduced veracity scores that distinguish between genuine errors and natural data fluctuations, conditioned on the available covariate information in the dataset. We propose a simple yet efficient filtering procedure for eliminating potential errors, and establish theoretical guarantees for our method. We also contribute a new error detection benchmark involving 5 regression datasets with real-world numerical errors (for which the true values are also known). In this benchmark and additional simulation studies, our method identifies incorrect values with better precision/recall than other approaches.

Researchers have proposed numerous text-to-SQL techniques to streamline data analytics and accelerate the development of data-driven applications. To compare these techniques and select the best one for deployment, the community depends on public benchmarks and their leaderboards. Since these benchmarks heavily rely on human annotations during question construction and answer evaluation, the validity of the annotations is crucial. In this paper, we conduct an empirical study that (i) benchmarks annotation error rates for two widely used text-to-SQL benchmarks, BIRD and Spider 2.0-Snow, and (ii) corrects a subset of the BIRD development (Dev) set to measure the impact of annotation errors on text-to-SQL agent performance and leaderboard rankings. Through expert analysis, we show that BIRD Mini-Dev and Spider 2.0-Snow have error rates of 52.8% and 62.8%, respectively. We re-evaluate all 16 open-source agents from the BIRD leaderboard on both the original and the corrected BIRD Dev subsets. We show that performance changes range from -7% to 31% (in relative terms) and rank changes range from -9 to +9 positions. We further assess whether these impacts generalize to the full BIRD Dev set. We find that the rankings of agents on the uncorrected subset correlate strongly with those on the full Dev set (Spearman's r_s=0.85, p=3.26e-5), whereas they correlate weakly with those on the corrected subset (Spearman's r_s=0.32, p=0.23). These findings show that annotation errors can significantly distort reported performance and rankings, potentially misguiding research directions or deployment choices. Our code and data are available at https://github.com/uiuc-kang-lab/text_to_sql_benchmarks.

Machine learning has demonstrated remarkable performance over finite datasets, yet whether the scores over the fixed benchmarks can sufficiently indicate the model's performance in the real world is still in discussion. In reality, an ideal robust model will probably behave similarly to the oracle (e.g., the human users), thus a good evaluation protocol is probably to evaluate the models' behaviors in comparison to the oracle. In this paper, we introduce a new robustness measurement that directly measures the image classification model's performance compared with a surrogate oracle (i.e., a foundation model). Besides, we design a simple method that can accomplish the evaluation beyond the scope of the benchmarks. Our method extends the image datasets with new samples that are sufficiently perturbed to be distinct from the ones in the original sets, but are still bounded within the same image-label structure the original test image represents, constrained by a foundation model pretrained with a large amount of samples. As a result, our new method will offer us a new way to evaluate the models' robustness performance, free of limitations of fixed benchmarks or constrained perturbations, although scoped by the power of the oracle. In addition to the evaluation results, we also leverage our generated data to understand the behaviors of the model and our new evaluation strategies.

The appearance of Jev marked the era of System One Models, foundation models that return structured decisions with probability distributions rather than text. Aside from low cost and great speed, Jev's central promise is that these probabilities are calibrated: such claim is not backed by any public test and available external benchmarks evaluate confidence calibration, not whether every returned option probability has the right numerical meaning. To tackle this issue, we introduce Sys1Cal-v1, a dataset of True/False questions about a proposition A for which the exact probability P(A) is known by construction. Each item is queried through the three Jev primitives - Noul, Choice and Score - and evaluated by total variation distance from the ground-truth distribution, which can be used to estimate a soft accuracy of System One Models. We showcase the utility of Sys1Cal-v1 as a benchmark dataset by evaluating Jev and SemIf, an open-source Choice-style baseline. In this work, however, we focus even more deeply on Jev, by studying the calibration of its Score and Choice answers. In particular, we discover a peculiar behaviour that can be explained by assuming that Jev suppresses a third truth value, going beyond True and False. In other words, in Choice answers, P(A) and P(neg A) are presented as if P(A)+P(neg A)=1, while a term P(U)neq0 is missing in the sum. Recovering P(U) leads to an improvement of median soft accuracy in Choice answers from 0.771 to 0.978, suggesting that, even in binary decisions, Jev wants to answer with a third option:``I don't know''.

Visual language model (VLM) agents show great promise in automating end-to-end (E2E) web testing against requirements in natural language. However, the probabilistic nature of language models can have inherent hallucinations. Therefore, given a detected inconsistency between the requirement and the web application, it is hard to distinguish whether it stems from the hallucination or a real application bug. Addressing this issue presents two core technical challenges: the implicit oracle inference challenge, where the agent must act as its own oracle to implicitly decide if the application's behavior is correct without guidance, and the probabilistic inference challenge, where an LLM's inconsistent reasoning undermines its trustworthiness as an oracle. Existing LLM-based approaches fail to capture such implicit oracles, either by treating any page navigation that doesn't crash as a success, or by checking each state in isolation, thus missing bugs dependent on context from prior steps. We introduce WebTestPilot, an LLM-based agent designed to address these challenges. WebTestPilot uses (1) a symbolization layer which detects and symbolizes critical GUI elements on the web application into symbols (i.e., variables) and (2) translates natural language specification into a sequence of steps, each of which is equipped with inferred pre- and post-conditions over the symbols as an oracle. This oracle captures data, temporal, and causal dependencies, enabling the validation of implicit requirements. To advance research in this area, we build a benchmark of bug-injected web apps for evaluating NL-to-E2E testing. The results show that WebTestPilot achieves a task completion rate of 99%, with 96% precision and 96% recall in bug detection, outperforming the best baseline (+70 precision, +27 recall). The agent generalizes across diverse natural language inputs and model scales.

When should an AI system's answer be trusted? Formal proof assistants offer certainty but cannot reach most of the problem distribution; scalar LLM judges offer coverage but produce opaque scores that cannot be audited after the fact and are subject to the same coherence issues as any LLM. We present Theoria, a verification architecture that closes this gap. A candidate solution is rewritten into a sequence of typed state transitions, each licensed by an explicit justification, whether that be a citation, computation, or problem-given fact, and every transition is independently auditable. The foundational invariant is completeness of change: every difference between consecutive proof states must be accounted for, so hidden premises surface as unlicensed mutations rather than passing silently. On HLE-Verified Gold (185 text-only expert problems), Theoria certifies 105 at 91.4% strict precision (Wilson 95% CI [84.5%, 95.4%]). Every certification produces a human readable proof trace in which each step can be independently challenged. Holistic LLM judges achieve comparable precision at matched coverage but fail on different problems (Jaccard 0.14-0.36), making the approaches complementary. On 95 adversarial poisoned proofs across 15 domains, structured judges catch 94.7% versus 83.2% for holistic judging (p= 0.0017). The overall 11.5 pp gap concentrates in hidden premises (90.6% vs. 62.5%, a 28 pp difference) and fabricated citations (100% vs. 90%), the error classes where the formal analysis predicts an advantage; performance is identical on arithmetic and theorem-misapplication errors, where no advantage is predicted. On GPQA Diamond (n= 65), certified precision is 97.1% (Wilson CI [85.1%, 99.5%]).

Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards. Recent work agrees these checkers are weak and patches them by hand---extra input distributions, fuzzing recipes, tighter tolerances---with no way to measure whether any patch suffices. We introduce mutation analysis as an adequacy metric for kernel-benchmark oracles: deterministic rules inject 10{,}303 compilable faults into verified CUDA implementations of 188 KernelBench problems, 7{,}384 of them with an independent kill witness; any test protocol is scored by the fraction it detects. The official check misses one in six witnessed faults (16.9%), deterministically, and the misses are skewed by family: 8.7% of arithmetic faults escape, but 78.6% of precision faults do. The metric explains why (a tolerance blind band growing with reduction size; a measured ceiling on input aggressiveness set by legitimate floating-point variance), audits the strongest existing patch (KernelBench-Verified's gain splits into +4.0 points from hidden inputs and +4.5 from tighter tolerance, a split its authors could not compute), and exposes a published fuzzing recipe that rejects correct kernels 107 times. Optimizing suites over the kill matrix reaches 98.0% detection with two inputs per problem (94.8% held-out), and the measurement's fault taxonomy teaches a test generator more than the raw faults themselves. Across 48 whole architectures, the blindness grows with scale, concentrating in deep homogeneous pipelines, and two problems prove unrefereeable: their official references violate the benchmark's own tolerance against fp64. We release everything as https://huggingface.co/datasets/Elfsong/KernelBench-M{KernelBench-M}.

We introduce SciTaRC, an expert-authored benchmark for question answering over scientific tables that targets composite, multi-step reasoning. To enable fine-grained diagnostic analysis beyond end-task accuracy, SciTaRC pairs each question with a manually constructed reasoning plan and explicit complexity metrics. State-of-the-art models fail on at least 23% of these questions, while highly capable open-weight models like Llama-3.3-70B collapse on 65.5% of the benchmark. Error analysis shows that, in zero-shot settings, failures are driven primarily by question comprehension, where models misinterpret the scientific query and derive the wrong reasoning objective. To determine whether overcoming this gap is sufficient, we use the structured plans to decouple strategy formulation from execution. Surprisingly, providing oracle step-by-step plans yields only limited gains and fails to eliminate the performance gap. This reveals a substantial execution bottleneck: both natural language and code-based methods struggle to reliably carry out long-horizon computational chains over structured data. Ultimately, SciTaRC serves as a rigorous diagnostic testbed for studying both planning and execution in scientific table reasoning.

Developers increasingly delegate real maintenance work to product-grade coding agents, and many state tasks in their native language, in the style of a customer request rather than a curated English issue. We introduce RuBench 1.0, a benchmark of 25 tasks mined from recent fix commits in five live open-source repositories (aiohttp, aiogram, Laravel, NestJS, Fastify), each specified natively in Russian -- written from scratch, not translated -- and judged by the upstream maintainer's regression tests, which we withhold from release. All fix commits postdate the training-data cutoffs of every evaluated model. Round 1 evaluates Claude Code with Opus 4.8, Sonnet 5, and Haiku 4.5, and Codex CLI with GPT-5.5 (3 independent runs each; pass@1 with task-level uncertainty); the best configuration resolves 78.7% of tasks. Auditing full trajectories of an hors-concours configuration (Claude Code + Fable 5), we caught the product silently substituting the model on 20% of tasks via an official safeguard fallback -- evidence that the deployed product, not the model, is the unit actually measured. Version 2 adds Round 2: seven further configurations on the same frozen set under a per-configuration freshness gate -- the Russian-market agents SourceCraft CLI (ds, legacy) and Koda CLI (koda-pro), Antigravity with Gemini 3.1 Pro and 3.5 Flash, and Codex CLI with GPT-5.6 Sol and Luna. SourceCraft's flagship resolves 68.1% (N=23), above GPT-5.5 and both Gemini rows. A tool-call contamination re-audit of all 437 Round-2 trajectories finds the Russian and Gemini columns clean (0/293 cells) while flagging systematic oracle-hunting in the GPT-5.6 family (8/69 and 13/75 cells), including one case of mining a prior round's artifacts from the run machine's disk; honest scores are published alongside raw ones. We release statements, metadata, trajectories, and diffs; oracles are withheld with a SHA-256 manifest.

Large language models (LLMs) have demonstrated strong performance in translating natural language questions into SQL queries (Text-to-SQL). In contrast, small language models (SLMs) ranging from 0.5B to 1.5B parameters currently underperform on Text-to-SQL tasks due to their limited logical reasoning capabilities. However, SLMs offer inherent advantages in inference speed and suitability for edge deployment. To explore their potential in Text-to-SQL applications, we leverage recent advancements in post-training techniques. Specifically, we used the open-source SynSQL-2.5M dataset to construct two derived datasets: SynSQL-Think-916K for SQL generation and SynSQL-Merge-Think-310K for SQL merge revision. We then applied supervised fine-tuning and reinforcement learning-based post-training to the SLM, followed by inference using a corrective self-consistency approach. Experimental results validate the effectiveness and generalizability of our method, SLM-SQL. On the BIRD development set, the five evaluated models achieved an average improvement of 31.4 points. Notably, the 0.5B model reached 56.87\% execution accuracy (EX), while the 1.5B model achieved 67.08\% EX. We will release our dataset, model, and code to github: https://github.com/CycloneBoy/slm_sql.

Making language models solve constraint problems reliably often means having them translate the problem into a formal specification and delegating the search to a sound solver. But the translation is itself a language-model task, and an unfaithful translation makes the solver faithfully solve the wrong problem. Existing pipelines repair only translations that crash, returning the solver's error message and falling silent when the program runs but is wrong. We replace the error message with a proof: when the generated program is unsatisfiable, we extract a minimal unsatisfiable core over the model's own constraints and hand it back the exact set that cannot hold together, a leakage-free signal that localizes the fault. On a new benchmark of 77 problems with an exact oracle, translation to Answer Set Programming is faithful on six of seven domains and fails only on aggregate coverage scheduling, which concentrates the translation tax in one diagnosable pattern. A minimal core, rather than a bare error, is what stops a weaker model from fabricating solutions to infeasible problems, cutting fabrication from 79% to 7%. A strong chain-of-thought baseline meanwhile matches the symbolic route on accuracy, so the route's value is not accuracy but certificates and its refusal to fabricate.

We present a hardware-oriented description of GoldenFloat (GF), a static-split floating-point family generated by a single closed rule, and three concrete artefacts: (i) an open multi-width RTL generator covering GF4-GF256 with a continuous-integration differential sweep against a correctly-rounded reference; (ii) an integer-backed Lucas-exact accumulator path verified at 500-digit precision for n = 1, ..., 256; and (iii) a GF16 FPGA codec passing a 35-of-35 testbench at 323 MHz on Artix-7 (Xilinx XC7A35T). For each total width N >= 4, the exponent width is e = round((N-1)/phi^2) with fraction f = N-1-e and phi = (1+sqrt(5))/2. The rule reproduces the realised exponent widths of nine formats (9/9) and extends consistently to GF128, GF512, GF1024. The rule is positioned alongside posit, takum, OCP-MX, and the IEEE P3109 multi-width float draft. We make no per-rung accuracy or superiority claim against any of them. The breadth/toolchain-coherence framing is recorded as an open conjecture with a pre-registered falsification path. A falsification ledger (FL-002) records open questions and the experiments that would settle them. An RTL-correctness erratum dated 2026-05-31 is reported; the fabricated TTSKY26b dies carry the defective multiplier portfolio, and the corrected generator is the regeneration baseline.

Evaluating ASR systems for Indian languages is challenging due to spelling variations, suffix splitting flexibility, and non-standard spellings in code-mixed words. Traditional Word Error Rate (WER) often presents a bleaker picture of system performance than what human users perceive. Better aligning evaluation with real-world performance requires capturing permissible orthographic variations, which is extremely challenging for under-resourced Indian languages. Leveraging recent advances in LLMs, we propose a framework for creating benchmarks that capture permissible variations. Through extensive experiments, we demonstrate that OIWER, by accounting for orthographic variations, reduces pessimistic error rates (an average improvement of 6.3 points), narrows inflated model gaps (e.g., Gemini-Canary performance difference drops from 18.1 to 11.5 points), and aligns more closely with human perception than prior methods like WER-SN by 4.9 points.

This study investigates various approaches to using Large Language Models (LLMs) for Text-to-SQL program synthesis, focusing on the outcomes and insights derived. Employing the popular Text-to-SQL dataset, spider, the goal was to input a natural language question along with the database schema and output the correct SQL SELECT query. The initial approach was to fine-tune a local and open-source model to generate the SELECT query. After QLoRa fine-tuning WizardLM's WizardCoder-15B model on the spider dataset, the execution accuracy for generated queries rose to a high of 61%. With the second approach, using the fine-tuned gpt-3.5-turbo-16k (Few-shot) + gpt-4-turbo (Zero-shot error correction), the execution accuracy reached a high of 82.1%. Of all the incorrect queries, most can be categorized into a seven different categories of what went wrong: selecting the wrong columns or wrong order of columns, grouping by the wrong column, predicting the wrong values in conditionals, using different aggregates than the ground truth, extra or too few JOIN clauses, inconsistencies in the Spider dataset, and lastly completely incorrect query structure. Most if not all of the queries fall into these categories and it is insightful to understanding where the faults still lie with LLM program synthesis and where they can be improved.

Large language models (LLMs) have demonstrated strong capabilities in translating natural language questions about relational databases into SQL queries. In particular, test-time scaling techniques such as Self-Consistency and Self-Correction can enhance SQL generation accuracy by increasing computational effort during inference. However, these methods have notable limitations: Self-Consistency may select suboptimal outputs despite majority votes, while Self-Correction typically addresses only syntactic errors. To leverage the strengths of both approaches, we propose CSC-SQL, a novel method that integrates Self-Consistency and Self-Correction. CSC-SQL selects the two most frequently occurring outputs from parallel sampling and feeds them into a merge revision model for correction. Additionally, we employ the Group Relative Policy Optimization (GRPO) algorithm to fine-tune both the SQL generation and revision models via reinforcement learning, significantly enhancing output quality. Experimental results confirm the effectiveness and generalizability of CSC-SQL. On the BIRD development set, our 3B model achieves 65.28% execution accuracy, while the 7B model achieves 69.19%. The code will be open sourced at https://github.com/CycloneBoy/csc_sql.

Text-to-SQL benchmarks play a crucial role in evaluating the progress made in the field and the ranking of different models. However, accurately matching a model-generated SQL query to a reference SQL query in a benchmark fails for various reasons, such as underspecified natural language queries, inherent assumptions in both model-generated and reference queries, and the non-deterministic nature of SQL output under certain conditions. In this paper, we conduct an extensive study of several prominent cross-domain text-to-SQL benchmarks and re-evaluate some of the top-performing models within these benchmarks, by both manually evaluating the SQL queries and rewriting them in equivalent expressions. Our evaluation reveals that attaining a perfect performance on these benchmarks is unfeasible due to the multiple interpretations that can be derived from the provided samples. Furthermore, we find that the true performance of the models is underestimated and their relative performance changes after a re-evaluation. Most notably, our evaluation reveals a surprising discovery: a recent GPT4-based model surpasses the gold standard reference queries in the Spider benchmark in our human evaluation. This finding highlights the importance of interpreting benchmark evaluations cautiously, while also acknowledging the critical role of additional independent evaluations in driving advancements in the field.

A retrieval-augmented QA system can return different answers after an index expansion even when its requested model identifier, prompt, retrieval policy, evidence depth, rendering, and exposed generation controls are held fixed. Aggregate accuracy may hide these changes when gains and losses cancel, while ordinary generation variability makes one-shot comparisons overstate update effects. We call the hidden phenomenon accuracy-blind answer churn and introduce the Snapshot Compatibility Audit, which estimates excess answer churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement. We instantiate it by expanding one frozen FineWeb prefix from one to seven shards. In a preregistered 400-question Natural Questions study, normalized-exact and blinded-semantic excess churn are 6.44 and 10.25 percentage points while exact-match accuracy changes by only -1.50 points. A post-hoc analysis finds repeat-stable semantic flips on 40/400 questions. A separately preregistered 200-question TriviaQA study yields smaller, directionally consistent excess churn while exact-match accuracy moves in the opposite direction. An outcome-blind post-hoc 100-question subset replication with a second DeepSeek generator and serving configuration finds 8.75 pp of semantic excess churn even as exact match rises by 3.00 percentage points. Answer-level compatibility can therefore fail without a conspicuous or consistently directed utility shift. Retrieval-augmented releases should audit compatibility alongside utility.

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

Text-to-SQL, the task of translating natural language questions into SQL queries, is part of various business processes. Its automation, which is an emerging challenge, will empower software practitioners to seamlessly interact with relational databases using natural language, thereby bridging the gap between business needs and software capabilities. In this paper, we consider Large Language Models (LLMs), which have achieved state of the art for various NLP tasks. Specifically, we benchmark Text-to-SQL performance, the evaluation methodologies, as well as input optimization (e.g., prompting). In light of the empirical observations that we have made, we propose two novel metrics that were designed to adequately measure the similarity between SQL queries. Overall, we share with the community various findings, notably on how to select the right LLM on Text-to-SQL tasks. We further demonstrate that a tree-based edit distance constitutes a reliable metric for assessing the similarity between generated SQL queries and the oracle for benchmarking Text2SQL approaches. This metric is important as it relieves researchers from the need to perform computationally expensive experiments such as executing generated queries as done in prior works. Our work implements financial domain use cases and, therefore contributes to the advancement of Text2SQL systems and their practical adoption in this domain.

Evaluating large language models (LLMs) typically requires thousands of benchmark items, making the process expensive, slow, and increasingly impractical at scale. Existing evaluation protocols rely on average accuracy over fixed item sets, treating all items as equally informative despite substantial variation in difficulty and discrimination. We introduce ATLAS, an adaptive testing framework based on Item Response Theory (IRT) that estimates model ability using Fisher information-guided item selection. ATLAS reduces the number of required items by up to 90% while maintaining measurement precision. For instance, it matches whole-bank ability estimates using only 41 items (0.157 MAE) on HellaSwag (5,600 items). We further reconstruct accuracy from ATLAS's ability estimates and find that reconstructed accuracies closely match raw accuracies across all five benchmarks, indicating that ability θ preserves the global performance structure. At the same time, θ provides finer discrimination within accuracy-equivalent models: among more than 3,000 evaluated models, 23-31% shift by more than 10 rank positions, and models with identical accuracies receive meaningfully different ability estimates. Code and calibrated item banks are available at https://github.com/Peiyu-Georgia-Li/ATLAS.git.

While large language models (LLMs) perform well on table tasks, they still make data referencing errors (DREs), i.e., incorrectly citing or omitting table values, despite understanding the table structure. Beyond final-answer accuracy, DREs directly compromise the correctness and reliability of intermediate reasoning steps. Yet prior studies have only offered limited, small-scale analyses. In this work, we present the first systematic evaluation of tabular data referencing errors across different models and tasks. Our results show that DREs occur across all tested models (1.7B to 20B parameters). Furthermore, we demonstrate that incorporating data referencing as a critic significantly improves answer accuracy up to 12.0%, through critic-based filtering and rejection sampling. Finally, we trained a lightweight 4B-parameter critic model that achieves an average F1 score of 78.2% in detecting both in-distribution and out-of-distribution DREs, and effectively assists inference for larger models.

Training large language models (LLMs) with synthetic reasoning data has become a popular approach to enhancing their reasoning capabilities, while a key factor influencing the effectiveness of this paradigm is the quality of the generated multi-step reasoning data. To generate high-quality reasoning data, many recent methods generate synthetic reasoning paths and filter them based on final answer correctness, often overlooking flaws in intermediate reasoning steps. To enhance the verification of intermediate reasoning steps, prior work primarily resorts to code execution or symbolic reasoning engines. However, code-based validation is restricted to code or mathematical tasks, and reasoning engines require a well-structured and complete context. As a result, existing methods fail to function effectively in natural language reasoning tasks that involve ambiguous or incomplete contexts. In these tasks, synthetic data still lack reliable checks for verifying each reasoning step. To address this challenge, we introduce ORACLE, a structured data generation framework inspired by syllogistic reasoning. ORACLE integrates the generative strengths of LLMs with symbolic supervision: the LLM produces step-wise reasoning contexts, while a symbolic reasoning engine verifies the validity of each intermediate step. By employing a unified prompting template to elicit modular reasoning chains, ORACLE enables fine-grained, step-level validation, facilitating the construction of high-quality multi-step reasoning data. Across six logical, factual, and commonsense reasoning benchmarks, our ORACLE consistently outperforms strong baselines on multiple models.

Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit "think in English" is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.

Large language models (LLMS) have shown increasing effectiveness in Text-to-SQL tasks. However, another closely related problem, Cross-System SQL Translation (a.k.a., SQL-to-SQL), which adapts a query written for one database system (e.g., MySQL) into its equivalent one for another system (e.g., ClickHouse), is of great practical importance but remains underexplored. Existing SQL benchmarks are not well-suited for SQL-to-SQL evaluation, which (1) focus on a limited set of database systems (often just SQLite) and (2) cannot capture many system-specific SQL dialects (e.g., customized functions, data types, and syntax rules). Thus, in this paper, we introduce PARROT, a Practical And Realistic BenchmaRk for CrOss-System SQL Translation. PARROT comprises 598 translation pairs from 38 open-source benchmarks and real-world business services, specifically prepared to challenge system-specific SQL understanding (e.g., LLMS achieve lower than 38.53% accuracy on average). We also provide multiple benchmark variants, including PARROT-Diverse with 28,003 translations (for extensive syntax testing) and PARROT-Simple with 5,306 representative samples (for focused stress testing), covering 22 production-grade database systems. To promote future research, we release a public leaderboard and source code at: https://code4db.github.io/parrot-bench/.

As LLM-based agents are deployed in production systems, understanding their behavioral consistency (whether they produce similar action sequences when given identical tasks) becomes critical for reliability. We study consistency in the context of SWE-bench, a challenging software engineering benchmark requiring complex, multi-step reasoning. Comparing Claude~4.5~Sonnet, GPT-5, and Llama-3.1-70B across 50 runs each (10 tasks times 5 runs), we find that across models, higher consistency aligns with higher accuracy: Claude achieves the lowest variance (CV: 15.2\%) and highest accuracy (58\%), GPT-5 is intermediate (CV: 32.2\%, accuracy: 32\%), and Llama shows the highest variance (CV: 47.0\%) with lowest accuracy (4\%). However, within a model, consistency can amplify both correct and incorrect interpretations. Our analysis reveals a critical nuance: consistency amplifies outcomes rather than guaranteeing correctness. 71\% of Claude's failures stem from "consistent wrong interpretation": making the same incorrect assumption across all runs. Interestingly, GPT-5 achieves similar early strategic agreement as Claude (diverging at step 3.4 vs.\ 3.2) but exhibits 2.1times higher variance, suggesting that divergence timing alone does not determine consistency. These findings suggest that for production deployment, interpretation accuracy matters more than execution consistency, with implications for agent evaluation and training.

Operations Research practitioners debug infeasible models through an iterative process: inspecting Irreducible Infeasible Subsystems ( IIS), identifying constraint conflicts, and repairing formulations until feasibility is restored. Existing LLM benchmarks mostly treat OR as one-shot translation from problem descriptions to solver code, omitting this diagnostic loop. We formalize infeasible-model repair as a solver-in-the-loop Markov Decision Process in which each action triggers solver re-execution and IIS recomputation, yielding deterministic, verifiable feedback. We introduce ORLoopBench, a benchmark suite with two components: OR-Debug-Bench releases 5,362 LP/MILP repair instances, while OR-Bias-Bench evaluates closed-form operational decision rationality across inventory settings. Solver-verified RLVR training enables an 8B model to surpass frontier APIs on LP repair (95.3% vs 92.4% RR @5), improves diagnostic behavior, and transfers to MILP repair. The same evaluation exposes semantic drift in whole-model code regeneration: feasible regenerated MILPs can solve the wrong problem. Process-level evaluation with solver oracles enables targeted training for reliable OR self-correction.

We introduce the Oracle-MNIST dataset, comprising of 28times 28 grayscale images of 30,222 ancient characters from 10 categories, for benchmarking pattern classification, with particular challenges on image noise and distortion. The training set totally consists of 27,222 images, and the test set contains 300 images per class. Oracle-MNIST shares the same data format with the original MNIST dataset, allowing for direct compatibility with all existing classifiers and systems, but it constitutes a more challenging classification task than MNIST. The images of ancient characters suffer from 1) extremely serious and unique noises caused by three-thousand years of burial and aging and 2) dramatically variant writing styles by ancient Chinese, which all make them realistic for machine learning research. The dataset is freely available at https://github.com/wm-bupt/oracle-mnist.

In this work, we explore uncertainty estimation as a proxy for correctness in LLM-generated code. To this end, we adapt two state-of-the-art techniques from natural language generation -- one based on entropy and another on mutual information -- to the domain of code generation. Given the distinct semantic properties of code, we introduce modifications, including a semantic equivalence check based on symbolic execution. Our findings indicate a strong correlation between the uncertainty computed through these techniques and correctness, highlighting the potential of uncertainty estimation for quality assessment. Additionally, we propose a simplified version of the entropy-based method that assumes a uniform distribution over the LLM's responses, demonstrating comparable effectiveness. Using these techniques, we develop an abstention policy that prevents the model from making predictions when uncertainty is high, reducing incorrect outputs to near zero. Our evaluation on the LiveCodeBench shows that our approach significantly outperforms a baseline relying solely on LLM-reported log-probabilities.

We investigate large language model performance across five orders of magnitude of compute scaling in eleven recent model architectures. We show that average benchmark performance, aggregating over many individual tasks and evaluations as in the commonly-used BIG-Bench dataset, is decently predictable as a function of training compute scale. Specifically, when extrapolating BIG-Bench Hard performance across one order of magnitude in compute, we observe average absolute errors of 6 percentage points (pp). By contrast, extrapolation for individual BIG-Bench tasks across an order of magnitude in compute yields higher average errors of 18pp. Nonetheless, individual task performance remains significantly more predictable than chance. Overall, our work suggests compute scaling provides a promising basis to forecast AI capabilities in diverse benchmarks, though predicting performance in specific tasks poses challenges.

Many real-world questions over business, legal, and scientific corpora are natural-language versions of database-style queries over records latent in text. Existing retrieval-augmented generation (RAG) systems are optimized primarily for semantic relevance, but retrieving plausible passages does not guarantee correct query execution. We introduce QO-Bench, a diagnostic benchmark for query-operator question answering over typed event tuples. The benchmark covers 22,984 news articles and 614 corporate events across 18 query templates, evaluated on 785 questions. Each gold answer is deterministically computed from typed event tuples and scored by recall, with answers matched to the gold tuples by exact match rather than an LLM judge. This design enables operator-level diagnosis such as joins and intersection. We evaluate RAG, ReAct RAG, GraphRAG, and information-extraction-to-SQL under matched conditions, with a long-context oracle ceiling to isolate retrieval failure. A two-axis framework -- index-time preservation versus query-time execution -- predicts where each paradigm fails, and the results bear it out: systems retrieve relevant text but discard the typed values operators need, and the deployable paradigm ranking inverts across operators, with similarity retrieval leading on filter/project and extraction-to-SQL on intersection and counting. Even given the gold evidence, a long-context oracle stays far from saturated, so operator execution -- not retrieval alone -- is a core bottleneck that a stronger answer model does not remove. QO-Bench reframes the goal from passage relevance to query-operator-preserving retrieval.

Large language models (LLMs) often present answers with high apparent confidence despite lacking an explicit mechanism for reasoning about certainty or truth. While existing benchmarks primarily evaluate single-turn accuracy, truthfulness or confidence calibration, they do not capture how models behave when their responses are challenged in interactive settings. We introduce the Certainty Robustness Benchmark, a two-turn evaluation framework that measures how LLMs balance stability and adaptability under self-challenging prompts such as uncertainty ("Are you sure?") and explicit contradiction ("You are wrong!"), alongside numeric confidence elicitation. Using 200 reasoning and mathematics questions from LiveBench, we evaluate four state-of-the-art LLMs and distinguish between justified self-corrections and unjustified answer changes. Our results reveal substantial differences in interactive reliability that are not explained by baseline accuracy alone: some models abandon correct answers under conversational pressure, while others demonstrate strong resistance to challenge and better alignment between confidence and correctness. These findings identify certainty robustness as a distinct and critical dimension of LLM evaluation, with important implications for alignment, trustworthiness and real-world deployment.

With significant efforts in recent studies, LLM-as-a-Judge has become a cost-effective alternative to human evaluation for assessing the text generation quality in a wide range of tasks. However, there still remains a reliability gap between LLM-as-a-Judge and human evaluation. One important reason is the lack of guided oracles in the evaluation process. Motivated by the role of reference pervasively used in classic text evaluation, we introduce RevisEval, a novel text generation evaluation paradigm via the response-adapted references. RevisEval is driven by the key observation that an ideal reference should maintain the necessary relevance to the response to be evaluated. Specifically, RevisEval leverages the text revision capabilities of large language models (LLMs) to adaptively revise the response, then treat the revised text as the reference (response-adapted reference) for the subsequent evaluation. Extensive experiments demonstrate that RevisEval outperforms traditional reference-free and reference-based evaluation paradigms that use LLM-as-a-Judge across NLG tasks and open-ended instruction-following tasks. More importantly, our response-adapted references can further boost the classical text metrics, e.g., BLEU and BERTScore, compared to traditional references and even rival the LLM-as-a-Judge. A detailed analysis is also conducted to confirm RevisEval's effectiveness in bias reduction, the impact of inference cost, and reference relevance.

The recent successful paradigm of solving logical reasoning problems with tool-augmented large language models (LLMs) leverages translation of natural language (NL) statements into First-Order Logic~(FOL) and external theorem provers. However, the correctness of FOL statements, comprising operators and text, often go unverified due to the lack of a reliable evaluation metric for comparing generated and ground-truth FOLs. In this paper, we conduct a comprehensive study on the sensitivity of existing NL-, FOL-, and graph-based metrics to capture differences between a sampled FOL and its corresponding ground-truth. We then measure the alignment between a metric-based ranking of FOL outputs and a strong LLM as-a-judge. To do this, we first apply operator and text-based perturbations to ground-truth FOL statements to assess metric sensitivity. We then evaluate metric robustness by comparing the metrics against LLMs judgment. Our empirical findings highlight a clear oversensitivity in the n-gram metric BLEU for text perturbations. The operator perturbation affects the semantic graph metric Smatch++ for structural changes, and the FOL metric for specific operator changes. We observe a closer alignment between BertScore and LLM judgement, proving the importance of semantic evaluation. Additionally, we show that combining metrics enhances both robustness and sensitivity compared to using individual metrics.

Text-to-SQL parsing, which aims at converting natural language instructions into executable SQLs, has gained increasing attention in recent years. In particular, Codex and ChatGPT have shown impressive results in this task. However, most of the prevalent benchmarks, i.e., Spider, and WikiSQL, focus on database schema with few rows of database contents leaving the gap between academic study and real-world applications. To mitigate this gap, we present Bird, a big benchmark for large-scale database grounded in text-to-SQL tasks, containing 12,751 pairs of text-to-SQL data and 95 databases with a total size of 33.4 GB, spanning 37 professional domains. Our emphasis on database values highlights the new challenges of dirty database contents, external knowledge between NL questions and database contents, and SQL efficiency, particularly in the context of massive databases. To solve these problems, text-to-SQL models must feature database value comprehension in addition to semantic parsing. The experimental results demonstrate the significance of database values in generating accurate text-to-SQLs for big databases. Furthermore, even the most effective text-to-SQL models, i.e. ChatGPT, only achieves 40.08% in execution accuracy, which is still far from the human result of 92.96%, proving that challenges still stand. Besides, we also provide an efficiency analysis to offer insights into generating text-to-efficient-SQLs that are beneficial to industries. We believe that BIRD will contribute to advancing real-world applications of text-to-SQL research. The leaderboard and source code are available: https://bird-bench.github.io/.

In recent years,Text-to-SQL, the problem of automatically converting questions posed in natural language to formal SQL queries, has emerged as an important problem at the intersection of natural language processing and data management research. Large language models (LLMs) have delivered impressive performance when used in an off-the-shelf performance, but still fall significantly short of expected expert-level performance. Errors are especially probable when a nuanced understanding is needed of database schemas, questions, and SQL clauses to do proper Text-to-SQL conversion. We introduce SelECT-SQL, a novel in-context learning solution that uses an algorithmic combination of chain-of-thought (CoT) prompting, self-correction, and ensemble methods to yield a new state-of-the-art result on challenging Text-to-SQL benchmarks. Specifically, when configured using GPT-3.5-Turbo as the base LLM, SelECT-SQL achieves 84.2% execution accuracy on the Spider leaderboard's development set, exceeding both the best results of other baseline GPT-3.5-Turbo-based solutions (81.1%), and the peak performance (83.5%) of the GPT-4 result reported on the leaderboard.

Text-to-SQL enables users to interact with databases using natural language, simplifying the retrieval and synthesis of information. Despite the remarkable success of large language models (LLMs) in translating natural language questions into SQL queries, widespread deployment remains limited due to two primary challenges. First, the effective use of text-to-SQL models depends on users' understanding of the model's capabilities-the scope of questions the model can correctly answer. Second, the absence of abstention mechanisms can lead to incorrect SQL generation going unnoticed, thereby undermining trust in the model's output. To enable wider deployment, it is crucial to address these challenges in model design and enhance model evaluation to build trust in the model's output. To this end, we introduce TrustSQL, a novel comprehensive benchmark designed to evaluate text-to-SQL reliability-defined as a model's ability to correctly handle any type of input question by generating correct SQL queries for feasible questions and abstaining from generating infeasible ones (e.g., due to schema incompatibility or functionalities beyond SQL). We evaluate existing methods using a novel penalty-based scoring metric with two modeling approaches: (1) pipeline-based methods combining SQL generators with infeasible question detectors and SQL error detectors for abstention; and (2) unified methods using a single model for the entire task. Our experimental results reveal that achieving high scores under severe penalties requires significant effort and provide a new perspective on developing text-to-SQL models for safer deployment. TrustSQL is available at https://github.com/glee4810/TrustSQL.

Text-to-SQL, which involves translating natural language into Structured Query Language (SQL), is crucial for enabling broad access to structured databases without expert knowledge. However, designing models for such tasks is challenging due to numerous factors, including the presence of 'noise,' such as ambiguous questions and syntactical errors. This study provides an in-depth analysis of the distribution and types of noise in the widely used BIRD-Bench benchmark and the impact of noise on models. While BIRD-Bench was created to model dirty and noisy database values, it was not created to contain noise and errors in the questions and gold queries. We found that noise in questions and gold queries are prevalent in the dataset, with varying amounts across domains, and with an uneven distribution between noise types. The presence of incorrect gold SQL queries, which then generate incorrect gold answers, has a significant impact on the benchmark's reliability. Surprisingly, when evaluating models on corrected SQL queries, zero-shot baselines surpassed the performance of state-of-the-art prompting methods. We conclude that informative noise labels and reliable benchmarks are crucial to developing new Text-to-SQL methods that can handle varying types of noise. All datasets, annotations, and code are available at https://github.com/niklaswretblad/the-effects-of-noise-in-text-to-SQL.

Search-augmented LLM agents can produce deep research reports (DRRs), but verifying claim-level factuality remains challenging. Existing fact-checkers are primarily designed for general-domain, factoid-style atomic claims, and there is no benchmark to test whether such verifiers transfer to DRRs. Yet building such a benchmark is itself difficult. We first show that static expert-labeled benchmarks are brittle in this setting: in a controlled study with PhD-level specialists, unassisted experts achieve only 60.8% accuracy on a hidden micro-gold set of verifiable claims. We propose Evolving Benchmarking via Audit-then-Score (AtS), where benchmark labels and rationales are explicitly revisable: when a verifier disagrees with the current benchmark, it must submit evidence; an auditor adjudicates the dispute; and accepted revisions update the benchmark before models are scored. Across four AtS rounds, expert micro-gold accuracy rises to 90.9%, indicating experts are substantially more reliable as auditors than as one-shot labelers. We instantiate AtS as DeepFact-Bench, a versioned DRR factuality benchmark with auditable rationales, and DeepFact-Eval, a document-level verification agent (with a grouped lite variant) that outperforms existing verifiers on DeepFact-Bench and transfers well to external factuality datasets.

Why do language models sometimes prefer correct statements even when trained on mixed-quality data? We introduce the Compression--Consistency Principle: next-token prediction favors hypotheses that allow shorter and more internally consistent descriptions of the training data. Truth bias emerges only when false alternatives are structurally harder to compress. We test this using small GPT-2-style character-level transformers (3.5M--86M parameters) on synthetic math corpora with controlled mixtures of correct and incorrect rules. In the random-error setting, models strongly prefer correct completions in paired evaluation: 83.1% accuracy at balanced data and 67.0% even when correct rules appear in only 10% of the corpus. Replacing random errors with a coherent but mathematically incorrect rule system largely eliminates the preference (near-chance accuracy). In a more natural-language-like synthetic world, the effect is weaker but still present (57.7%). Additional experiments show that embedding verification steps can restore preference for correctness even at small scale, while increasing the number of consistent rules produces a graded improvement in accuracy. Our results suggest that what appears as a "truth bias" is largely a side effect of compression pressure and preference for internal consistency, rather than an intrinsic drive toward truth. Full code and data are available at https://github.com/Rai220/compression-drives-truth.

Text-to-SQL models allow users to interact with a database more easily by generating executable SQL statements from natural-language questions. Despite recent successes on simpler databases and questions, current Text-to-SQL methods still suffer from low execution accuracy on industry-scale databases and complex questions involving domain-specific business logic. We present PaVeRL-SQL, a framework that combines Partial-Match Rewards and Verbal Reinforcement Learning to drive self-improvement in reasoning language models (RLMs) for Text-to-SQL. To handle practical use cases, we adopt two pipelines: (1) a newly designed in-context learning framework with group self-evaluation (verbal-RL), using capable open- and closed-source large language models (LLMs) as backbones; and (2) a chain-of-thought (CoT) RL pipeline with a small backbone model (OmniSQL-7B) trained with a specially designed reward function and two-stage RL. These pipelines achieve state-of-the-art (SOTA) results on popular Text-to-SQL benchmarks -- Spider, Spider 2.0, and BIRD. For the industrial-level Spider2.0-SQLite benchmark, the verbal-RL pipeline achieves an execution accuracy 7.4\% higher than SOTA, and the CoT pipeline is 1.4\% higher. RL training with mixed SQL dialects yields strong, threefold gains, particularly for dialects with limited training data. Overall, PaVeRL-SQL delivers reliable, SOTA Text-to-SQL under realistic industrial constraints. The code is available at https://github.com/PaVeRL-SQL/PaVeRL-SQL.

We present PLSQLBench, to our knowledge the first benchmark for evaluating whether LLMs can write executable PL/SQL programs, with correctness measured through execution-based tests. Existing LLM evaluations largely target general-purpose code generation or declarative text-to-SQL, leaving procedural database programming underexplored. PLSQLBench contains 2,865 instances: 2,594 single-turn tasks and 271 multi-turn conversations spanning 978 turns. The benchmark combines complex schema-grounded tasks over enterprise-style Spider 2 databases, simpler schema-grounded tasks derived from Spider, and MBPP-derived procedural problems, covering varying levels of database grounding and procedural complexity. Experiments with eight LLMs reveal recurring difficulties in schema grounding, PL/SQL dialect fidelity, procedural control flow, exception handling, and cross-turn consistency. Tool-augmented LLM agents improve performance on several schema-grounded evaluations, although substantial gaps remain. These results highlight procedural database programming capabilities not directly assessed by conventional code generation or text-to-SQL benchmarks. Our code is available at https://github.com/oracle-samples/plsqlbench.

False-presupposition QA (FPQA) tests LLMs on their ability to identify false presuppositions in questions and abstain or correct them rather than reinforcing false assumptions. The common approach reduces the task to prompting LLMs to extract presuppositions and fact checking each presupposition. While the performance on dedicated benchmarks keeps improving, evaluation largely focuses on questions with false presuppositions (FPQs) while ignoring the performance on ``normal'' questions (TPQs). Since many benchmarks over-represent FPQs compared to their natural occurrence, the result is that performance on these benchmarks doesn't reflect real-world QA performance. Through extensive experiments across various model families, sizes, and benchmarks, we show that methods that perform better on FPQs tend to perform worse on TPQs. Our analysis reveals this is the result of weak fact checking modules that reject also true presuppositions. We hope our findings will help guide future work toward FPQA methods that generalize well to realistic settings.

Accurate translation from Natural Language to First-Order Logic (NL-to-FOL) underpins neurosymbolic AI systems and Natural Language Inference (NLI), making the quality of NL-to-FOL benchmarks essential -- yet these datasets have never been rigorously audited. Our first contribution is to present a systematic human inspection of the validation split of FOLIO and a subset of MALLS test instances, finding that approximately 39% and 36% of entries, respectively, contain incorrect FOL formalizations (i.e., ground truth labels), with additional rates of ambiguous NL sentences (16.4% and 48%) and incorrect NLI labels in FOLIO (8.4%). Our second contribution is to develop and release corrected ground truths for such datasets, showing that annotation errors distort model evaluation on a reference benchmark task: testing three state-of-the-art LLMs (Gemma~4 31B-it, Qwen3-30B-A3B, and GPT-4o-mini) with the corrected ground truths yields accuracy gains from +9 to +22 percentage points. Motivated by these findings, we propose an LLM-based framework to support humans in manual reviewing NL-to-FOL datasets. By directing reviewers toward the most error-prone instances, we empirically show that it is possible to achieve 90% dataset accuracy after reviewing fewer than 24% of instances, compared to over 70% required by unguided review. We release all human-verified annotations and the code for our framework.

LLM agents increasingly modify their own prompts, tools, middleware, resources, and execution harnesses at runtime. Such self-evolution can improve capability, but a successful mutation may leave persistent effects that cannot be safely reversed in states different from the one in which it was created. We introduce EvoUndo, a framework for representing, synthesizing, diagnosing, and independently verifying recoverability of model-generated self-modifications across counterfactual states. Across 600 unseen one-shot self-evolution tasks, we identify 197 capability-improving mutations that fail recoverability verification. Under the original recovery representation, conventional repair strategies recover 0/197 of these natural failures. Deterministic oracle analysis recovers 48/197 under the original recovery language L0, while the extended recovery calculus increases empirical oracle recovery to 191/197. A protocol-locked 2x2 grounding-by-expressivity intervention then separates two bottlenecks: exact state-address grounding increases successful recovery from 0/48 to 38/48 (79.2%) when the original language is sufficient, while extending the recovery language enables recovery on 142/143 (99.3%) failures in the oracle-defined S1 stratum. On the primary gpt-oss-120b backbone, adding exact-address diagnostics to the richer language reduces recovery to 133/143 (93.0%); a Qwen3.8-27B replication preserves the grounding and expressivity effects but not this negative interaction, indicating that the latter is model-dependent. These results indicate that reliable agent self-evolution requires co-designing verification, state grounding, witness semantics, and recovery-language expressivity rather than relying on iterative prompting alone.

Randomized smoothing (RS) is one of the prominent techniques to ensure the correctness of machine learning models, where point-wise robustness certificates can be derived analytically. While RS is well understood for classification, its application to generative models is unclear, since their outputs are sequences rather than labels. We resolve this by connecting generative outputs to an oracle classification task and showing that RS can still be enabled: the final response can be classified as a discrete action (e.g., service-robot commands in VLAs), as harmful vs. harmless (content moderation or toxicity detection in VLMs), or even applying oracles to cluster answers into semantically equivalent ones. Provided that the error rate for the oracle classifier comparison is bounded, we develop the theory that associates the number of samples with the corresponding robustness radius. We further derive improved scaling laws analytically relating the certified radius and accuracy to the number of samples, showing that the earlier result of 2 to 3 orders of magnitude fewer samples sufficing with minimal loss remains valid even under weaker assumptions. Together, these advances make robustness certification both well-defined and computationally feasible for state-of-the-art VLMs, as validated against recent jailbreak-style adversarial attacks.

A robot that fails at a task faces the first decision in corrective dialogue: act on its own diagnosis, consult another onboard sensor, or interrupt a person. Choosing well requires knowing how much the robot's sensors reveal about the cause and how reliable the robot's own diagnosis is. We build a simulated benchmark in which every failure's true cause is known, because we injected it, and measure what each sensor reveals, with explicit checks against data leakage. Some failures are diagnosable from camera images; others only from the robot's force data (0.99 from force data, no image method above 0.55). We then test six open vision-language models. Their behavior tracks the surface of the prompt, not the evidence: moving the refusal option from last to first in the answer list collapses refusal rates from 78-100% to 0-6% in three of the six swept model-and-family pairs. Accuracy from frames stays at or below a majority-class baseline under every prompt variant, with or without worked examples, and stated confidence carries no information about correctness. Handing the same models the force data as ten lines of text produces the first above-baseline diagnoses, in four of the six models: much of the failure reflects missing sensor data, not missing ability. We pose the choice as a three-action decision problem, act, consult your own sensors, or ask a human, whose optimal policy follows from measured accuracy. The models do not follow it, and their ask rates ignore a fourfold change in question cost. One question to a human still lifts them from that baseline to roughly the answerer's own reliability (0.70-0.81 when they ask). The decision to ask should be tied to measured accuracy and stated costs, not to the model's confidence.

Standard accuracy benchmarks are designed to test how closely large language models (LLMs) approach correct answers, but are not suitable for testing whether LLMs stick with a correct answer when that answer is challenged by a plausible counter-argument. We introduce a controlled protocol for evaluating answer stability: after a model answers a multiple-choice question correctly, we challenge the model's answer with a coherent argument for an incorrect option and measure whether the model flips. The setup a) isolates argumentative content from overt social pressure and b) varies argument length, self-attribution, and cross-model source. Across seven frontier models and 57 MMLU subjects, flip rates range from 17.5% to 97.3%, revealing large differences in stability that are not captured by accuracy metrics alone. We find that self-attribution consistently increases flip rates (mean +7.1pp, up to +18.7pp). Also, pooling wrong-answer arguments across models and selecting the most effective one per question yields stronger adversarial challenges than relying on any single source model. We further construct MaxFlip, a curated challenge set that amplifies flips by up to +23.6pp over standard self-generated challenges. We release the protocol, challenge records, and MaxFlip to support stability evaluation alongside standard accuracy benchmarks. Materials are available at https://github.com/nafisenik/WhoFlips and https://hf.co/datasets/nafisehNik/WhoFlips.