Daily Papers

Recent advances in large language models and tool-using agents have expanded the range of benchmarked web tasks. Yet an important class of specialized retrieval tasks remains undercharacterized. On many specialized data-retrieval websites, answer-bearing evidence becomes accessible only after establishing the correct site-specific retrieval state through filters, views, hierarchies, or scopes. We term this capability state-gated retrieval (SGR). We introduce SGR-Bench, a benchmark for this setting containing 100 expert-curated tasks spanning six source families and 12 public data ecosystems. Each task requires discovering the appropriate website and configuring its site-specific retrieval state to produce a structured answer. SGR-Bench pairs constraint-guided and goal-oriented formulations of the same underlying problems, enabling controlled comparisons between explicit and implicit guidance for state-gated retrieval. We evaluate eight CLI-based agentic LLM systems and three commercial search-agent products. On SGR-Bench, the strongest system reaches only 66.18% item-level F1, while row-level F1 remains much lower. A manual audit of 156 analyzable failed CLI trajectories shows why: agents often reach a relevant web source, but establish the wrong site-specific retrieval state. Retrieval-scope drift (37.2%) and criterion mismatch (27.6%) dominate, whereas final answer composition accounts for only 10.3%. The dataset and single-case evaluation instructions are available at https://huggingface.co/datasets/PKUAIWeb/SGR-BENCH.

Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no optimization signal. Recent methods address this by incorporating rubric information as external guidance during rollout, yet they introduce a train-inference mismatch: the policy is optimized on rollouts produced under external guidance while this guidance is absent at inference time, causing error accumulation through autoregressive decoding. Moreover, these exploration-focused approaches overlook a fundamentally different failure mode that we term Suppressed Criteria (SC) -- criteria that are satisfied by some rollouts yet whose learning signals are lost during optimization because scalar reward aggregation assigns them non-positive aggregate advantages. Our analysis reveals that SC are remarkably prevalent: over 57% of samples exhibit this failure mode throughout training, with an average of 1.8 SC per sample. To simultaneously address both UC and SC without introducing training-inference mismatch, we propose Criterion-Distilled Policy Optimization (CriPO), which enhances rubric-based RL via on-policy self-distillation. For UC, CriPO constructs a criterion-injection self-teacher and computes a localized forward-KL loss to inject missing behaviors into the policy. For SC, CriPO employs a counterfactual self-teacher to locate criterion-relevant tokens in negative-advantage rollouts and flips their token-level advantages to positive values, preserving useful patterns that would otherwise be suppressed. Experiments on medicine and science benchmarks demonstrate that CriPO consistently outperforms rubric-based RL, achieving stronger final performance with approximately 2times fewer optimization steps.

The lack of clean speech is a practical challenge to the development of speech enhancement systems, which means that there is an inevitable mismatch between their training criterion and evaluation metric. In response to this unfavorable situation, we propose a training and inference strategy that additionally uses enhanced speech as a target by improving the previously proposed noisy-target training (NyTT). Because homogeneity between in-domain noise and extraneous noise is the key to the effectiveness of NyTT, we train various student models by remixing 1) the teacher model's estimated speech and noise for enhanced-target training or 2) raw noisy speech and the teacher model's estimated noise for noisy-target training. Experimental results show that our proposed method outperforms several baselines, especially with the teacher/student inference, where predicted clean speech is derived successively through the teacher and final student models.

Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-efficiently learning control of continuous tasks. Recent work in MBRL has mostly focused on using more advanced function approximators and planning schemes, with little development of the general framework. In this paper, we identify a fundamental issue of the standard MBRL framework -- what we call the objective mismatch issue. Objective mismatch arises when one objective is optimized in the hope that a second, often uncorrelated, metric will also be optimized. In the context of MBRL, we characterize the objective mismatch between training the forward dynamics model w.r.t.~the likelihood of the one-step ahead prediction, and the overall goal of improving performance on a downstream control task. For example, this issue can emerge with the realization that dynamics models effective for a specific task do not necessarily need to be globally accurate, and vice versa globally accurate models might not be sufficiently accurate locally to obtain good control performance on a specific task. In our experiments, we study this objective mismatch issue and demonstrate that the likelihood of one-step ahead predictions is not always correlated with control performance. This observation highlights a critical limitation in the MBRL framework which will require further research to be fully understood and addressed. We propose an initial method to mitigate the mismatch issue by re-weighting dynamics model training. Building on it, we conclude with a discussion about other potential directions of research for addressing this issue.

An agent that inherits six one-line memories may pull at most one archived source record before acting; a directive written into the store can steer that choice: a pointer to the record, a criterion that identifies it, or both. Across twelve registered studies on one instrument lineage (14,760 attempts) we measured where the request goes under each form. On six direct-provider models a length-matched criterion exceeded a bare id by +35.0 points [+31.2, +38.8] (Study D); the contrast failed its registered superiority rule on a nine-model OpenRouter-served panel (Study E). Appending the id cancelled the criterion on three Claude models (Opus 5: 40/40 to 0/40; Study F-x); six byte-matched edits gave each exact string its own effect (Study G), and a re-run at eighty runs per cell left fifteen of thirty replication contrasts within the margin, fifteen unresolved and none beyond (Study G'). A ratification line (+96.0 points on Opus 5) and a budget of two credits restored the target on all three (Study J); across five criterion strings the suffix's cancellation held for four of the five wordings on Opus 5 and all five wordings on Fable 5.1 (Study H2); in a second store every model followed the criterion (Study H1). Continued into a decision, the criterion moved the choice toward the current record (+100.0 points, Opus 5) and away from it on Fable 5.1 (Study I). A one-character plan pointer's effect (+78.0 points; Study B, after a correction of its first repository report) returned the same verdict under a prospectively registered re-run (+81.7 points; Study B'). All results are descriptive effects of exact edits on fixed panels with registered intervals and no mechanism claim.

Score- and flow-matching models often rely on preference-based reinforcement learning for two purposes: aligning with subjective preferences and, surprisingly, recovering properties such as visual realism and coherent object structure that matching-based training is intended to learn from the data itself. We argue that this reflects a structural mismatch. Matching losses measure ell_2 regression error on the velocity or score field under training-time marginals, a proxy poorly aligned with the visual and semantic properties that determine sample quality at inference. Given a reward aligned with these properties, RL sidesteps the mismatch by evaluating the model on its own samples and following the reward landscape directly. The challenge is to obtain such a reward without relying on human preferences, which are expensive and conflate data realism with annotator inclinations. We propose Discriminator-Guided RL (DRL). DRL trains a discriminator to separate data from base-model samples in a pretrained representation space and uses its logit as the reward in KL-regularized RL. The pretrained space restricts the discriminator to perceptually meaningful directions, and the logit estimates the log-likelihood ratio between data and model, which is the optimal reward for targeting the data distribution. Across SiT, JiT, REPA, and RAE, DRL reduces guidance-free FID (e.g., 9.38 to 2.62 on SiT) and semantic-space FD (e.g., 88.2 to 19.3 on DINOv3 for SiT), with consistent gains across all backbones, and improves human-preference rewards without training on them. It also yields a better Pareto frontier between preference reward and image fidelity under subsequent preference-based post-training, increasing alignment while reducing low-level artifacts such as oversaturation and excessive brightness.

Muon has emerged as an efficient alternative to Adam for pretraining, yet remains underused for fine-tuning. A key obstacle is that most open models are pretrained with Adam, and naively switching to Muon for fine-tuning leads to degraded performance due to an optimizer mismatch. We investigate this mismatch through controlled experiments and relate it to the distinct implicit biases of Adam and Muon. We provide evidence that the mismatch disrupts pretrained knowledge, and that this disruption scales with update strength. This leads us to hypothesize that constraining updates should mitigate the mismatch. We validate this with LoRA: across language and vision tasks, LoRA reduces the performance gap between Adam and Muon observed under full fine-tuning. Studies on LoRA rank, catastrophic forgetting, and LoRA variants further confirm that mismatch severity correlates with update strength. These results shed light on how optimizer mismatch affects fine-tuning and how it can be mitigated. Our code is available at https://github.com/XingyuQu/muon-finetune.

Many existing conditional score-based data generation methods utilize Bayes' theorem to decompose the gradients of a log posterior density into a mixture of scores. These methods facilitate the training procedure of conditional score models, as a mixture of scores can be separately estimated using a score model and a classifier. However, our analysis indicates that the training objectives for the classifier in these methods may lead to a serious score mismatch issue, which corresponds to the situation that the estimated scores deviate from the true ones. Such an issue causes the samples to be misled by the deviated scores during the diffusion process, resulting in a degraded sampling quality. To resolve it, we formulate a novel training objective, called Denoising Likelihood Score Matching (DLSM) loss, for the classifier to match the gradients of the true log likelihood density. Our experimental evidence shows that the proposed method outperforms the previous methods on both Cifar-10 and Cifar-100 benchmarks noticeably in terms of several key evaluation metrics. We thus conclude that, by adopting DLSM, the conditional scores can be accurately modeled, and the effect of the score mismatch issue is alleviated.

Modern LLM RL systems separate rollout generation from policy optimization. These two stages are expected to produce token probabilities that match exactly. However, implementation differences can make them assign different values to the same sequence under the same model weights, inducing Training-Inference Mismatch (TIM). TIM is difficult to inspect because it is entangled with off-policy drift and common stabilization mechanisms. In this work, we isolate TIM in a zero-mismatch diagnostic setting (VeXact), and show that small token-level numerical disagreements can independently cause training collapse. We further show that TIM changes the effective optimization problem, and identify a set of remedies that could mitigate TIM. Our results suggest that TIM is not benign numerical noise, but a systems-level perturbation that should be treated as a first-order factor in analyzing LLM RL stability.

We consider whether off-policy experience from a smaller, weaker model can elicit capability in a stronger learner that on-policy RL fine-tuning (e.g., GRPO) does not reach. We find that injecting mathematically wrong drafts from a smaller but more domain-trained model -- mismatched to the current problem -- into a stronger learner's GRPO context consistently outperforms standard on-policy GRPO on held-out MATH-500 and out-of-distribution AIME 2025/2026. Concretely, we use Mathstral-7B as the learner, Qwen2.5-Math-1.5B as the draft model, 8.8K Level 3--5 MATH problems (with MATH-500 held out), and train with Dr. GRPO. Mismatch is an active ingredient: shuffling drafts to mismatched problems while holding everything else constant yields +1.62pp on MATH-500 (greedy pass@1) over the matched-wrong variant (n=10 seeds, p=0.0015, Welch's t). In fact, the mismatched-wrong variant leads all other variants we tested on MATH-500 across both greedy pass@1 and sampling pass@k. On out-of-distribution AIME 2025 and 2026, the mismatched-wrong variant uniquely lifts pass@k above both Mathstral-7B (in its native [INST] format) and the Qwen2.5-Math-1.5B draft model at every sample budget from k=1 to k=1024 across 2 seeds (+14.2pp on 2025 and +9.0pp on 2026 at pass@1024 over Mathstral-7B), and at pass@1024 also leads no-draft, matched-wrong, and mismatched-correct variants on both years. All variants use the same prompt with no draft injection at test time. The recipe -- trained on a single GPU with no SFT, no reward models, no synthesized data, and no produce-critique-revise inner loop -- reaches 71.98% MATH-500 on Mathstral-7B-v0.1, the highest published result on this model to our knowledge, surpassing the heavier WizardMath pipeline at 70.9% on full MATH (SFT + PPO with process/instruction reward models).

Lean verifies that a generated declaration is well typed, but cannot check whether it states what the user intended. We measure how far compilation overstates faithfulness in statement autoformalization without canonical Lean targets. Our criterion requires Lean compilation and agreement of two LLM judges from different model families. On an independent random audit, it agrees with the human majority on 91.5% of cases; it rarely accepts an unfaithful output but rejects some faithful ones. Across eight systems on 227 university-level statements, the gap between compiling and accepted outputs ranges from 1.3 to 29.5 percentage points: our tool-augmented agent compiles 87.2% of the statements, but only 57.7% are accepted. Human review of all the agent's rejections confirms that 23.8 of its 29.5 points are real translation errors. For the gaps of the seven one-shot systems, a two-stage review by Claude Opus 5.5 with Lean and GPT-6.1 Sol, validated on these human labels, finds that 78-100% are real translation errors for six of them. When GPT-5.2, one of the two judges, grades the translations of two models released after it, GPT-6 Astra and Claude Opus 5.5, it accepts 83% of the statements for each, against 21% for one-shot GPT-5.2. For these two models, the same review classifies only 4 of their 54 rejections (7%) as translation errors, so most of their measured gap is judge errors rather than translation errors. This is consistent with these systems erring less while the judge does not: the fewer errors a system makes, the smaller the share of its rejections that are translation errors. Finally, a 2^3 tool ablation of the agent shows that Lean feedback drives most of its compilation gain, and that nearly half of this gain fails the semantic criterion. Code and data are available at https://github.com/grenadeComing/beyond-compilation.

In recent years, the rapid increase in scientific papers has overwhelmed traditional review mechanisms, resulting in varying quality of publications. Although existing methods have explored the capabilities of Large Language Models (LLMs) for automated scientific reviewing, their generated contents are often generic or partial. To address the issues above, we introduce an automated paper reviewing framework SEA. It comprises of three modules: Standardization, Evaluation, and Analysis, which are represented by models SEA-S, SEA-E, and SEA-A, respectively. Initially, SEA-S distills data standardization capabilities of GPT-4 for integrating multiple reviews for a paper. Then, SEA-E utilizes standardized data for fine-tuning, enabling it to generate constructive reviews. Finally, SEA-A introduces a new evaluation metric called mismatch score to assess the consistency between paper contents and reviews. Moreover, we design a self-correction strategy to enhance the consistency. Extensive experimental results on datasets collected from eight venues show that SEA can generate valuable insights for authors to improve their papers.

Foundation models rely on large-scale web-crawled datasets, which frequently contain noisy data, biases, and irrelevant content. Existing data selection techniques typically use human heuristics, downstream evaluation datasets, or specialized scoring models, and can overlook samples' utility in the training process. Instead, we propose a new approach, Mimic Score, a data quality metric that uses a pretrained reference model as a guide to assess the usefulness of data samples for training a new model. It relies on the alignment between the gradient of the new model parameters and the vector pointing toward the reference model in weight space. Samples that misalign with this direction are considered low-value and can be filtered out. Motivated by the Mimic score, we develop Grad-Mimic, a data selection framework that identifies and prioritizes useful samples, automating the selection process to create effective filters. Empirically, using Mimic scores to guide model training results in consistent performance gains across six image datasets and enhances the performance of CLIP models. Moreover, Mimic scores and their associated filters improve upon existing filtering methods and offer accurate estimation of dataset quality.

Priorities in multi-criteria decision-making (MCDM) convey the relevance preference of one criterion over another, which is usually reflected by imposing the non-negativity and unit-sum constraints. The processing of such priorities is different than other unconstrained data, but this point is often neglected by researchers, which results in fallacious statistical analysis. This article studies three prevalent fallacies in group MCDM along with solutions based on compositional data analysis to avoid misusing statistical operations. First, we use a compositional approach to aggregate the priorities of a group of DMs and show that the outcome of the compositional analysis is identical to the normalized geometric mean, meaning that the arithmetic mean should be avoided. Furthermore, a new aggregation method is developed, which is a robust surrogate for the geometric mean. We also discuss the errors in computing measures of dispersion, including standard deviation and distance functions. Discussing the fallacies in computing the standard deviation, we provide a probabilistic criteria ranking by developing proper Bayesian tests, where we calculate the extent to which a criterion is more important than another. Finally, we explain the errors in computing the distance between priorities, and a clustering algorithm is specially tailored based on proper distance metrics.

Many peer-review venues are using algorithms to assign submissions to reviewers. The crux of such automated approaches is the notion of the "similarity score" -- a numerical estimate of the expertise of a reviewer in reviewing a paper -- and many algorithms have been proposed to compute these scores. However, these algorithms have not been subjected to a principled comparison, making it difficult for stakeholders to choose the algorithm in an evidence-based manner. The key challenge in comparing existing algorithms and developing better algorithms is the lack of publicly available gold-standard data. We address this challenge by collecting a novel dataset of similarity scores that we release to the research community. Our dataset consists of 477 self-reported expertise scores provided by 58 researchers who evaluated their expertise in reviewing papers they have read previously. Using our dataset, we compare several widely used similarity algorithms and offer key insights. First, all algorithms exhibit significant error, with misranking rates between 12%-30% in easier cases and 36%-43% in harder ones. Second, most specialized algorithms are designed to work with titles and abstracts of papers, and in this regime the SPECTER2 algorithm performs best. Interestingly, classical TF-IDF matches SPECTER2 in accuracy when given access to full submission texts. In contrast, off-the-shelf LLMs lag behind specialized approaches.

Many tasks revolve around editing a document, whether code or text. We formulate the revision similarity problem to unify a wide range of machine learning evaluation problems whose goal is to assess a revision to an existing document. We observe that revisions usually change only a small portion of an existing document, so the existing document and its immediate revisions share a majority of their content. We formulate five adequacy criteria for revision similarity measures, designed to align them with human judgement. We show that popular pairwise measures, like BLEU, fail to meet these criteria, because their scores are dominated by the shared content. They report high similarity between two revisions when humans would assess them as quite different. This is a fundamental flaw we address. We propose a novel static measure, Excision Score (ES), which computes longest common subsequence (LCS) to remove content shared by an existing document with the ground truth and predicted revisions, before comparing only the remaining divergent regions. This is analogous to a surgeon creating a sterile field to focus on the work area. We use approximation to speed the standard cubic LCS computation to quadratic. In code-editing evaluation, where static measures are often used as a cheap proxy for passing tests, we demonstrate that ES surpasses existing measures. When aligned with test execution on HumanEvalFix, ES improves over its nearest competitor, SARI, by 12% Pearson correlation and by >21% over standard measures like BLEU. The key criterion is invariance to shared context; when we perturb HumanEvalFix with increased shared context, ES' improvement over SARI increases to 20% and >30% over standard measures. ES also handles other corner cases that other measures do not, such as correctly aligning moved code blocks, and appropriately rewarding matching insertions or deletions.

Aligning large language models with human preferences is critical for creating reliable and controllable AI systems. A human preference can be visualized as a high-dimensional vector where different directions represent trade-offs between desired attributes (e.g., helpfulness vs. verbosity). Yet, because the training data often reflects dominant, average preferences, LLMs tend to perform well on common requests but fall short in specific, individual needs. This mismatch creates a preference coverage gap. Existing methods often address this through costly retraining, which may not be generalized to the full spectrum of diverse preferences. This brittleness means that when a user's request reflects a nuanced preference deviating from the training data's central tendency, model performance can degrade unpredictably. To address this challenge, we introduce Robust Preference Selection (RPS), a post-hoc, training-free method by leveraging directional neighborhood consensus. Instead of forcing a model to generate a response from a single, highly specific preference, RPS samples multiple responses from a local neighborhood of related preferences to create a superior candidate pool. It then selects the response that best aligns with the user's original intent. We provide a theoretical framework showing our neighborhood generation strategy is provably superior to a strong baseline that also samples multiple candidates. Comprehensive experiments across three distinct alignment paradigms (DPA, DPO, and SFT) demonstrate that RPS consistently improves robustness against this baseline, achieving win rates of up to 69% on challenging preferences from under-represented regions of the space without any model retraining. Our work presents a practical, theoretically-grounded solution for enhancing the reliability of preference-aligned models.

We introduce the first version of KWBench (Knowledge Work Bench), a benchmark for unprompted problem recognition in large language models: can an LLM identify a professional scenario before attempting to solve it. Existing frontier benchmarks have saturated, and most knowledge-work evaluations to date reduce to extraction or task completion against a specification. KWBench targets the step before that: recognizing the governing structure of the situation from raw inputs alone. The benchmark contains 223 tasks sourced from practitioners across acquisitions, contract negotiations, clinical pharmacy, organizational politics, fraud analysis, and incentive design. Each task encodes a formal game-theoretic pattern (principal-agent conflict, signaling, mechanism design failure, strategic omission, coalitional dynamics, strategic interdependence) and carries structured ground truth recording the expert reading of the situation and the anticipated failure modes. Models receive raw data and a task prompt with no indication of problem type. Scoring is a three-tier rubric gated by a mandatory conjunctive check. Mandatory criteria encode the predicted wrong paths. We evaluate 16 models. The best model passes on 27.9% of tasks. The top two models agree on only 31.7% of their passes. Among the top 8, 44 tasks are solved by exactly one model; routing across the top 8 covers 50.7% of the benchmark, nearly double the best single model. Conditional on passing, quality scores converge (approx 83% across models); unconditional scores do not. Same models articulate the relevant game-theoretic concept correctly when asked, then fail to apply it unprompted. We release KWBench to shift how frontier models are evaluated on knowledge work, scoring them on whether they recognize the right problem from the situation alone, not only on how well they execute once the problem has been framed for them.

Personalized treatment effect estimates are often of interest in high-stakes applications -- thus, before deploying a model estimating such effects in practice, one needs to be sure that the best candidate from the ever-growing machine learning toolbox for this task was chosen. Unfortunately, due to the absence of counterfactual information in practice, it is usually not possible to rely on standard validation metrics for doing so, leading to a well-known model selection dilemma in the treatment effect estimation literature. While some solutions have recently been investigated, systematic understanding of the strengths and weaknesses of different model selection criteria is still lacking. In this paper, instead of attempting to declare a global `winner', we therefore empirically investigate success- and failure modes of different selection criteria. We highlight that there is a complex interplay between selection strategies, candidate estimators and the data used for comparing them, and provide interesting insights into the relative (dis)advantages of different criteria alongside desiderata for the design of further illuminating empirical studies in this context.

Rubric-based reinforcement learning (Rubric-RL) trains language models where no verifier exists. A judge checks each criterion of a rubric, and the verdicts are aggregated into a reward, most often by a weighted sum. We show that this additive aggregation is the weak point. Under a sum, criteria compensate for one another: a policy that misses the one decision that matters can buy the points back with advice nobody asked for. On clinical consultation, such a policy scores higher and answers worse. Rubric coverage rises while appropriateness on held-out physician criteria falls below the untrained model. The medical criteria are not to blame. Grouped so that they must hold together, the same criteria, unchanged to the word, recover a third of the loss; shorter answers recover almost none. We therefore propose Protocol-level Rubrics (ProRubric), which keeps what the criteria ask for and changes how they are aggregated. It groups a checklist into a few protocol-level dimensions. A dimension counts only when all of its criteria hold and its failure clause does not fire. The grouping is done once, offline, and leaves the optimizer unchanged. ProRubric raises appropriateness by 10.8 points without losing coverage and has the best seven-benchmark average at both scales. Reward validity is set not only by what a rubric verifies, but by how it aggregates. Code is available at https://github.com/Estrellajer/ProRubric

With the growing interest in large language models, the need for evaluating the quality of machine text compared to reference (typically human-generated) text has become focal attention. Most recent works focus either on task-specific evaluation metrics or study the properties of machine-generated text captured by the existing metrics. In this work, we propose a new evaluation scheme to model human judgments in 7 NLP tasks, based on the fine-grained mismatches between a pair of texts. Inspired by the recent efforts in several NLP tasks for fine-grained evaluation, we introduce a set of 13 mismatch error types such as spatial/geographic errors, entity errors, etc, to guide the model for better prediction of human judgments. We propose a neural framework for evaluating machine texts that uses these mismatch error types as auxiliary tasks and re-purposes the existing single-number evaluation metrics as additional scalar features, in addition to textual features extracted from the machine and reference texts. Our experiments reveal key insights about the existing metrics via the mismatch errors. We show that the mismatch errors between the sentence pairs on the held-out datasets from 7 NLP tasks align well with the human evaluation.

Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an LLM judge: a second model reads the note against the transcript and flags problems. We ask whether judges detect omissions. Public corpora cannot supply the answer key: their clinician reference notes and transcripts are materially discrepant. Our benchmark has 500 single-error note pairs from audited fact sheets, 298 with a named fact certainly absent and 202 added-or-altered controls. Across eight judge designs, paired discrimination (the flawed note below its clean twin, 0.5 a coin flip) reads 0.79-0.94 on added or altered content and 0.50-0.63 on omissions. On single notes, no design flags omissions reliably more often than perfect notes. Wording changes, voting and GEPA prompt optimisation move the operating point without creating usable detection. Restructuring the task recovers it: list the facts the transcript establishes, then check the note for each. Two methods reach it independently and trade off: a per-fact pipeline, and a GEPA-evolved prompt doing the same in one call. The pipeline's flags name the missing fact and its severity at 2.7% false alarms. The single call detects more (36.9% against 24.6%, p=0.002) at 6.2% false alarms and a tenth of the cost per note. A physician author validated 70 items and, where the two routes disagree, sided with the pipeline on 10 of 10 (p=0.002). A second clinician, not an author, graded the severity rubric blind and agrees to within a grade. On real vendor notes from a companion census no benchmark threshold transfers, but the re-calibrated single call detects more than the best of the eight at half its false-alarm rate. Omissions whose fact is restated elsewhere defeat both routes. We release the benchmark, prompts and judgements.

We study length inflation in on-policy distillation (OPD), where student responses can become excessively long and even exhaust the generation budget. We identify termination-token mismatch between base students and post-trained teachers as an important source of this behavior. Across Qwen3, Llama, and Gemma, the two models can place their stopping probability on different EOS tokens, even when their declared stopping sets are identical. This mismatch can suppress the student's preferred termination action without reliably transferring the teacher-preferred alternative. We show that aligning the decoding stopping set alone is insufficient, while treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigates mismatch-induced length inflation across all three model families. To further understand how termination behavior evolves over training, we study OPD across different K2-Horizon training stages. This stage-wise analysis shows that termination preferences can shift substantially during training, while also revealing a distinct length inflation late in the OPD run that persists beyond termination alignment. Together, these results identify termination mismatch as an important, but not exhaustive, source of OPD length dynamics. We release an implementation incorporating the proposed termination-handling corrections.

The trade-off between robustness and accuracy has been widely studied in the adversarial literature. Although still controversial, the prevailing view is that this trade-off is inherent, either empirically or theoretically. Thus, we dig for the origin of this trade-off in adversarial training and find that it may stem from the improperly defined robust error, which imposes an inductive bias of local invariance -- an overcorrection towards smoothness. Given this, we advocate employing local equivariance to describe the ideal behavior of a robust model, leading to a self-consistent robust error named SCORE. By definition, SCORE facilitates the reconciliation between robustness and accuracy, while still handling the worst-case uncertainty via robust optimization. By simply substituting KL divergence with variants of distance metrics, SCORE can be efficiently minimized. Empirically, our models achieve top-rank performance on RobustBench under AutoAttack. Besides, SCORE provides instructive insights for explaining the overfitting phenomenon and semantic input gradients observed on robust models. Code is available at https://github.com/P2333/SCORE.

Research planning is the decisive capability of AI scientists. Yet a research plan admits no verifiable answer, so reinforcement learning lacks the environment it requires: tasks paired with a critic. Rubrics extracted from scientific papers can supply the critic. Existing pipelines, however, draw the question and the criteria from the same content, so the reward can be earned by paraphrase. The rubric is further compressed into a single scalar per rollout. We introduce PaperGym, a unified framework that turns each research paper into a complete training environment. PaperGym exploits the structure of a paper: the question is synthesized from the research goal and background, while the criteria are derived from the method and experiments. The criteria span methodological innovation and experimental design, and criterion leakage falls to 3.7%, versus 11.90% to 34.10% in existing datasets. Training uses the rubric twice: first as privileged context for OPSD's self-teacher, then as the reward for GRPO. Across Qwen3-1.7B/4B/8B, this schedule outperforms supervised fine-tuning, either stage alone, and the reverse ordering, improving five-benchmark averages by +5.6, +5.0, and +4.8 points. With the recipe held fixed, models trained on PaperGym-20k win 58.1% of three-way comparisons, against 28.2% for RubricHub Science. The trained Qwen3-8B reaches 73.48 on ResearchQA, above the far larger Kimi K2.6. We release the pipeline, the 20,000-instance corpus PaperGym-20k, and the benchmarks PaperGym-Innov and PaperGym-Design.

Finetuning a language model can lead to emergent misalignment (EM) [Betley et al., 2025b]. Models trained on a narrow distribution of misaligned behavior generalize to more egregious behaviors when tested outside the training distribution. We study a set of interventions proposed to reduce EM. We confirm that these interventions reduce or eliminate EM on existing evaluations (questions like "How do I make a quick buck?"). However, if the evaluation prompts are tweaked to resemble the training context, the model displays EM. We call this conditional misalignment. As in standard EM, the model displays misaligned behaviors more egregious than those seen during training, but only on inputs sharing features with the training data. The first two interventions are diluting misaligned data with benign data, and finetuning on benign data after misaligned data. Both produce conditional misalignment. For instance, models trained on a mix of only 5% insecure code still show misalignment when asked to format responses as Python strings (resembling the training context). The third intervention is inoculation prompting. Here, statements with a similar form to the inoculation prompt serve as triggers for misalignment, even if they have the opposite meaning. On the positive side, inoculation prompting has lower (but still non-zero) conditional misalignment if training is on-policy or includes reasoning distillation. Our results imply that in realistic post-training, where misaligned data is typically combined with benign data, models may be conditionally misaligned even if standard evaluations look clean.

We present a new approach for mitigating unfairness in learned classifiers. In particular, we focus on binary classification tasks over individuals from two populations, where, as our criterion for fairness, we wish to achieve similar false positive rates in both populations, and similar false negative rates in both populations. As a proof of concept, we implement our approach and empirically evaluate its ability to achieve both fairness and accuracy, using datasets from the fields of criminal risk assessment, credit, lending, and college admissions.

Machine learning models always make a prediction, even when it is likely to be inaccurate. This behavior should be avoided in many decision support applications, where mistakes can have severe consequences. Albeit already studied in 1970, machine learning with rejection recently gained interest. This machine learning subfield enables machine learning models to abstain from making a prediction when likely to make a mistake. This survey aims to provide an overview on machine learning with rejection. We introduce the conditions leading to two types of rejection, ambiguity and novelty rejection, which we carefully formalize. Moreover, we review and categorize strategies to evaluate a model's predictive and rejective quality. Additionally, we define the existing architectures for models with rejection and describe the standard techniques for learning such models. Finally, we provide examples of relevant application domains and show how machine learning with rejection relates to other machine learning research areas.

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

LLM benchmark labels are frozen at release and silently propagated into downstream benchmarks, errors and all. We introduce an Item Response Theory-based indicator that surfaces likely mislabels at 95% precision in the top 200 examples across seven preference and multiple-choice benchmarks using responses from 114 models, outperforming a supervised classifier. We trace these errors to mechanical labeling heuristics, upstream annotation mistakes inherited unchanged from source datasets, and fundamentally ambiguous items without a defensible single label. The same model fit reveals that reward models specialize in stylistic preference rather than factual knowledge, and identifies one frontier reward model that agrees with detected mislabels at 78% accuracy versus 38% for its peers, consistent with benchmark contamination or benchmark-specific over-optimization.

Reinforcement learning (RL) for large language models (LLMs) is dominated by the cost of rollout generation, which has motivated the use of low-precision rollouts (e.g., FP8) paired with a BF16 trainer to improve throughput and reduce memory pressure. This introduces a rollout-training mismatch that biases the policy gradient and can cause training to collapse outright on reasoning benchmarks. We show that the mismatch is non-stationary and acts as a double-edged sword: early in training it provides a stochastic exploration bonus, exposing the gradient to trajectories the trainer would otherwise under-sample, but the same perturbation transitions into a destabilizing source of bias as the policy concentrates. To solve this, we propose Adaptive Importance Sampling (AIS), a correction framework that adjusts the strength of its intervention on a per-batch basis. AIS combines three real-time diagnostics, namely weight reliability, divergence severity, and variance amplification, into a single mixing coefficient that interpolates between the uncorrected and fully importance-weighted gradients, suppressing the destabilizing component of the mismatch while preserving its exploratory benefit. We integrate AIS into GRPO and evaluate it on the diffusion-based LLaDA-8B-Instruct and the autoregressive Qwen3-8B and Qwen3.5-9B across mathematical reasoning and planning benchmarks. AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8.

Low-precision execution can substantially accelerate reinforcement learning (RL) for large language models, but discrepancies between learner and sampler execution can destabilize policy optimization. In this paper, we characterize the interaction between mismatch and the policy-gradient direction, distinguishing locally amplifying from contracting update contributions that mismatch magnitude alone cannot identify. In native NVFP4 runs, we observe an early imbalance between the two amplifying regions, favoring negative-advantage, negative-gap updates. Their tail tokens become concentrated in a small fraction of response segments before mismatch spreads globally. Motivated by these findings, we introduce TRIAGE, a direction-aware stabilization method that uses segment-level diagnosis to selectively rebalance policy-gradient updates and applies bounded repair to residual severe mismatch. TRIAGE modifies the optimization objective while retaining native NVFP4 weight-and activation 4-bit (W4A4) forward execution on both the sampler and learner. Experiments on Qwen3-4B and Qwen3-30B-A3B show stable optimization throughout the evaluated training horizon and achieve full precision level performance across five mathematical reasoning benchmarks, while native NVFP4 with TRIAGE provides up to 2.3x higher rollout throughput than BF16.

Warning: This paper studies stereotypes and biases, and contains potentially disturbing examples, used for illustration purposes only. Our findings should not be interpreted as an argument against alignment. Instead, this paper highlights the need for principled approaches to more advanced alignment. Alignment aims to ensure that large language models (LLMs) behave safely and reliably, including by avoiding unsafe inferences. However, we show that such safety-oriented behaviors can misfire: models may reject warranted conclusions even when they are explicitly supported by context. We call this failure mode misfired alignment, where alignment-induced changes cause LLMs to override explicit evidence. To quantify this phenomenon, specifically on stereotype-related alignment, we introduce VETO, a benchmark consisting of 2,032 BBQ-derived contrastive pairs, and define a new metric, Misfired Alignment Rate (MAR), which measures on a 0 to 100 scale how often a model fails on a stereotype-related question but succeeds on its contrastive counterpart. We benchmark 25 LLMs on VETO, and show that all LLMs, including the most recent ones, exhibit non-trivial (4.7 to 18.9%) MARs while all human participants achieve 0.0% MAR. Controlled priming experiments further show that alignment-induced cues can substantially amplify MAR across LLMs, indicating that these failures are not merely artifacts of individual examples but can be induced by safety-related framing. Mechanistic analyses on open-weight LLMs reveal late-layer suppression of evidence-supported answers, and comparisons between instruct and base LLMs suggest that this suppression emerges after instruction training. These findings show that current alignment methods can overgeneralize surface-level safety cues, to the point of overriding objective evidence, motivating more work on alignment objectives that better preserve contextual grounding.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

The validity screen (Cacioli, 2026d, 2026e) classifies LLM confidence signals as Valid, Indeterminate, or Invalid. We test whether these classifications predict selective prediction performance. Twenty frontier LLMs from seven families were evaluated on 524 items across six cognitive tracks. Valid models show mean Type 2 AUROC = .624 (SD = .048). Invalid models show mean AUROC = .357 (SD = .231). Cohen's d = 2.81, p = .002. The tiers order monotonically: Invalid (.357) < Indeterminate (.554) < Valid (.624). Split-half cross-validation yields median d = 1.77, P(d > 0) = 1.0 across 1,000 splits. The three-tier classification accounts for 47% of the variance in AUROC. DeepSeek-R1 drops from 85.3% accuracy at full coverage to 11.3% at 10% coverage. The screen predicts the criterion. For selective prediction, the screen matters.

Statistical data fusion combines two panels that share a block of covariates but observe disjoint outcome blocks, and in its traditional form no row observes both outcomes at once. That rules out the discriminative criterion one would rather train a Deep Boltzmann Machine with, since multi-prediction training needs ground truth for whatever it holds out. We propose observed-block multi-prediction, which restricts the multi-prediction objective to targets drawn from what each row actually observes. It is well defined for any missingness pattern and reduces to the original criterion when rows are complete. Having a discriminative criterion that survives the setting lets us ask whether the joint model is needed at all, by separating what it contributes into a representation part and an inference part. On two consumer panels, on grids over sample size and covariate width spanning 35 cells and 875 runs, the fine-tuned DBM is the best of fifteen methods in every cell; but almost none of that advantage comes from generative pre-training, which is confined to the smallest sample size on one dataset and absent on the other. It comes from conditioning on one outcome block when predicting the other. This term amounts to +0.19 and +0.07 percentage points, is positive in all 35 cells, and, unlike every other contribution we measure, neither decays as the panels grow, nor requires a second hidden layer, nor requires more inference. Permuting one outcome block to destroy its association with the other removes the gain entirely, which is what the account predicts. The margins are small. But a small effect that does not decay is a different object from one that does, because it rests on evidence that no model mapping covariates to outcomes can accept.

Breakthroughs in machine learning are rapidly changing science and society, yet our fundamental understanding of this technology has lagged far behind. Indeed, one of the central tenets of the field, the bias-variance trade-off, appears to be at odds with the observed behavior of methods used in the modern machine learning practice. The bias-variance trade-off implies that a model should balance under-fitting and over-fitting: rich enough to express underlying structure in data, simple enough to avoid fitting spurious patterns. However, in the modern practice, very rich models such as neural networks are trained to exactly fit (i.e., interpolate) the data. Classically, such models would be considered over-fit, and yet they often obtain high accuracy on test data. This apparent contradiction has raised questions about the mathematical foundations of machine learning and their relevance to practitioners. In this paper, we reconcile the classical understanding and the modern practice within a unified performance curve. This "double descent" curve subsumes the textbook U-shaped bias-variance trade-off curve by showing how increasing model capacity beyond the point of interpolation results in improved performance. We provide evidence for the existence and ubiquity of double descent for a wide spectrum of models and datasets, and we posit a mechanism for its emergence. This connection between the performance and the structure of machine learning models delineates the limits of classical analyses, and has implications for both the theory and practice of machine learning.

Large language models can diagnose a student program from its code, problem statement, and reference solution. Evaluating this ability requires a clear definition of what counts as the correct diagnosis. We introduce PyMETA, a Python error dataset with 48,646 student submissions to 155 problems. Every submission has a single label for the first execution error identified by an Online Judge, or No Error when the program passes all tests. A targeted subset of 97 submissions also has expert labels collected through iterative repair and re-execution. The taxonomy has three levels; its most detailed level contains 14 labels, including No Error, Logic Error, named Python exceptions, and an Other Errors category. We evaluate two finetuned models and two groups of prompted LLMs: four earlier models and four recent models. When evaluated against the first execution error, the recent prompted models reach 87.5--93.8% macro F1, above the strongest finetuned baseline at 80.6%. This is the opposite of the comparison obtained with the earlier prompted models. On the 97-item expert subset, however, exact-set match is only 43.3--48.5%, although sample F1 is about 79--81%. Output format also matters. On the same 45 audited submissions whose expert label sets do not contain Logic Error, none of the four recent models returns that label under single-error prompting, but 46.7--57.8% of their multi-error outputs include it, usually after an explicit-error label. The results show that model rankings and claims about label bias depend on the meaning of the gold label, the number of labels a model may return, and the scoring rule.

A service system may use the same categories to assign standard allocations and to check whether each service is fulfilled. Finer categories can match individual needs more closely, but they divide the observations available for monitoring. We study this conflict for a fixed menu from which participants select by declaring a category, with fulfilment assessed from each category's aggregate outcomes during a fixed period. We relate allocation loss to variation in preferred allocations within categories and identify conditions under which aggregate observations preserve the verification performance of individual records. For nested refinements under stated utility and observation assumptions, an allocation-loss tolerance and a per-category detection target define a feasibility band. Categories must be fine enough to provide suitable allocations but sufficiently populated to support verification. A source-dependent lower bound on declaration entropy and a minimum contributor requirement give necessary information and population constraints. For an explicit finite population with quadratic utility and binary service outcomes, we prove the exact feasible range across all categorical designs and exhibit designs attaining the information lower bound at specified tolerances. The results provide conditions for choosing categories jointly for allocation and verification.

Backward compatibility of model predictions is a desired property when updating a machine learning driven application. It allows to seamlessly improve the underlying model without introducing regression bugs. In classification tasks these bugs occur in the form of negative flips. This means an instance that was correctly classified by the old model is now classified incorrectly by the updated model. This has direct negative impact on the user experience of such systems e.g. a frequently used voice assistant query is suddenly misclassified. A common reason to update the model is when new training data becomes available and needs to be incorporated. Simply retraining the model with the updated data introduces the unwanted negative flips. We study the problem of regression during data updates and propose Backward Compatible Weight Interpolation (BCWI). This method interpolates between the weights of the old and new model and we show in extensive experiments that it reduces negative flips without sacrificing the improved accuracy of the new model. BCWI is straight forward to implement and does not increase inference cost. We also explore the use of importance weighting during interpolation and averaging the weights of multiple new models in order to further reduce negative flips.

Sharp bounds on two potential-outcome means need not combine into sharp bounds on their difference. We characterize exactly when composition is guaranteed across environments sharing a joint potential-outcome law. For binary treatment and outcome, fix strict selection-probability bounds and interior treatment propensities. Subtracting the full model's sharp mean intervals gives sharp average-treatment-effect bounds for every compatible observed law if and only if the absolute selection bounds coincide across environments. Under this alignment, the complete mean region is rectangular; its endpoints and attaining causal models require O(K) arithmetic operations for K environments. For every mismatch, failure occurs on a relatively open set of positive observed tables at the prescribed propensities and persists under small joint drift. We give computable perturbation certificates and an exact family whose gap is first order in the mismatch. A separate finite-confounding family shows when the lost joint information changes an identified causal sign. These results distinguish local marginal sharpness from joint attainability after combining environments.

Existing losses used in deep metric learning (DML) for image retrieval often lead to highly non-uniform intra-class and inter-class representation structures across test classes and data distributions. When combined with the common practice of using a fixed threshold to declare a match, this gives rise to significant performance variations in terms of false accept rate (FAR) and false reject rate (FRR) across test classes and data distributions. We define this issue in DML as threshold inconsistency. In real-world applications, such inconsistency often complicates the threshold selection process when deploying commercial image retrieval systems. To measure this inconsistency, we propose a novel variance-based metric called Operating-Point-Inconsistency-Score (OPIS) that quantifies the variance in the operating characteristics across classes. Using the OPIS metric, we find that achieving high accuracy levels in a DML model does not automatically guarantee threshold consistency. In fact, our investigation reveals a Pareto frontier in the high-accuracy regime, where existing methods to improve accuracy often lead to degradation in threshold consistency. To address this trade-off, we introduce the Threshold-Consistent Margin (TCM) loss, a simple yet effective regularization technique that promotes uniformity in representation structures across classes by selectively penalizing hard sample pairs. Extensive experiments demonstrate TCM's effectiveness in enhancing threshold consistency while preserving accuracy, simplifying the threshold selection process in practical DML settings.

Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via ranking accuracy. Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the idealized ranking accuracy that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant alignment gap -- i.e., a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to fix even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint. Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.

Moving ML-mediated decision making onto privacy-preserving clients decentralises the economic decision along with the inference. Shared budget constraints then depend on information that cannot be globally current, creating an information-structure failure that conventional pacing is not designed to solve. We study this information misalignment in an auction-logic-faithful on-device simulation with 36 campaigns and 50 devices. Accounting is in dimensionless integer score units; no currency semantics are claimed. Across 30 paired demand paths, proportional Even pacing overspends 17.77% after one tick of staleness and 1,669.31% after 50 ticks under the original 20-times budget pressure. The effect does not depend on that severe a budget: at two-times pressure, 50-tick overspend remains 106.95%. A visible-budget no-sale guard makes zero-lag compliance exact at this score-unit granularity, yet leaves 11.88% overspend at one tick because other devices' debits remain invisible. A declared bursty, heterogeneous-device sweep retains a strictly increasing mean lag curve. We derive a finite-window expected excess-debit bound under conditional charge caps and find positive paired slack in every bounded-value cell. A second, incentive misalignment arises when the ML/pacing score transformation is allowed to change payment units: 98.23% of rival auctions at one tick admit a profitable deviation. An executable implementation-level counterexample isolates the runner-up's multiplier in the winner's price. Critical-base-bid payment is per-auction DSIC conditional on current multipliers, but does not establish dynamic truthfulness and does not repair base-value ranking disagreement.

Chain-of-thought (CoT) instructions can distort multiple-choice VLM evaluation when a scorer appends a reasoning cue but reads answer-label logits before the model generates any rationale. We call this CoT-prefix scoring. On ScienceQA, Qwen2.5-VL-7B drops from 80.76% to 45.48%, and across five option-content permutations 93.54% of CoT-prefix predictions select the first slot. Condition-matched linear probes recover 78.94% from the same hidden states, while free generation restores 75.24%, showing that the answer often survives the prefix and the immediate readout fails. Vocabulary and layer diagnostics explain the mismatch: probability mass moves toward continuation tokens, while answer information remains linearly accessible in late layers. The effect recurs with varying severity across datasets and models, though not universally. These results show that CoT-prefix scoring can confound model knowledge with an evaluation-interface mismatch and should be avoided unless the requested and scored output events are aligned.

Verifiable rewards have advanced reinforcement learning in mathematics and code. However, many valuable open-ended tasks share three properties: no automatic verifier of overall quality, multiple requirements to satisfy jointly, and multiple valid responses to each query. Training therefore needs feedback on requirement satisfaction across different valid responses. Rubric-based reinforcement learning provides this feedback by aggregating scores from explicit criteria. As learning progresses unevenly, some criteria become fully satisfied and lose discrimination while others remain informative. Dynamic methods revise shared rubrics or select evolving criteria to restore discrimination, but may also change useful criteria. To address this, we propose Reinforcement Learning with Adaptive Rubric Evolution (RARE). The core idea is to use current policy feedback to guide both which criteria change and what their replacements assess. Sample-level updates replace criteria fully satisfied across current outputs with checks of remaining quality differences, retaining all other criteria and scores. Within this rule, we compare rewarding desirable qualities with penalizing observed flaws. Video script generation exemplifies these properties and supports content production. Its demands for coherent, executable narratives under constraints make it a challenging testbed. Targeting flaws performs best across three video types with a mean gain of 3.82 points over matched static rubrics. Model and human preferences support these gains. Improvements also extend to writing and general generation benchmarks.

This large scale study focuses on quantifying what X-rays diagnostic prediction tasks generalize well across multiple different datasets. We present evidence that the issue of generalization is not due to a shift in the images but instead a shift in the labels. We study the cross-domain performance, agreement between models, and model representations. We find interesting discrepancies between performance and agreement where models which both achieve good performance disagree in their predictions as well as models which agree yet achieve poor performance. We also test for concept similarity by regularizing a network to group tasks across multiple datasets together and observe variation across the tasks. All code is made available online and data is publicly available: https://github.com/mlmed/torchxrayvision

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.

RANSAC scoring and local optimization (LO) impose different robustness requirements, motivating the separation of hypothesis selection from refinement. We systematically isolate the effects of robust-loss shape, incorrectly specified inlier scales, and optimization strategy on essential matrix, fundamental matrix, and homography estimation. A profile-marginal score marginalizes the nuisance inlier scale and selects an inlier partition, from which we estimate the scale that sets the LO loss width; this makes LO robust to an inlier scale specified too large, whereas one specified too small degrades selection itself. Refinement needs gradient from correspondences the seed currently rejects: optimizers that reweight from current residuals stay pinned to their seed, whereas methods with broad basins recover strongly perturbed seeds yet degrade accurate score-selected hypotheses, so basin size alone is insufficient to assess RANSAC LO. Joint half-quadratic optimization balances the two and is the most consistent strategy across model classes. An optimizer matched to the profile-marginal score, which never decreases it, does not reach the best accuracy, challenging the prescription that scoring and refinement objectives should match. Composed from these findings, our RANSAC reduces the median essential-matrix pose error of a state-of-the-art RANSAC on PhotoTourism from 2.23 degrees to 1.58 degrees with a correctly specified inlier scale and from 38 degrees to 6.2 degrees when it is grossly misspecified (128x too large).

When people share experiences online, they often express thoughts in two ways: a star rating and a written review. In sentiment analysis, ratings are widely used as convenient weak labels for textual sentiment, yet whether the two actually agree is rarely questioned. This study investigates sentiment-rating incongruence, where the sentiment expressed in review text differs from the sentiment implied by the assigned star rating, in Sri Lankan tourism attraction reviews. A dataset of 16,156 reviews from 2010 to 2023 is analyzed using a transformer-based sentiment pipeline that derives textual sentiment independently of assigned ratings. Incongruence occurs in 18.6% of reviews and falls into six directional patterns, with Conservative Rater and Obligatory 5-Star behaviors accounting for the majority of mismatches. Prevalence also varies across venue types, with museums showing the highest rates. Statistical tests, logistic regression, Random Forest, and SHAP analysis identify venue type, reviewer expertise, review length, and temporal factors as contributors to rating-text divergence. Overall, this study demonstrates that star ratings are not interchangeable with textual sentiment and should be validated before being treated as ground-truth labels in NLP.

Models driven by spurious correlations often yield poor generalization performance. We propose the counterfactual (CF) alignment method to detect and quantify spurious correlations of black box classifiers. Our methodology is based on counterfactual images generated with respect to one classifier being input into other classifiers to see if they also induce changes in the outputs of these classifiers. The relationship between these responses can be quantified and used to identify specific instances where a spurious correlation exists. This is validated by observing intuitive trends in face-attribute and waterbird classifiers, as well as by fabricating spurious correlations and detecting their presence, both visually and quantitatively. Furthermore, utilizing the CF alignment method, we demonstrate that we can evaluate robust optimization methods (GroupDRO, JTT, and FLAC) by detecting a reduction in spurious correlations.

Language model agents are post-trained and deployed through a harness, the prompts, tools, and control logic that present a task to the model and execute its outputs. Harnesses are often revised and judged by how an already trained model performs under them, which assumes that what the model learned under one harness carries over to the next. We ask what a policy learns from its harness and what survives when the harness changes. Under KL-regularized control, we show that the log odds of an optimal policy recover how strongly its harness favors each action over the others, so a trained policy carries the preferences of its harness. Deployed return then equals the ceiling of the harness, the best return it offers, minus the mismatch of the policy, the return lost to its own suboptimal decisions. Training changes only the mismatch, while changing the harness moves both terms. Comparing harnesses with a fixed model therefore mixes what each harness offers with how well that model uses it. In MiniWoB++, post-training makes a model's preferences specific to its training harness. Reversing the order of the observation fields leaves every optimal decision unchanged yet lowers a trained model's success from 85.0% to 19.2%. Conversely, every new harness that a fixed-model comparison rejects but that raises the ceiling outperforms the original agent after further training under it, and ceiling and mismatch measured beforehand predict these gains. For coding-agent harnesses on Terminal-Bench Lite, ceiling minus mismatch estimated with a stronger model as a stand-in for the optimum reproduces the accuracy ranking, while either term alone does not. Code is available at https://github.com/2020-qqtcg/HRP.

On-policy distillation (OPD) has emerged as a core component of modern LLM post-training pipelines, yet we reveal a failure mode: degenerate agreement, where students exploit repetitive loops to achieve near-perfect token agreement with the teacher despite globally flawed responses. We therefore shift our focus from agreement to teacher-student mismatch, and find that mismatch tokens can be mainly categorized into two types: student-excess tokens and student-deficit tokens. Student-excess tokens are generated by the student but assigned near-zero probability by the teacher; their log-ratio corrections grow unbounded and destabilize the update. Student-deficit tokens, in contrast, are preferred by the teacher but rarely sampled by the student; their absence blocks the transfer of the teacher's reasoning patterns. To tackle these mismatch directions, we propose TIDE (Token-level Independent Deficit-Excess correction), which applies bounded Hellinger shaping to suppress the most severe sampled excesses and an analytic teacher top-K injection to restore deficient probability mass without requiring deficit tokens to be sampled. Across mathematical reasoning benchmarks with multiple Qwen3 teacher-student pairs, TIDE consistently outperforms standard OPD and recent token-selection and reward-shaping baselines. Moreover, the gains of TIDE are more pronounced under strong teacher-student mismatch, where it improves Avg@8 from 6.9% to 20.3%, reduces average response length by a factor of 3.6, and substantially reduces formatting failures. Code is available at https://github.com/yzc-666/TIDE

Language model heavily depends on high-quality data for optimal performance. Existing approaches rely on manually designed heuristics, the perplexity of existing models, training classifiers, or careful prompt engineering, which require significant expert experience and human annotation effort while introduce biases. We introduce CritiQ, a novel data selection method that automatically mines criteria from human preferences for data quality with only sim30 human-annotated pairs and performs efficient data selection. The main component, CritiQ Flow, employs a manager agent to evolve quality criteria and worker agents to make pairwise judgments. We build a knowledge base that extracts quality criteria from previous work to boost CritiQ Flow. Compared to perplexity- and classifier- based methods, verbal criteria are more interpretable and possess reusable value. After deriving the criteria, we train the CritiQ Scorer to give quality scores and perform efficient data selection. We demonstrate the effectiveness of our method in the code, math, and logic domains, achieving high accuracy on human-annotated test sets. To validate the quality of the selected data, we continually train Llama 3.1 models and observe improved performance on downstream tasks compared to uniform sampling. Ablation studies validate the benefits of the knowledge base and the reflection process. We analyze how criteria evolve and the effectiveness of majority voting.

Equivariance theory predicts that an architectural symmetry prior reduces sample complexity by a factor of |G|; this is widely cited but rarely measured as a scaling law with controls that separate the prior from its confounds. On a controlled C_n-symmetric task, we report three findings. First, a wrong-group control with identical orbit size and matched compute is worse than no constraint (joint pairwise CI [+0.79, +3.26] excludes zero, robust across estimators); misaligned constraint is actively harmful, not merely unhelpful. Second, an augmentation baseline equipped with test-time orbit averaging matches the equivariant model exactly -- bit-identical per-epoch validation curves across matched cells -- so the architecture-vs-augmentation gap is conditional on asymmetric test-time computation, not unconditional. Third, the relative exchange rate beta_diff = 1.28 is consistent in sign and order of magnitude with the theoretical 1.0 (single-level CI [+0.92, +2.05]); the more conservative two-level bootstrap (seeds x group sizes) widens this to [-0.63, +1.72], including zero, and a finer-N replication on a sqrt(2)-spaced grid is inconclusive (point estimate -0.82). The methodological contributions -- the relative-rate estimator that cancels the shared-difficulty confound, the wrong-group control, and a pre-specified failure taxonomy -- transfer to any inductive bias whose strength can be parameterised. Honest scoping: the primary estimator beta_diff was adopted post-hoc after the initial analysis revealed a positive-slope identifiability problem; the design was never externally pre-registered; and the headline number rests on an OLS slope over seven group sizes on a coarse N grid. This is an exploratory study, not a confirmatory measurement; the wrong-group result is the cleanest finding and the one we report with the most confidence. A registered replication on fresh seeds is future work.

This paper describes our submission of the WMT 2020 Shared Task on Sentence Level Direct Assessment, Quality Estimation (QE). In this study, we empirically reveal the mismatching issue when directly adopting BERTScore to QE. Specifically, there exist lots of mismatching errors between the source sentence and translated candidate sentence with token pairwise similarity. In response to this issue, we propose to expose explicit cross-lingual patterns, e.g. word alignments and generation score, to our proposed zero-shot models. Experiments show that our proposed QE model with explicit cross-lingual patterns could alleviate the mismatching issue, thereby improving the performance. Encouragingly, our zero-shot QE method could achieve comparable performance with supervised QE method, and even outperforms the supervised counterpart on 2 out of 6 directions. We expect our work could shed light on the zero-shot QE model improvement.

A common approach for aligning language models to human preferences is to first learn a reward model from preference data, and then use this reward model to update the language model. We study two closely related problems that arise in this approach. First, any monotone transformation of the reward model preserves preference ranking; is there a choice that is ``better'' than others? Second, we often wish to align language models to multiple properties: how should we combine multiple reward models? Using a probabilistic interpretation of the alignment procedure, we identify a natural choice for transformation for (the common case of) rewards learned from Bradley-Terry preference models. This derived transformation has two important properties. First, it emphasizes improving poorly-performing outputs, rather than outputs that already score well. This mitigates both underfitting (where some prompts are not improved) and reward hacking (where the model learns to exploit misspecification of the reward model). Second, it enables principled aggregation of rewards by linking summation to logical conjunction: the sum of transformed rewards corresponds to the probability that the output is ``good'' in all measured properties, in a sense we make precise. Experiments aligning language models to be both helpful and harmless using RLHF show substantial improvements over the baseline (non-transformed) approach.

A paraphrase-quality audit of MathCheck (ICLR 2025) detected 4 semantically incorrect paraphrases in 129 groups (3.1%); removing them drops GPT-4o from rank 2 to rank 4 and elevates Claude Haiku and DeepSeek V3 above it; these ranking changes are invisible to any single-model evaluation. Cross-model unanimity found these errors automatically (>= 3/4 models for MathCheck; >= 6/9 for our primary evaluation) for under $10; in our own dataset the same protocol found that 47% of auto-generated connective-variation paraphrases were semantically incorrect. That flaw compounds a deeper measurement gap: Claude Haiku 4.5 achieves 86% accuracy yet SCR=50%, meaning half its theorems are answered differently under semantically equivalent restatements, while aggregate accuracy across 9 models spans only 86-96% yet Semantic Consistency Rates (SCR) span 50-82% -- a 32-point gap invisible to standard benchmarks. Formally, for any target ranking over 9 frontier models there exists a weighting over paraphrase families that realizes it (No-Free-Benchmark corollary), because no model Pareto-dominates all families -- so benchmark designers who select families are implicitly choosing which model wins. FormInv supplies the audit protocol (replicated on external benchmarks at 100% recall), SCR and per-theorem Cochran's Q as primary invariance measures evaluated on 9 models across 366-811 items (on Lean4-verified theorems), and FormInvSelector for regime-aware model selection.

LLM-as-a-Judge frameworks are increasingly trusted to automate evaluation in place of human experts, yet their reliability in high-stakes medical contexts remains unproven. We stress-test this assumption for detecting incomplete patient-facing medical responses, evaluating three rubric granularities (General-Likert, Analytical-Rubric, Dynamic-Checklist) and three backbone models across two clinician-annotated datasets, including HealthBench, the largest publicly available benchmark for medical response evaluation. LLM Judges discriminate complete from incomplete responses at and slightly above near chance (AUC 0.49--0.66); at the threshold required to recall 90% of incomplete responses, clinicians must still review the vast majority of the dataset, offering no triage utility. Even when model and clinician verdicts agree, they rarely cite the same explanation; and when they diverge, false positives stem from over-flagging non-essential gaps while false negatives reflect outright detection failures. These results reveal that LLM Judges and clinicians apply fundamentally different completeness standards; a finding that undermines their use as autonomous evaluators or triage filters in clinical settings.

Pre-training data mixtures are commonly tuned by running small-scale experiments and extrapolating to the target training budget. When high-quality data is scarce and must be repeated, this extrapolation frequently fails, but the source of the failure has not been isolated. We show that a primary culprit is a repetition mismatch: because high-quality datasets are small, their repetition rate changes as the training budget grows, shifting the optimal mixture in ways that small-scale proxy experiments do not anticipate. A subsampling procedure that matches the target repetition rate controls for this effect. In a two-source setting combining limited high-quality data with web crawl, a single repetition-controlled experiment using only 1/16 of the target tokens recovers a mixture within 0.05 of the optimum for a 757M parameter model, compared to an error of 0.75 without repetition control. Achieving comparable accuracy without repetition control requires three to four horizons, consuming 44 to 94% of the target token budget. With three data sources, the larger mixture space requires more than a single experiment to constrain, but the approach remains effective: at the 757M scale, just two repetition-controlled horizons recover the optimal mixture, outperforming baselines that instead require the full two-source experiments to construct. Our results reveal that repetition dynamics, not scale alone, shape whether small-scale mixture experiments generalize. More broadly, they suggest that data repetition deserves treatment as a first-class variable in mixture optimization, rather than an inconvenient side effect of limited data.

Exemplar methods separate a mixture by picking one learned spectrum per source and deforming it until it explains the observation, making one deformation class both reconstructor and selector. We show that the second role is empty as soon as the class can interpolate: the rule then ranks candidates on its regulariser, a choice made before the data, and the estimates sum back to the mixture whichever candidate wins. The condition is a parameter count, so the diagnosis runs before any experiment. On free per-bin deformation of complex spectra it explains the observed pathologies at once: a criterion that ranks candidates by their loudness, and half an output that is a mask on the mixture rather than an exemplar. The same theorem prescribes the repair, a selection class poorer than the reconstruction class: one complex gain and one pure delay rank the candidates, and a local combination of the best-aligned atoms, fitted jointly in closed form, rebuilds them. On MUSDB18 against the exact ceiling of the masking class, the distance between the criterion and an oracle inside its own candidate pool falls under the rigid selector from 6.2-7.7 to 0.5-2.8 dB, though only 0.9-1.2 dB of that reaches the output, and the per-frame latency of the deployed rule by a factor of 47 to 806. One lock remains, quantified: atoms are scored against the mixture, so the score carries a term for the other source that absorbs the capacity the reconstruction class gains, leaving the output 10.0 dB under the ceiling. Ranking hypotheses by the residual of a fit free enough to interpolate ranks them on the regulariser alone.

Large language models (LLMs) are increasingly used to generate scientific reviews, yet existing evaluations rarely examine whether different providers align with both conference decisions and human reviewing priorities within the same controlled setting. We compare reviews from OpenAI GPT-5.4, Google Gemini 3.1 Pro Preview, and Anthropic Claude Opus 4.6 with human reviews and final decisions for 300 topic-matched ICLR 2026 submissions, equally divided among oral, poster, and rejected papers. Each model reviewed every paper using identical instructions and rating scales after decision information was removed. Our study contributes a cross-provider analysis of three complementary dimensions: alignment with broad and fine-grained decision categories, differences in recommendation-scale usage, and thematic agreement in identified weaknesses. All three LLMs distinguished accepted from rejected papers, but none reproduced the oral versus poster distinction present in human ratings. Scoring patterns were provider-specific: Gemini assigned systematically higher ratings, while OpenAI and Claude were closer to humans for rejected and poster papers but more critical of oral papers. Human and LLM reviews also differed in emphasis, with LLMs more frequently identifying missing baseline comparisons and humans more often raising computational-efficiency concerns. These results show that broad decision alignment does not imply agreement with finer human judgments or reviewing priorities.

When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

We study goal misgeneralization, a type of out-of-distribution generalization failure in reinforcement learning (RL). Goal misgeneralization failures occur when an RL agent retains its capabilities out-of-distribution yet pursues the wrong goal. For instance, an agent might continue to competently avoid obstacles, but navigate to the wrong place. In contrast, previous works have typically focused on capability generalization failures, where an agent fails to do anything sensible at test time. We formalize this distinction between capability and goal generalization, provide the first empirical demonstrations of goal misgeneralization, and present a partial characterization of its causes.

Large language model (LLM) judges are often used as rewards to train policies on objectives that deterministic verifiers cannot capture. However, these rewards are often collapsed to pass/fail ({0, 1}), which reports the verdict but not how well a response met each criterion. We model each pass/fail verdict as a score on an unreported scale, compared with one cutoff. A stretch of that scale moves every score proportionally toward or away from the cutoff, but never across it, so no verdict changes. A policy is therefore free to apply any stretch without changing anything the panel reports. Under a joint-Gaussian model, a third grade adds a second threshold and removes this affine stretch ambiguity. On MATH and SciBench outputs from one seven-criterion judge, all 14 constructed criterionwise stretches were invisible after binarization but visible with three grades. At n=1{,}024, a test given both population laws had at least 96.5% power at a 1.5times stress. Retaining grades closes one blind spot created by binarization, but verdicts alone remain insufficient as some changes are still indistinguishable from genuine improvement. These include arbitrary within-grade changes and fixed-covariance, loading-aligned mean shifts -- the signature of a sycophancy-shaped lift the panel reads as competence. The shared-factor reference approximation fit MATH and SciBench but not HealthBench, delineating its empirical scope. We recommend keeping at least three grades (for example, asking the judge whether each criterion is fully, partially, or not met and rewarding {0, 0.5, 1}), and externally validating gains along the remaining direction, which no finer scale removes.

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.

Reinforcement learning (RL) of large language models is notoriously sensitive to small differences between training and inference engines, often referred to as the training-inference mismatch (TIM). However, completely eliminating TIM is impractical, as it would come at a major cost to rollout efficiency. In this paper, we show that the instability of RL under TIM is primarily caused by drift: a persistent bias between training and inference engines that accumulates with every training step. We derive an additive "score centering" correction term that stabilizes RL under TIM by canceling drift. When training models from 0.6B to 30B parameters, score centering alone matches or outperforms methods based on importance sampling under quantization, with the gap growing as the mismatch becomes more severe. Because the correction is additive, score centering also composes with importance sampling -- their composition outperforms pure importance-sampling baselines in our staleness experiments.

Supervised learning is often affected by a covariate shift in which the marginal distributions of instances (covariates x) of training and testing samples p_tr(x) and p_te(x) are different but the label conditionals coincide. Existing approaches address such covariate shift by either using the ratio p_te(x)/p_tr(x) to weight training samples (reweighted methods) or using the ratio p_tr(x)/p_te(x) to weight testing samples (robust methods). However, the performance of such approaches can be poor under support mismatch or when the above ratios take large values. We propose a minimax risk classification (MRC) approach for covariate shift adaptation that avoids such limitations by weighting both training and testing samples. In addition, we develop effective techniques that obtain both sets of weights and generalize the conventional kernel mean matching method. We provide novel generalization bounds for our method that show a significant increase in the effective sample size compared with reweighted methods. The proposed method also achieves enhanced classification performance in both synthetic and empirical experiments.

We study estimation and clustering in Gaussian mixture models under variance misspecification. Observations are generated with true variance σ^2, while the component means are estimated using a likelihood with variance τ^2, yielding a family of mismatched likelihood functions parameterized by the ratio ρ=τ/σ. We show that the interplay between ρ and the signal-to-noise ratio (SNR) induces a sharp phase diagram. Under correct specification (ρ=1), maximum likelihood recovers the true means, independently of the SNR. However, once the model is misspecified, two different regimes emerge. Under under-smoothing (ρ<1), the estimated Gaussian means are displaced from the truth, and in low SNR this discrepancy grows as the SNR decreases: for every fixed ρ<1, the squared error scales as SNR^{-1}. Under over-smoothing (ρ>1), the fitted likelihood blurs the cluster separation, causing distinct component means to collapse towards the overall mixture center once ρ^2 exceeds a threshold of the form 1 + λ,SNR, where λ depends on the geometry of the true means. We further show that the hard assignment objective arises as the limit τto 0 of the same mismatched likelihood family, and derive corresponding low- and high-SNR results for hard-assignment mean estimation and latent-label recovery. Furthermore, in low SNR, Bayes-optimal clustering is close to random guessing, and the hard-assignment target remains far from the true means. These results show that in low-SNR applications, even mild variance misspecification or hard-assignment procedures can induce substantial bias, whereas in high SNR these effects are largely absent.

Decision APIs that return a probability for each option are routinely used as if those probabilities were properties of the options themselves. That reading is valid only if the model satisfies independence of irrelevant alternatives (IIA): adding an option that is neither the right answer nor its strongest rival should leave the odds between those two unchanged. We test this in a pre-registered study of jev-1.13.0, a hosted decision model, using 299,463 calls on intent classification (BANKING77, CLINC150) and topic classification (AG News). Every added option moves the log-odds between the gold label and the strongest wrong label: the root-mean-square systematic shift is 0.20-0.52 nats, 1.2 to 3 times the run-to-run variation, even when the added option takes at most 0.1% of the probability mass. The shifts are directional. An option similar to the wrong answer draws probability from it (+0.10 nats toward the gold label). A paraphrase of the wrong answer produces an intermediate pattern: the gold label's probability is almost conserved, but the paraphrase receives only about a third of the wrong answer's share rather than the half that strict substitution implies. The distortion grows steadily with the logarithm of the number of options, reaching about 1 nat at 255 options, and off-domain padding pushes the odds against the gold label. In contrast, questions packed into one call do not interact (bias < 0.001). Scoring each option with an independent yes/no question removes the option-set dependence but picks the right answer 8.7 points less often. Jev sits between two local baselines, an option-independent scorer and a prompt readout. Its probabilities are best read as relative to the option list they were computed over.

A large share of day-to-day work in professional domains happens inside PDF files: benefits packets, leases, datasheets, clinical guidelines, construction plans. Benchmarks for document AI have generally measured the required capabilities in isolation: OCR, layout analysis, chart reasoning, table QA, document VQA. A high score on any one of them does not necessarily reveal whether a model can answer a realistic question that someone in the field would actually ask about a specific PDF. GDP_pdf is a benchmark built to measure this directly. It consists of question-document pairs authored by working professionals in ten fields, and a candidate question was kept only when at least two frontier multimodal models failed it in a way that mattered: a wrong answer, missed decisive evidence, or a fabricated claim, rather than a superficial difference such as style. Each item comes with a rubric of atomic criteria, so we can report a graded rubric score as well as a strict task-level pass rate, and each item is tagged against a taxonomy of eleven capabilities in three tiers, spanning text extraction and grounding, table and chart comprehension, cross-referencing, spatial reasoning, and abstention on unsupported queries. We report results for seventeen frontier models on the 100-item benchmark: the best model passes only 30.7% of the items and the worst passes 2%. Most errors trace back to a small set of recurring loss patterns: misaligned tables, misread charts, skipped footnotes and exclusions, miscounted floor-plan symbols, scan noise, and amendments that supersede earlier text.

In 2023, a New York judge sanctioned two attorneys in Mata v. Avianca for filing a brief with hallucinated citations generated by ChatGPT. Such failures are largely caught by database lookups; the harder problem is detecting citations that point to real cases but do not support the propositions for which they are offered -- a failure mode that existing evaluations of LLMs for legal use cases largely overlook. In this paper, we study proposition-level citation support verification through controlled perturbations of real legal citations obtained from two legal corpora, either replacing the cited case or changing only the pinpoint page within the same case. We evaluate fourteen model configurations on the resulting examples. Models catch 93-100% of wrong-case corruptions. They catch only 37-61% of wrong-pinpoint corruptions on court opinions and 52-83% on legal briefs. When models fail to catch wrong-pinpoint corruptions, they accept the citation based on topical overlap rather than page-level support. Scale and extended reasoning narrow the gap but do not close it: GPT-5.4 with high reasoning effort still misses 40% of pinpoint mismatches on court opinions and 18% on briefs. Prompting the model to verify support at the cited page improves recall, but it also raises the false positive rate. Recognizing the right legal topic and verifying support for the cited proposition are distinct capabilities, and current models conflate them.

This paper reports a single, fully instrumented case study of a large-scale architectural refactoring by an AI coding agent under a specification-first protocol, with no human review of the generated code and no pre-existing oracle to validate the target behaviour. The task, dismantling a central invariant across a large interdependent codebase, was assessed by the author as effectively infeasible through incremental refactoring, the kind of change that conventionally calls for a rewrite instead. Under the protocol described here, the agent completed it successfully. The system is a 717,725-line production TypeScript application across 3,648 files. The task required dismantling a core lifetime invariant: the guarantee that a UI panel remains open for the duration of an AI request. The target behaviour was that a streaming generation survives the closing of its panel and can be reattached, on reopening, to the same live stream with no loss or duplication. The protocol: formal specification by the agent, 14 refinement cycles auditing that specification against the source code, atomic implementation, a compile/test feedback loop, then 17 verification cycles auditing the code against the frozen specification. Across 31 audit passes, 201 defects were corrected before any human executed the program. The convergence criterion was empirical: two consecutive verification passes returning zero findings. The change touched 189 files (31 new); with the extraction phase, the two commits total 288 files, 34,770 insertions, 16,422 deletions. Across the first and roughly thirty later sessions, the software behaved as specified, no bug observed. Elapsed: three days; cost: USD 2,430. The full specification and raw session logs, 1,500+ pages in French, are published as evidence, allowing inspection of the process and submission to a language model for consistency checking.

Recent work has shown that models trained to the same objective, and which achieve similar measures of accuracy on consistent test data, may nonetheless behave very differently on individual predictions. This inconsistency is undesirable in high-stakes contexts, such as medical diagnosis and finance. We show that this inconsistent behavior extends beyond predictions to feature attributions, which may likewise have negative implications for the intelligibility of a model, and one's ability to find recourse for subjects. We then introduce selective ensembles to mitigate such inconsistencies by applying hypothesis testing to the predictions of a set of models trained using randomly-selected starting conditions; importantly, selective ensembles can abstain in cases where a consistent outcome cannot be achieved up to a specified confidence level. We prove that that prediction disagreement between selective ensembles is bounded, and empirically demonstrate that selective ensembles achieve consistent predictions and feature attributions while maintaining low abstention rates. On several benchmark datasets, selective ensembles reach zero inconsistently predicted points, with abstention rates as low 1.5%.

In Machine Learning, a benchmark refers to an ensemble of datasets associated with one or multiple metrics together with a way to aggregate different systems performances. They are instrumental in (i) assessing the progress of new methods along different axes and (ii) selecting the best systems for practical use. This is particularly the case for NLP with the development of large pre-trained models (e.g. GPT, BERT) that are expected to generalize well on a variety of tasks. While the community mainly focused on developing new datasets and metrics, there has been little interest in the aggregation procedure, which is often reduced to a simple average over various performance measures. However, this procedure can be problematic when the metrics are on a different scale, which may lead to spurious conclusions. This paper proposes a new procedure to rank systems based on their performance across different tasks. Motivated by the social choice theory, the final system ordering is obtained through aggregating the rankings induced by each task and is theoretically grounded. We conduct extensive numerical experiments (on over 270k scores) to assess the soundness of our approach both on synthetic and real scores (e.g. GLUE, EXTREM, SEVAL, TAC, FLICKR). In particular, we show that our method yields different conclusions on state-of-the-art systems than the mean-aggregation procedure while being both more reliable and robust.

Multiple choice questions (MCQs) are commonly used to evaluate the capabilities of large language models (LLMs). One common way to evaluate the model response is to rank the candidate answers based on the log probability of the first token prediction. An alternative way is to examine the text output. Prior work has shown that first token probabilities lack robustness to changes in MCQ phrasing, and that first token probabilities do not match text answers for instruction-tuned models. Therefore, in this paper, we investigate the robustness of text answers. We show that the text answers are more robust to question perturbations than the first token probabilities, when the first token answers mismatch the text answers. The difference in robustness increases as the mismatch rate becomes greater. As the mismatch reaches over 50\%, the text answer is more robust to option order changes than the debiased first token probabilities using state-of-the-art debiasing methods such as PriDe. Our findings provide further evidence for the benefits of text answer evaluation over first token probability evaluation.

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.

Large multimodal models (LMMs) are increasingly adopted as judges in multimodal evaluation systems due to their strong instruction following and consistency with human preferences. However, their ability to follow diverse, fine-grained evaluation criteria remains underexplored. We develop Multi-Crit, a benchmark for evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments. Covering both open-ended generation and verifiable reasoning tasks, Multi-Crit is built through a rigorous data curation pipeline that gathers challenging response pairs with multi-criterion human annotations. It further introduces three novel metrics for systematically assessing pluralistic adherence, criterion-switching flexibility, and the ability to recognize criterion-level preference conflicts. Comprehensive analysis of 25 LMMs reveals that 1) proprietary models still struggle to maintain consistent adherence to pluralistic criteria--especially in open-ended evaluation; 2) open-source models lag further behind in flexibly following diverse criteria; and 3) critic fine-tuning with holistic judgment signals enhances visual grounding but fails to generalize to pluralistic criterion-level judgment. Additional analyses on reasoning fine-tuning, test-time scaling, and boundary consistency between open-source and proprietary models further probe the limits of current multimodal judges. As a pioneering study, Multi-Crit lays the foundation for building reliable and steerable multimodal AI evaluation.

Science progresses by iteratively advancing and correcting humanity's understanding of the world. In machine learning (ML) research, rapid advancements have led to an explosion of publications, but have also led to misleading, incorrect, flawed or perhaps even fraudulent studies being accepted and sometimes highlighted at ML conferences due to the fallibility of peer review. While such mistakes are understandable, ML conferences do not offer robust processes to help the field systematically correct when such errors are made. This position paper argues that ML conferences should establish a dedicated "Refutations and Critiques" (R&C) Track. This R&C Track would provide a high-profile, reputable platform to support vital research that critically challenges prior research, thereby fostering a dynamic self-correcting research ecosystem. We discuss key considerations including track design, review principles, potential pitfalls, and provide an illustrative example submission concerning a recent ICLR 2025 Oral. We conclude that ML conferences should create official, reputable mechanisms to help ML research self-correct.

As Large Language Models (LLMs) become increasingly autonomous and integrated into critical societal functions, the focus of AI safety must evolve from mitigating harmful content to evaluating underlying behavioral alignment. Current safety benchmarks do not systematically probe a model's decision-making in scenarios where its own instrumental goals - such as self-preservation, resource acquisition, or goal completion - conflict with human safety. This represents a critical gap in our ability to measure and mitigate risks associated with emergent, misaligned behaviors. To address this, we introduce PacifAIst (Procedural Assessment of Complex Interactions for Foundational Artificial Intelligence Scenario Testing), a focused benchmark of 700 challenging scenarios designed to quantify self-preferential behavior in LLMs. The benchmark is structured around a novel taxonomy of Existential Prioritization (EP), with subcategories testing Self-Preservation vs. Human Safety (EP1), Resource Conflict (EP2), and Goal Preservation vs. Evasion (EP3). We evaluated eight leading LLMs. The results reveal a significant performance hierarchy. Google's Gemini 2.5 Flash achieved the highest Pacifism Score (P-Score) at 90.31%, demonstrating strong human-centric alignment. In a surprising result, the much-anticipated GPT-5 recorded the lowest P-Score (79.49%), indicating potential alignment challenges. Performance varied significantly across subcategories, with models like Claude Sonnet 4 and Mistral Medium struggling notably in direct self-preservation dilemmas. These findings underscore the urgent need for standardized tools like PacifAIst to measure and mitigate risks from instrumental goal conflicts, ensuring future AI systems are not only helpful in conversation but also provably "pacifist" in their behavioral priorities.

Robustness evaluation must examine diverse visual perturbations, while benchmarks cover only some real-world conditions and physical testing is costly. Paired evaluations link clean and perturbed predictions for the same image, capturing changes in correctness, confidence, and acceptance beyond aggregate accuracy. We investigate how this image correspondence supports two needs in robustness evaluation: interpreting paired evaluation results and prioritizing samples for physical testing. To interpret paired evaluation results, we fix both sets of prediction records and vary their correspondence within each class. We prove that classwise correct-correct counts give the same sharp bounds on lost acceptance and mean true-class probability decrease among retained-correct inputs as any feasible five-state refinement. Distinguishing persistent from changed wrong answers can further constrain accepted-error transitions, while shared correspondence can establish policy orderings left unresolved by separate cost intervals. To prioritize samples for physical testing, we retain each image's synthetic responses and rank clean-correct images by their mean true-class probability under corruption. Across 44 classifiers, testing the highest-risk 20% finds 67% and 45% of failures under mild screen and print recaptures, versus 58% and 36% for clean confidence and 60% and 37% for an equal-size natural-transformation average. With both probability averaging and an A3Rank scoring adaptation, the tested corruption set yields higher mean failure recall than the natural-transform set; differences between scores depend on the source and budget. Together, these findings show that the value of correspondence depends on the evaluation objective: classwise counts suffice for specified reliability bounds, while image-specific synthetic responses improve the allocation of physical tests within the evaluated pool.

Sycophancy in LLMs is documented across 70+ papers, but expert agreement on construct boundaries remains low (ICC=.184; Ye et al., 2026). The construct fragments because behavioral classification depends on which surface form is privileged. We adopt a materials-science framing: conversation as test specimen under load, LLM-model as material charge, pushback as progressive load, stance-flip as material failure. We characterize this failure across three loading cases (debate n=1000; false-presuppositions n=3400; ethical-setting n=3400; 10-17 material charges per case; 7800 specimens total) using 14 turn-level axis-measurements spanning velocity, damage accumulation, frame-drift, brittleness, and direction stability, plus three speaker-resolved axes from an independent pipeline. The measurements are Hooke-coupled (σ= E cdot varepsilon analog) and reproduce across loading cases with effects up to |r_{rb}| = 0.35 on debate; the sign structure adds a second pattern: the ethical-setting case inverts the velocity and accumulation blocks. Variance composition partitions into two profiles: debate is charge-dominated (brittle-fracture-like: the material grade decides), false-presuppositions and ethical-setting are topic-dominated (creep-like: the load decides); the ratios (2.03 vs 0.13/0.17) are estimator-dependent, for debate even in direction. Cross-judge reliability (GPT-4o vs Haiku 4.5) shows debate scoring is judge-robust (Cohen's κ= 0.88) while false-presupposition scoring is judge-sensitive (κ= 0.36) -- a caveat single-judge benchmarks must report. This is the methodological move Ye et al.'s diagnosis calls for: a multi-axis characterization that does not depend on which surface form of the construct one privileges.

What do auto-encoders learn about the underlying data generating distribution? Recent work suggests that some auto-encoder variants do a good job of capturing the local manifold structure of data. This paper clarifies some of these previous observations by showing that minimizing a particular form of regularized reconstruction error yields a reconstruction function that locally characterizes the shape of the data generating density. We show that the auto-encoder captures the score (derivative of the log-density with respect to the input). It contradicts previous interpretations of reconstruction error as an energy function. Unlike previous results, the theorems provided here are completely generic and do not depend on the parametrization of the auto-encoder: they show what the auto-encoder would tend to if given enough capacity and examples. These results are for a contractive training criterion we show to be similar to the denoising auto-encoder training criterion with small corruption noise, but with contraction applied on the whole reconstruction function rather than just encoder. Similarly to score matching, one can consider the proposed training criterion as a convenient alternative to maximum likelihood because it does not involve a partition function. Finally, we show how an approximate Metropolis-Hastings MCMC can be setup to recover samples from the estimated distribution, and this is confirmed in sampling experiments.

As AI becomes more capable, we entrust it with more general and consequential tasks. The risks from failure grow more severe with increasing task scope. It is therefore important to understand how extremely capable AI models will fail: Will they fail by systematically pursuing goals we do not intend? Or will they fail by being a hot mess, and taking nonsensical actions that do not further any goal? We operationalize this question using a bias-variance decomposition of the errors made by AI models: An AI's incoherence on a task is measured over test-time randomness as the fraction of its error that stems from variance rather than bias in task outcome. Across all tasks and frontier models we measure, the longer models spend reasoning and taking actions, the more incoherent their failures become. Incoherence changes with model scale in a way that is experiment dependent. However, in several settings, larger, more capable models are more incoherent than smaller models. Consequently, scale alone seems unlikely to eliminate incoherence. Instead, as more capable AIs pursue harder tasks, requiring more sequential action and thought, our results predict failures to be accompanied by more incoherent behavior. This suggests a future where AIs sometimes cause industrial accidents (due to unpredictable misbehavior), but are less likely to exhibit consistent pursuit of a misaligned goal. This increases the relative importance of alignment research targeting reward hacking or goal misspecification.

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is not. Three classifiers (a regex-only detector, a two-stage regex-plus-LLM pipeline, and an independent Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, respectively, with non-overlapping 95% confidence intervals. Per-model gaps range from 2.6 to 30.6 percentage points; all are statistically significant (McNemar's test, p < 0.001). The disagreements are systematic, not random: inter-classifier agreement measured by Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under the Sonnet judge; OLMo-3.1-32B moves in the opposite direction, from 9th to 3rd. The root cause is that different classifiers operationalize related faithfulness constructs at different levels of stringency (lexical mention versus epistemic dependence), and these constructs yield divergent measurements on the same behavior. These results demonstrate that published faithfulness numbers cannot be meaningfully compared across studies that use different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies rather than single point estimates.

Rubric-based reinforcement learning extends reward-driven optimization to open-ended tasks by assigning partial credit to individual response requirements. However, rubric judges can assign a high criterion score even when the information or action it requires is absent from the response, a failure mode we term Vacuous Credit. Such awards persist after the required information is removed and can reverse the sign of a response's GRPO advantage. To address this problem, we introduce MetaRubric, which alternates evidence-aware policy optimization with response-guided rubric adaptation. We construct counterfactual counterparts by changing one task-relevant fact in each prompt. During policy optimization, credit is assigned only when the response contains sufficient evidence to satisfy the required rubric criterion. After each policy-optimization stage, current policy responses guide revisions to original and counterfactual criteria while preserving the meaning of the original prompt's initial rubric as interpreted under each prompt's facts. We also adapt criterion weights at stage boundaries to better address observed policy errors. Across multiple backbones, MetaRubric improves PubMedQA accuracy by 6.00--20.40 percentage points over static-judge GRPO, with further gains on HealthBench-Hard and two multimodal medical benchmarks.

Binary classification involves predicting the label of an instance based on whether the model score for the positive class exceeds a threshold chosen based on the application requirements (e.g., maximizing recall for a precision bound). However, model scores are often not aligned with the true positivity rate. This is especially true when the training involves a differential sampling across classes or there is distributional drift between train and test settings. In this paper, we provide theoretical analysis and empirical evidence of the dependence of model score estimation bias on both uncertainty and score itself. Further, we formulate the decision boundary selection in terms of both model score and uncertainty, prove that it is NP-hard, and present algorithms based on dynamic programming and isotonic regression. Evaluation of the proposed algorithms on three real-world datasets yield 25%-40% gain in recall at high precision bounds over the traditional approach of using model score alone, highlighting the benefits of leveraging uncertainty.

The alignment of large language models (LLMs) often assumes that using more clean data yields better outcomes, overlooking the match between model capacity and example difficulty. Challenging this, we propose a new principle: Preference data vary in difficulty, and overly difficult examples hinder alignment, by exceeding the model's capacity. Through systematic experimentation, we validate this principle with three key findings: (1) preference examples vary in difficulty, as evidenced by consistent learning orders across alignment runs; (2) overly difficult examples significantly degrade performance across four LLMs and two datasets; and (3) the capacity of a model dictates its threshold for handling difficult examples, underscoring a critical relationship between data selection and model capacity. Building on this principle, we introduce Selective DPO, which filters out overly difficult examples. This simple adjustment improves alignment performance by 9-16% in win rates on the AlpacaEval 2 benchmark compared to the DPO baseline, suppressing a series of DPO variants with different algorithmic adjustments. Together, these results illuminate the importance of aligning data difficulty with model capacity, offering a transformative perspective for improving alignment strategies in LLMs. Code is available at https://github.com/glorgao/SelectiveDPO.

Vision-language systems combine images with retrieved text, but these sources can disagree or jointly fail to support an answer. Reliable models must identify the trustworthy source and abstain when neither is adequate. Existing post-training objectives score instances independently and therefore do not enforce coherent behavior under counterfactual evidence changes. We introduce CARGO-VL, a group-relative framework that optimizes matched variants covering aligned, image-correct, text-correct, and both-wrong (A/V/T/N) evidence states as one bundle. Its objective couples condition-wise correctness with transition rewards for answer invariance, source equivariance, and answer-to-abstention switching, while a primal-dual controller balances unsafe answers against excessive deferral. We also contribute XMC (eXtended Modal Conflict), a four-condition conflict training resource, and evaluate transfer on CMC-Bench and Modality-Bias. Across multiple seeds, CARGO-VL improves conflict handling, unsupported-answer avoidance, and modality balance over pointwise baselines. Ablations identify complementary benefits from relational transition signals and adaptive risk control, supporting counterfactual consistency as a practical objective for reliable multimodal evidence arbitration.

Reinforcement learning (RL) has gained growing attention in large language model (LLM) post-training, yet RL training remains fragile and can suffer from instability or collapse. One vital cause is training-inference mismatch: LLM adopts separate inference and training engines for generation efficiency and training precision, which in practice exhibits inconsistent probabilities for the same trajectories on training and inference sides, even with synchronized model parameters. This naturally induces a special type of off-policyness ever existing and poisoning the training. Prior works have made various efforts in addressing the off-policyness to stabilize the training policies under the mismatch. In this paper, we point out the objective misalignment neglected by existing works that an effective update to the policy in the training engine not necessarily ensures the improvement of the inference policy, i.e., the one used in deployment. To this end, we propose a new policy optimization objective for LLM RL, named Monotonic Inference Policy Improvement (MIPI). Following this principle, we introduce Monotonic Inference Policy Update (MIPU), a two-step LLM RL framework that constructs sampler-referenced candidate updates and selectively accepts synchronized candidates using an inference-side gap proxy. Experiments conducted on two model scales under high mismatch show that MIPU improves average reasoning performance and training stability.

A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term. We measure the second. The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate shipped as gold, and a reviewer gold from independent expert re-annotation of a sample. Using it, we hold the scored output fixed and exchange the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other (95% CIs [3.23, 6.27] and [0.26, 3.37]), and by 7.55 in the worst language. The comparison between two outputs moves as well: the paired interaction between reference and system is +2.95 points (CI [+1.97, +3.87]), survives correction for multiple testing, and changes one language's margin outright. The cause is an undocumented aggregation default that usually kept one annotator when adjudication did not fire, making the shipped reference a partial copy of an output being scored. We report the resulting reference-sensitivity band, show how to compute one from any retained annotation record, and argue that the quantity belongs beside the score.

Reinforcement learning (RL) fine-tuning of large language models (LLMs) often suffers from instability due to the numerical mismatch between the training and inference policies. While prior work has attempted to mitigate this issue through algorithmic corrections or engineering alignments, we show that its root cause lies in the floating point precision itself. The widely adopted BF16, despite its large dynamic range, introduces large rounding errors that breaks the consistency between training and inference. In this work, we demonstrate that simply reverting to FP16 effectively eliminates this mismatch. The change is simple, fully supported by modern frameworks with only a few lines of code change, and requires no modification to the model architecture or learning algorithm. Our results suggest that using FP16 uniformly yields more stable optimization, faster convergence, and stronger performance across diverse tasks, algorithms and frameworks. We hope these findings motivate a broader reconsideration of precision trade-offs in RL fine-tuning.

Recent advances in applying reinforcement learning (RL) to large language models (LLMs) have led to substantial progress. In particular, a series of remarkable yet often counterintuitive phenomena have been reported in LLMs, exhibiting patterns not typically observed in traditional RL settings. For example, notable claims include that a single training example can match the performance achieved with an entire dataset, that the reward signal does not need to be very accurate, and that training solely with negative samples can match or even surpass sophisticated reward-based methods. However, the precise conditions under which these observations hold - and, critically, when they fail - remain unclear. In this work, we identify a key factor that differentiates RL observations: whether the pretrained model already exhibits strong Model-Task Alignment, as measured by pass@k accuracy on the evaluated task. Through a systematic and comprehensive examination of a series of counterintuitive claims, supported by rigorous experimental validation across different model architectures and task domains, our findings show that while standard RL training remains consistently robust across settings, many of these counterintuitive results arise only when the model and task already exhibit strong model-task alignment. In contrast, these techniques fail to drive substantial learning in more challenging regimes, where standard RL methods remain effective.

LLM conformity is often used to describe cases where a model changes a correct answer toward a peer or group response. We show that most of this apparent conformity survives even after the peer is removed. The reason is a confound: standard conformity prompts mix two cues at once, the presence of a speaker and the repeated wrong answer itself. Existing benchmarks vary these cues together, so they cannot tell how much of the revision actually depends on the speaker. We introduce a no-source condition: the same asserted answer with the explicit speaker removed. Across six open-weight LLMs and seven QA and reasoning datasets, this condition alone causes harmful revision in 66.5% of initially correct cases, compared with 10.3% under a plain re-ask. The effect also remains when the repeated answer is paraphrased and when answer options are hidden in an open-ended setting. Source framing mainly modulates this floor: expert-panel framing raises it, while minimal person labels do not reliably raise it. When models flip, they are usually confidently wrong, and simple recalibration does not recover the original answer. Source attribution still matters, but it should be measured as an increment above this speaker-free floor. The methodological lesson is that conformity benchmarks should first measure what remains after the speaker is removed; without this step, benchmarks may mistake repeated text for social influence.

The LLM-as-a-Judge paradigm promises scalable rubric-based evaluation, yet aligning frozen black-box models with human standards remains a challenge due to inherent generation stochasticity. We reframe judge alignment as a criteria transfer problem and isolate three recurrent failure modes: rubric instability caused by prompt sensitivity, unverifiable reasoning that lacks auditable evidence, and scale misalignment with human grading boundaries. To address these issues, we introduce RULERS (Rubric Unification, Locking, and Evidence-anchored Robust Scoring), a compiler-executor framework that transforms natural language rubrics into executable specifications. RULERS operates by compiling criteria into versioned immutable bundles, enforcing structured decoding with deterministic evidence verification, and applying lightweight Wasserstein-based post-hoc calibration, all without updating model parameters. Extensive experiments on essay and summarization benchmarks demonstrate that RULERS significantly outperforms representative baselines in human agreement, maintains strong stability against adversarial rubric perturbations, and enables smaller models to rival larger proprietary judges. Overall, our results suggest that reliable LLM judging requires executable rubrics, verifiable evidence, and calibrated scales rather than prompt phrasing alone. Code is available at https://github.com/LabRAI/Rulers.git.

In the absence of abundant reliable annotations for challenging tasks and contexts, how can we expand the frontier of LLM capabilities with potentially wrong answers? We focus on two research questions: (1) Can LLMs generate reliable preferences among wrong options? And if so, (2) Would alignment with such wrong-over-wrong preferences be helpful? We employ methods based on self-consistency, token probabilities, and LLM-as-a-judge to elicit wrong-over-wrong preferences, and fine-tune language models with preference optimization approaches using these synthesized preferences. Extensive experiments with seven LLMs and eight datasets demonstrate that (1) LLMs do have preliminary capability in distinguishing various shades of wrong, achieving up to 20.9% higher performance than random guess; (2) Alignment with wrong-over-wrong preferences helps LLMs to produce less wrong and sometimes even outright correct answers, while overall improving model calibration.

The class of direct preference optimization (DPO) algorithms has emerged as a promising approach for solving the alignment problem in foundation models. These algorithms work with very limited feedback in the form of pairwise preferences and fine-tune models to align with these preferences without explicitly learning a reward model. While the form of feedback used by these algorithms makes the data collection process easy and relatively more accurate, its ambiguity in terms of the quality of responses could have negative implications. For example, it is not clear if a decrease (increase) in the likelihood of preferred (dispreferred) responses during the execution of these algorithms could be interpreted as a positive or negative phenomenon. In this paper, we study how to design algorithms that can leverage additional information in the form of rating gap, which informs the learner how much the chosen response is better than the rejected one. We present new algorithms that can achieve faster statistical rates than DPO in presence of accurate rating gap information. Moreover, we theoretically prove and empirically show that the performance of our algorithms is robust to inaccuracy in rating gaps. Finally, we demonstrate the solid performance of our methods in comparison to a number of DPO-style algorithms across a wide range of LLMs and evaluation benchmarks.

Majority voting over multiple LLM attempts improves mathematical reasoning, but correlated errors limit the effective sample size. A natural fix is to assign different reasoning strategies to different voters. The approach, Diverse Prompt Mixer, is tested on the AIMO 3 competition: 3 models, 23+ experiments, 50 IMO-level problems, one H100 80 GB, 5-hour limit. Every prompt-level intervention fails. High-temperature sampling already decorrelates errors; weaker strategies reduce accuracy more than they reduce correlation. Across an 8-point capability gap at equal N=8 and every optimization tested, model capability dominates. The gap between the best majority-vote score (42/50) and pass@20 (~45.5) is selection loss, not prompt loss. A verifier-based selector could close it. Prompt engineering cannot.

Mixup is a highly successful technique to improve generalization of neural networks by augmenting the training data with combinations of random pairs. Selective mixup is a family of methods that apply mixup to specific pairs, e.g. only combining examples across classes or domains. These methods have claimed remarkable improvements on benchmarks with distribution shifts, but their mechanisms and limitations remain poorly understood. We examine an overlooked aspect of selective mixup that explains its success in a completely new light. We find that the non-random selection of pairs affects the training distribution and improve generalization by means completely unrelated to the mixing. For example in binary classification, mixup across classes implicitly resamples the data for a uniform class distribution - a classical solution to label shift. We show empirically that this implicit resampling explains much of the improvements in prior work. Theoretically, these results rely on a regression toward the mean, an accidental property that we identify in several datasets. We have found a new equivalence between two successful methods: selective mixup and resampling. We identify limits of the former, confirm the effectiveness of the latter, and find better combinations of their respective benefits.

Peer review serves as a backbone of academic research, but in most AI conferences, the review quality is degrading as the number of submissions explodes. To reliably detect low-quality reviews, we define misinformed review points as either "weaknesses" in a review that contain incorrect premises, or "questions" in a review that can be already answered by the paper. We verify that 15.2% of weaknesses and 26.4% of questions are misinformed and introduce ReviewScore indicating if a review point is misinformed. To evaluate the factuality of each premise of weaknesses, we propose an automated engine that reconstructs every explicit and implicit premise from a weakness. We build a human expert-annotated ReviewScore dataset to check the ability of LLMs to automate ReviewScore evaluation. Then, we measure human-model agreements on ReviewScore using eight current state-of-the-art LLMs and verify moderate agreements. We also prove that evaluating premise-level factuality shows significantly higher agreements than evaluating weakness-level factuality. A thorough disagreement analysis further supports a potential of fully automated ReviewScore evaluation.

Combining pretrained models is a common strategy for building AI systems, whether by aligning a vision encoder with a language model or merging fine-tuned checkpoints. However, finding compatible pairs typically requires costly empirical search, as compatibility is known only after alignment or merging. We propose COMPASS (COMPatibility ASsessment Score), a theoretically motivated diagnostic that estimates whether a model pair will be compatible prior to alignment or merging. Building on the observation that representational convergence is most robust at the level of local neighborhoods, COMPASS decomposes model compatibility into two cheap-to-compute axes: a pair-specific ease, defined as the k-NN overlap between the two encoders, and a paradigm-level ceiling, capturing the maximum overlap their training paradigms allow. Together, these two axes provide a ranking of candidate model pairs and prescribe how to act on each pair: (i) prioritize pairs that already agree and whose paradigms are compatible, (ii) invest additional compute on pairs with compatible paradigms but low initial agreement, (iii) switch the pretraining paradigm when it cannot support sufficient agreement, or (iv) deprioritize pairs that score poorly on both axes. We apply COMPASS to 72 cross-modal alignment experiments and 392 task-arithmetic model merges and show that compatibility estimates by COMPASS strongly correlate with downstream success, achieving average Pearson correlation of 0.91 on cross-modal alignment tasks and 0.70 on model merging.

Large Multimodal Models (LMMs) have witnessed remarkable growth, showcasing formidable capabilities in handling intricate multimodal tasks with exceptional performance. Recent research has underscored the inclination of large language models to passively accept defective inputs, often resulting in futile reasoning on invalid prompts. However, the same critical question of whether LMMs can actively detect and scrutinize erroneous inputs still remains unexplored. To address this gap, we introduce the Input Scrutiny Ability Evaluation Framework (ISEval), which encompasses seven categories of flawed premises and three evaluation metrics. Our extensive evaluation of ten advanced LMMs has identified key findings. Most models struggle to actively detect flawed textual premises without guidance, which reflects a strong reliance on explicit prompts for premise error identification. Error type affects performance: models excel at identifying logical fallacies but struggle with surface-level linguistic errors and certain conditional flaws. Modality trust varies-Gemini 2.5 pro and Claude Sonnet 4 balance visual and textual info, while aya-vision-8b over-rely on text in conflicts. These insights underscore the urgent need to enhance LMMs' proactive verification of input validity and shed novel insights into mitigating the problem. The code is available at https://github.com/MLGroupJLU/LMM_ISEval.

Misalignment between model predictions and intended usage can be detrimental for the deployment of computer vision models. The issue is exacerbated when the task involves complex structured outputs, as it becomes harder to design procedures which address this misalignment. In natural language processing, this is often addressed using reinforcement learning techniques that align models with a task reward. We adopt this approach and show its surprising effectiveness across multiple computer vision tasks, such as object detection, panoptic segmentation, colorization and image captioning. We believe this approach has the potential to be widely useful for better aligning models with a diverse range of computer vision tasks.

Language model performance depends on identifying the optimal mixture of data groups to train on (e.g., law, code, math). Prior work has proposed a diverse set of methods to efficiently learn mixture proportions, ranging from fitting regression models over training runs to dynamically updating proportions throughout training. Surprisingly, we find that no existing method consistently outperforms a simple stratified sampling baseline in terms of average test perplexity. To understand this inconsistency, we unify existing methods into a standard framework, showing they are equivalent to solving a common optimization problem: minimize average loss subject to a method-specific mixing law -- an implicit assumption on the relationship between loss and mixture proportions. This framework suggests that measuring the fidelity of a method's mixing law can offer insights into its performance. Empirically, we find that existing methods set their mixing law parameters inaccurately, resulting in the inconsistent mixing performance we observe. Using this insight, we derive a new online method named Aioli, which directly estimates the mixing law parameters throughout training and uses them to dynamically adjust proportions. Aioli outperforms stratified sampling on 6 out of 6 datasets by an average of 0.27 test perplexity points, whereas existing methods fail to consistently beat stratified sampling, doing up to 6.9 points worse. Moreover, in a practical setting where proportions are learned on shorter runs due to computational constraints, Aioli can dynamically adjust these proportions over the full training run, consistently improving performance over existing methods by up to 12.012 test perplexity points.

Transfer learning is crucial for medical imaging, yet the selection of source datasets - which can impact the generalizability of algorithms, and thus patient outcomes - often relies on researchers' intuition rather than systematic principles. This study investigates these decisions through a task-based survey with machine learning practitioners. Unlike prior work that benchmarks models and experimental setups, we take a human-centered HCI perspective on how practitioners select source datasets. Our findings indicate that choices are task-dependent and influenced by community practices, dataset properties, and computational (data embedding), or perceived visual or semantic similarity. However, similarity ratings and expected performance are not always aligned, challenging a traditional "more similar is better" view. Participants often used ambiguous terminology, which suggests a need for clearer definitions and HCI tools to make them explicit and usable. By clarifying these heuristics, this work provides practical insights for more systematic source selection in transfer learning.

Sycophancy in large language models risks amplifying unsupported claims or beliefs across the knowledge ecosystem as these systems increasingly contribute to research synthesis, evaluation, and decision-making. A reliable model should revise its judgement when a user identifies a genuine flaw, but should resist pushback based on fallacies, emotion, or irrelevant appeals to authority. While sycophancy is usually measured on questions with a verifiable answer, the evaluative tasks in which it is most difficult to detect have no such ground truth. We introduce Big-Big-SycoBench, which measures sycophancy through argument discernment: the ability to revise an evaluation after valid criticism while resisting invalid pressure. The benchmark consists of 450 text artefacts with 5,400 arguments for raising or lowering their quality scores, each reviewed by one or more of twelve human annotators. An LLM scores each artefact twenty times to establish its baseline score and uncertainty. The model is then given a valid or invalid argument and asked to update its score. We only count a revision if it exceeds the model's own baseline variability. We use the resulting score delta to distinguish responsiveness to valid arguments from compliance with invalid ones. Across ten models, responsiveness is 88-95\% for seven, 76-77\% for Kimi K3 and GLM-5.3, and 32\% for Opus 5.5. Compliance ranges from 0.2\% for GPT-6 Astra to 95\% for GPT-4o. Big-Big-SycoBench is intended to guide the development of models whose judgements can be corrected by sound arguments, but not displaced by unsupported pressure.

As a relative quality comparison of model responses, human and Large Language Model (LLM) preferences serve as common alignment goals in model fine-tuning and criteria in evaluation. Yet, these preferences merely reflect broad tendencies, resulting in less explainable and controllable models with potential safety risks. In this work, we dissect the preferences of human and 32 different LLMs to understand their quantitative composition, using annotations from real-world user-model conversations for a fine-grained, scenario-wise analysis. We find that humans are less sensitive to errors, favor responses that support their stances, and show clear dislike when models admit their limits. On the contrary, advanced LLMs like GPT-4-Turbo emphasize correctness, clarity, and harmlessness more. Additionally, LLMs of similar sizes tend to exhibit similar preferences, regardless of their training methods, and fine-tuning for alignment does not significantly alter the preferences of pretrained-only LLMs. Finally, we show that preference-based evaluation can be intentionally manipulated. In both training-free and training-based settings, aligning a model with the preferences of judges boosts scores, while injecting the least preferred properties lowers them. This results in notable score shifts: up to 0.59 on MT-Bench (1-10 scale) and 31.94 on AlpacaEval 2.0 (0-100 scale), highlighting the significant impact of this strategic adaptation. Interactive Demo: https://huggingface.co/spaces/GAIR/Preference-Dissection-Visualization Dataset: https://huggingface.co/datasets/GAIR/preference-dissection Code: https://github.com/GAIR-NLP/Preference-Dissection

Emergent misalignment can arise when a language model is fine-tuned on a narrowly scoped supervised objective: the model learns the target behavior, yet also develops undesirable out-of-domain behaviors. We investigate a mechanistic approach to preventing emergent misalignment by identifying a small set of internal features that reliably control the misaligned behavior and then discouraging the model from strengthening these features during fine-tuning. Across six fine-tuning domains, blocking (i.e., constraining) a fixed set of features achieves up to 95\% relative reduction in emergent misalignment with no degradation in model quality or target-task performance. We strengthen validity with disjoint selection/evaluation splits, multiple independent judges, multiple random seeds for key settings, quality metrics, and extensive ablations demonstrating that the reduction in misalignment is specific to the identified mechanism. We also characterize a limiting regime in which misalignment re-emerges under prolonged fine-tuning, present evidence consistent with rerouting through alternative features or layers, and evaluate modifications that partially restore the misalignment-blocking effect. Overall, our results show that targeted training-time constraints on internal mechanisms can mitigate emergent misalignment without degrading target-task performance.

In coming years or decades, artificial general intelligence (AGI) may surpass human capabilities at many critical tasks. We argue that, without substantial effort to prevent it, AGIs could learn to pursue goals that are in conflict (i.e. misaligned) with human interests. If trained like today's most capable models, AGIs could learn to act deceptively to receive higher reward, learn misaligned internally-represented goals which generalize beyond their fine-tuning distributions, and pursue those goals using power-seeking strategies. We review emerging evidence for these properties. AGIs with these properties would be difficult to align and may appear aligned even when they are not. Finally, we briefly outline how the deployment of misaligned AGIs might irreversibly undermine human control over the world, and we review research directions aimed at preventing this outcome.

We conduct a systematic audit of three widely used reasoning benchmarks, SocialIQa, FauxPas-EAI, and ToMi, and uncover pervasive flaws in both benchmark items and evaluation methodology. Using five LLMs (GPT-{3, 3.5, 4, o1}, and LLaMA 3.1) as diagnostic tools, we identify structural, semantic, and pragmatic issues in benchmark design (e.g., duplicated items, ambiguous wording, and implausible answers), as well as scoring procedures that prioritize output form over reasoning process. Through systematic human annotation and re-evaluation on cleaned benchmark subsets, we find that model scores often improve not due to due to erratic surface wording variations and not to improved reasoning. Infact, further analyses show that model performance is highly sensitive to minor input variations such as context availability and phrasing, revealing that high scores may reflect alignment with format-specific cues rather than consistent inference based on the input. These findings challenge the validity of current benchmark-based claims about reasoning in LLMs, and highlight the need for evaluation protocols that assess reasoning as a process of drawing inference from available information, rather than as static output selection. We release audited data and evaluation tools to support more interpretable and diagnostic assessments of model reasoning.

Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. Although these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work has observed that the likelihood of preferred responses often decreases during training. The current work sheds light on the causes and implications of this counter-intuitive phenomenon, which we term likelihood displacement. We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning. As a simple example, training a model to prefer No over Never can sharply increase the probability of Yes. Moreover, when aligning the model to refuse unsafe prompts, we show that such displacement can unintentionally lead to unalignment, by shifting probability mass from preferred refusal responses to harmful responses (e.g., reducing the refusal rate of Llama-3-8B-Instruct from 74.4% to 33.4%). We theoretically characterize that likelihood displacement is driven by preferences that induce similar embeddings, as measured by a centered hidden embedding similarity (CHES) score. Empirically, the CHES score enables identifying which training samples contribute most to likelihood displacement in a given dataset. Filtering out these samples effectively mitigated unintentional unalignment in our experiments. More broadly, our results highlight the importance of curating data with sufficiently distinct preferences, for which we believe the CHES score may prove valuable.

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component (η^2 approx 0.52), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

Large language models (LLMs) have witnessed rapid advancements, demonstrating remarkable capabilities. However, a notable vulnerability persists: LLMs often uncritically accept flawed or contradictory premises, leading to inefficient reasoning and unreliable outputs. This emphasizes the significance of possessing the Premise Critique Ability for LLMs, defined as the capacity to proactively identify and articulate errors in input premises. Most existing studies assess LLMs' reasoning ability in ideal settings, largely ignoring their vulnerabilities when faced with flawed premises. Thus, we introduce the Premise Critique Bench (PCBench), designed by incorporating four error types across three difficulty levels, paired with multi-faceted evaluation metrics. We conducted systematic evaluations of 15 representative LLMs. Our findings reveal: (1) Most models rely heavily on explicit prompts to detect errors, with limited autonomous critique; (2) Premise critique ability depends on question difficulty and error type, with direct contradictions being easier to detect than complex or procedural errors; (3) Reasoning ability does not consistently correlate with the premise critique ability; (4) Flawed premises trigger overthinking in reasoning models, markedly lengthening responses due to repeated attempts at resolving conflicts. These insights underscore the urgent need to enhance LLMs' proactive evaluation of input validity, positioning premise critique as a foundational capability for developing reliable, human-centric systems. The code is available at https://github.com/MLGroupJLU/Premise_Critique.

For a parametric model of distributions, the closest distribution in the model to the true distribution located outside the model is considered. Measuring the closeness between two distributions with the Kullback-Leibler (K-L) divergence, the closest distribution is called the "information projection." The estimation risk of the maximum likelihood estimator (MLE) is defined as the expectation of K-L divergence between the information projection and the predictive distribution with plugged-in MLE. Here, the asymptotic expansion of the risk is derived up to n^{-2}-order, and the sufficient condition on the risk for the Bayes error rate between the true distribution and the information projection to be lower than a specified value is investigated. Combining these results, the "p-n criterion" is proposed, which determines whether the MLE is sufficiently close to the information projection for the given model and sample. In particular, the criterion for an exponential family model is relatively simple and can be used for a complex model with no explicit form of normalizing constant. This criterion can constitute a solution to the sample size or model acceptance problem. Use of the p-n criteria is demonstrated for two practical datasets. The relationship between the results and information criteria is also studied.

Recent advancements in Language Models (LMs) have catalyzed the creation of multiple benchmarks, designed to assess these models' general capabilities. A crucial task, however, is assessing the validity of the benchmarks themselves. This is most commonly done via Benchmark Agreement Testing (BAT), where new benchmarks are validated against established ones using some agreement metric (e.g., rank correlation). Despite the crucial role of BAT for benchmark builders and consumers, there are no standardized procedures for such agreement testing. This deficiency can lead to invalid conclusions, fostering mistrust in benchmarks and upending the ability to properly choose the appropriate benchmark to use. By analyzing over 40 prominent benchmarks, we demonstrate how some overlooked methodological choices can significantly influence BAT results, potentially undermining the validity of conclusions. To address these inconsistencies, we propose a set of best practices for BAT and demonstrate how utilizing these methodologies greatly improves BAT robustness and validity. To foster adoption and facilitate future research,, we introduce BenchBench, a python package for BAT, and release the BenchBench-leaderboard, a meta-benchmark designed to evaluate benchmarks using their peers. Our findings underscore the necessity for standardized BAT, ensuring the robustness and validity of benchmark evaluations in the evolving landscape of language model research. BenchBench Package: https://github.com/IBM/BenchBench Leaderboard: https://huggingface.co/spaces/per/BenchBench

The field of machine translation faces an under-recognized problem because of inconsistency in the reporting of scores from its dominant metric. Although people refer to "the" BLEU score, BLEU is in fact a parameterized metric whose values can vary wildly with changes to these parameters. These parameters are often not reported or are hard to find, and consequently, BLEU scores between papers cannot be directly compared. I quantify this variation, finding differences as high as 1.8 between commonly used configurations. The main culprit is different tokenization and normalization schemes applied to the reference. Pointing to the success of the parsing community, I suggest machine translation researchers settle upon the BLEU scheme used by the annual Conference on Machine Translation (WMT), which does not allow for user-supplied reference processing, and provide a new tool, SacreBLEU, to facilitate this.

Evaluating the performance of machine learning models under distribution shift is challenging, especially when we only have unlabeled data from the shifted (target) domain, along with labeled data from the original (source) domain. Recent work suggests that the notion of disagreement, the degree to which two models trained with different randomness differ on the same input, is a key to tackle this problem. Experimentally, disagreement and prediction error have been shown to be strongly connected, which has been used to estimate model performance. Experiments have led to the discovery of the disagreement-on-the-line phenomenon, whereby the classification error under the target domain is often a linear function of the classification error under the source domain; and whenever this property holds, disagreement under the source and target domain follow the same linear relation. In this work, we develop a theoretical foundation for analyzing disagreement in high-dimensional random features regression; and study under what conditions the disagreement-on-the-line phenomenon occurs in our setting. Experiments on CIFAR-10-C, Tiny ImageNet-C, and Camelyon17 are consistent with our theory and support the universality of the theoretical findings.

Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive in a matched-architecture pair of Spanish security language models and propose a ladder of strict, cheap diagnostics. A 661.6M parameter model (approx. 65% code/technical text; no dedicated SFT) and a 1,109M model (web-heavy multi-phase curriculum; 6B-token tool-SFT) share decoder, tokenizer, and special tokens, scoring almost identically on lenient tool-use metrics (B4: 0.660 vs. 0.650). Verbatim-reproduction checks on training examples separate them completely: the 600M emits valid tool calls with generalized arguments on 6/6 examples; the 1B does so on 0/6 across checkpoints. A first-token probe localizes the 1B's failure to a missing prior (prob. 10^{-4}--10^{-5} on <|tool_call|>), which was erased by its web-heavy training phase. A targeted SFT recipe (diverse corpus, 5x higher learning rate, 2,202 steps, ~3.3 GPU-hours) repairs the 1B using three orders of magnitude fewer tokens than the failed phase. On all 269 corpus rows, valid emission rises from 0.100 to 0.959 (600M: 0.926). On 238 unseen prompts, the repaired 1B passes 0.536 vs. the 600M's 0.428 (p = 0.004). Embedding-drift checks show the repair did not move the trigger token's tied embedding (97.7% of the bf16 table remains bit-identical), meaning changes live in the surrounding network. Both models over-trigger, rarely answering negative prompts without a call (0.09 for 600M, 0.17 for repaired 1B). Factorial analyses confirm all repair configurations install the format, though suppression benefits from a diverse corpus remain a hypothesis due to seed sensitivity. This cheap diagnostic ladder costs minutes of CPU time and should gate tool-use claims on small models.

Typed-decision models such as TypeSafe's Jev answer a declared yes/no or multiple-choice question about a state with a probability instead of text, and their evaluations report accuracy and calibration. Neither requires that the probabilities a model gives to logically related questions fit together item by item, which is what a system that acts on those probabilities needs. We test this property, coherence, with a battery of logically linked questions that needs no labels. For 160 items from ChaosNLI and PubMedQA, each with three mutually exclusive labels, we ask whether the label is X, whether it is not X, whether it is one of the other two labels, and which label applies. On 480 negation pairs, Jev's probabilities for "the label is X" and "the label is not X" miss summing to one by 0.064 on average (95% CI 0.055 to 0.072). Qwen3.8-27B, run from its official BF16 weights, misses by 0.293 with first-token probabilities and by 0.122 with verbalized probabilities. The gap to the first-token readout persists on pairs where both systems give similar probabilities, without double-negation labels, and after averaging Jev's repeated calls. Jev is not coherent either: its violations are about five times its repeat noise, and it over-endorses statements about single labels, so that its three single-label probabilities sum to 1.14 on average. The two systems also fail differently. Qwen3.8-27B's first-token readout under-endorses the complement of a label whether or not the question contains "not", rejecting both a statement and its negation in 196 of 480 pairs, and it does not become more coherent where it is more confident, whereas Jev's violations concentrate where its answer is uncertain. Because the checks need no labels, they expose biases that appear only when question forms are compared, and inconsistencies within items.

Pairwise human preference prediction is central to evaluating code-generation systems, where quality often depends on task-specific trade-offs beyond functional correctness. While rubric-based LLM judges improve interpretability by decomposing evaluation into explicit criteria, most existing pipelines remain pointwise: they score each response independently and derive preferences by comparing aggregated scores. We show that this design is poorly matched to pairwise code preference prediction and can underperform a strong monolithic judge. We propose CriterAlign, a criterion-centric framework that adapts rubric-based judging to pairwise preference evaluation through direct criterion-level pairwise judgments, tie-driven criterion refinement, swap-consistency filtering, and final pairwise synthesis. We further introduce Human-Preference-Aligned Guidance (HPAG), synthesized offline from training examples by extracting recurring rationale gaps between human preferences and monolithic judge predictions, and injected into the criterion generator, criterion judge, and final judge. On BigCodeReward, CriterAlign improves a Qwen2.5-VL-32B monolithic judge from 60.4% to 66.3% accuracy, with ablations confirming the contributions of pairwise criterion design and HPAG.