Daily Papers

Deep Neural Networks (DNNs) are being adopted as components in software systems. Creating and specializing DNNs from scratch has grown increasingly difficult as state-of-the-art architectures grow more complex. Following the path of traditional software engineering, machine learning engineers have begun to reuse large-scale pre-trained models (PTMs) and fine-tune these models for downstream tasks. Prior works have studied reuse practices for traditional software packages to guide software engineers towards better package maintenance and dependency management. We lack a similar foundation of knowledge to guide behaviors in pre-trained model ecosystems. In this work, we present the first empirical investigation of PTM reuse. We interviewed 12 practitioners from the most popular PTM ecosystem, Hugging Face, to learn the practices and challenges of PTM reuse. From this data, we model the decision-making process for PTM reuse. Based on the identified practices, we describe useful attributes for model reuse, including provenance, reproducibility, and portability. Three challenges for PTM reuse are missing attributes, discrepancies between claimed and actual performance, and model risks. We substantiate these identified challenges with systematic measurements in the Hugging Face ecosystem. Our work informs future directions on optimizing deep learning ecosystems by automated measuring useful attributes and potential attacks, and envision future research on infrastructure and standardization for model registries.

We study the role of batch size in stochastic conditional gradient methods under a μ-Kurdyka-Łojasiewicz (μ-KL) condition. Focusing on momentum-based stochastic conditional gradient algorithms (e.g., Scion), we derive a new analysis that explicitly captures the interaction between stepsize, batch size, and stochastic noise. Our study reveals a regime-dependent behavior: increasing the batch size initially improves optimization accuracy but, beyond a critical threshold, the benefits saturate and can eventually degrade performance under a fixed token budget. Notably, the theory predicts the magnitude of the optimal stepsize and aligns well with empirical practices observed in large-scale training. Leveraging these insights, we derive principled guidelines for selecting the batch size and stepsize, and propose an adaptive strategy that increases batch size and sequence length during training while preserving convergence guarantees. Experiments on NanoGPT are consistent with the theoretical predictions and illustrate the emergence of the predicted scaling regimes. Overall, our results provide a theoretical framework for understanding batch size scaling in stochastic conditional gradient methods and offer guidance for designing efficient training schedules in large-scale optimization.

Although natural language is the default medium for Large Language Models (LLMs), its limited expressive capacity creates a profound bottleneck for complex problem-solving. While recent advancements in AI have relied heavily on scaling, merely internalizing knowledge does not guarantee its effective application. Defining language representation as the linguistic and symbolic constructs used to map and model the real world, this paper argues that shaping schemas through advanced language representation is the next frontier for expanding LLM intelligence. We posit that an LLM's knowledge activation and organization -- its schema -- depends heavily on the structural and symbolic sophistication of the language used to represent a given task. This paper contributes both a formalization of this claim and the empirical evidence to support it. With a new formalization, we present multiple lines of evidence to support our position: Firstly, we review recent empirical practices and emerging methodologies that demonstrate the substantial performance gains achievable through deliberate language representation design, even without modifying model parameters or scale. Secondly, we conduct controlled experiments showing that LLM performance and its internal feature activations vary under different language representations of the same underlying task. Together, these findings highlight language representation design as a promising direction for future research.

The robot learning community has made great strides in recent years, proposing new architectures and showcasing impressive new capabilities; however, the dominant metric used in the literature, especially for physical experiments, is "success rate", i.e. the percentage of runs that were successful. Furthermore, it is common for papers to report this number with little to no information regarding the number of runs, the initial conditions, and the success criteria, little to no narrative description of the behaviors and failures observed, and little to no statistical analysis of the findings. In this paper we argue that to move the field forward, researchers should provide a nuanced evaluation of their methods, especially when evaluating and comparing learned policies on physical robots. To do so, we propose best practices for future evaluations: explicitly reporting the experimental conditions, evaluating several metrics designed to complement success rate, conducting statistical analysis, and adding a qualitative description of failures modes. We illustrate these through an evaluation on physical robots of several learned policies for manipulation tasks.

Foundation model (FM)-based AI agents are rapidly gaining adoption across diverse domains, but their inherent non-determinism and non-reproducibility pose testing and quality assurance challenges. While recent benchmarks provide task-level evaluations, there is limited understanding of how developers verify the internal correctness of these agents during development. To address this gap, we conduct the first large-scale empirical study of testing practices in the AI agent ecosystem, analyzing 39 open-source agent frameworks and 439 agentic applications. We identify ten distinct testing patterns and find that novel, agent-specific methods like DeepEval are seldom used (around 1%), while traditional patterns like negative and membership testing are widely adapted to manage FM uncertainty. By mapping these patterns to canonical architectural components of agent frameworks and agentic applications, we uncover a fundamental inversion of testing effort: deterministic components like Resource Artifacts (tools) and Coordination Artifacts (workflows) consume over 70% of testing effort, while the FM-based Plan Body receives less than 5%. Crucially, this reveals a critical blind spot, as the Trigger component (prompts) remains neglected, appearing in around 1% of all tests. Our findings offer the first empirical testing baseline in FM-based agent frameworks and agentic applications, revealing a rational but incomplete adaptation to non-determinism. To address it, framework developers should improve support for novel testing methods, application developers must adopt prompt regression testing, and researchers should explore barriers to adoption. Strengthening these practices is vital for building more robust and dependable AI agents.

The success of AI models relies on the availability of large, diverse, and high-quality datasets, which can be challenging to obtain due to data scarcity, privacy concerns, and high costs. Synthetic data has emerged as a promising solution by generating artificial data that mimics real-world patterns. This paper provides an overview of synthetic data research, discussing its applications, challenges, and future directions. We present empirical evidence from prior art to demonstrate its effectiveness and highlight the importance of ensuring its factuality, fidelity, and unbiasedness. We emphasize the need for responsible use of synthetic data to build more powerful, inclusive, and trustworthy language models.

Mechanistic interpretability seeks to understand the internal mechanisms of machine learning models, where localization -- identifying the important model components -- is a key step. Activation patching, also known as causal tracing or interchange intervention, is a standard technique for this task (Vig et al., 2020), but the literature contains many variants with little consensus on the choice of hyperparameters or methodology. In this work, we systematically examine the impact of methodological details in activation patching, including evaluation metrics and corruption methods. In several settings of localization and circuit discovery in language models, we find that varying these hyperparameters could lead to disparate interpretability results. Backed by empirical observations, we give conceptual arguments for why certain metrics or methods may be preferred. Finally, we provide recommendations for the best practices of activation patching going forwards.

In today's society, we are becoming increasingly dependent on software systems. However, we also constantly witness the negative impacts of buggy software. Program synthesis aims to improve software correctness by automatically generating the program given an outline of the expected behavior. For decades, program synthesis has been an active research field, with recent approaches looking to incorporate Large Language Models to help generate code. This paper explores the concept of LLM4TDD, where we guide Large Language Models to generate code iteratively using a test-driven development methodology. We conduct an empirical evaluation using ChatGPT and coding problems from LeetCode to investigate the impact of different test, prompt and problem attributes on the efficacy of LLM4TDD.

Large Language Models (LLMs) are foundational in language technologies, particularly in information retrieval (IR). Previous studies have utilized LLMs for query expansion, achieving notable improvements in IR. In this paper, we thoroughly explore the best practice of leveraging LLMs for query expansion. To this end, we introduce a training-free, straightforward yet effective framework called Multi-Text Generation Integration (MuGI). It leverages LLMs to generate multiple pseudo-references, integrating them with queries to enhance both sparse and dense retrievers. Our empirical findings reveal that: (1) Increasing the number of samples from LLMs benefits IR systems; (2) A balance between the query and pseudo-documents, and an effective integration strategy, is critical for high performance; (3) Contextual information from LLMs is essential, even boost a 23M model to outperform a 7B baseline model; (4) Pseudo relevance feedback can further calibrate queries for improved performance; and (5) Query expansion is widely applicable and versatile, consistently enhancing models ranging from 23M to 7B parameters. Our code and all generated references are made available at https://github.com/lezhang7/Retrieval_MuGI

ChatGPT has significantly impacted software development practices, providing substantial assistance to developers in a variety of tasks, including coding, testing, and debugging. Despite its widespread adoption, the impact of ChatGPT as an assistant in collaborative coding remains largely unexplored. In this paper, we analyze a dataset of 210 and 370 developers shared conversations with ChatGPT in GitHub pull requests (PRs) and issues. We manually examined the content of the conversations and characterized the dynamics of the sharing behavior, i.e., understanding the rationale behind the sharing, identifying the locations where the conversations were shared, and determining the roles of the developers who shared them. Our main observations are: (1) Developers seek ChatGPT assistance across 16 types of software engineering inquiries. In both conversations shared in PRs and issues, the most frequently encountered inquiry categories include code generation, conceptual questions, how-to guides, issue resolution, and code review. (2) Developers frequently engage with ChatGPT via multi-turn conversations where each prompt can fulfill various roles, such as unveiling initial or new tasks, iterative follow-up, and prompt refinement. Multi-turn conversations account for 33.2% of the conversations shared in PRs and 36.9% in issues. (3) In collaborative coding, developers leverage shared conversations with ChatGPT to facilitate their role-specific contributions, whether as authors of PRs or issues, code reviewers, or collaborators on issues. Our work serves as the first step towards understanding the dynamics between developers and ChatGPT in collaborative software development and opens up new directions for future research on the topic.

Large Language Models (LLMs) have shown potential to enhance software development through automated code generation and refactoring, reducing development time and improving code quality. This study empirically evaluates StarCoder2, an LLM optimized for code generation, in refactoring code across 30 open-source Java projects. We compare StarCoder2's performance against human developers, focusing on (1) code quality improvements, (2) types and effectiveness of refactorings, and (3) enhancements through one-shot and chain-of-thought prompting. Our results indicate that StarCoder2 reduces code smells by 20.1% more than developers, excelling in systematic issues like Long Statement and Magic Number, while developers handle complex, context-dependent issues better. One-shot prompting increases the unit test pass rate by 6.15% and improves code smell reduction by 3.52%. Generating five refactorings per input further increases the pass rate by 28.8%, suggesting that combining one-shot prompting with multiple refactorings optimizes performance. These findings provide insights into StarCoder2's potential and best practices for integrating LLMs into software refactoring, supporting more efficient and effective code improvement in real-world applications.

As innovation in deep learning continues, many engineers seek to adopt Pre-Trained Models (PTMs) as components in computer systems. Researchers publish PTMs, which engineers adapt for quality or performance prior to deployment. PTM authors should choose appropriate names for their PTMs, which would facilitate model discovery and reuse. However, prior research has reported that model names are not always well chosen - and are sometimes erroneous. The naming for PTM packages has not been systematically studied. In this paper, we frame and conduct the first empirical investigation of PTM naming practices in the Hugging Face PTM registry. We initiated our study with a survey of 108 Hugging Face users to understand the practices in PTM naming. From our survey analysis, we highlight discrepancies from traditional software package naming, and present findings on naming practices. Our findings indicate there is a great mismatch between engineers' preferences and practical practices of PTM naming. We also present practices on detecting naming anomalies and introduce a novel automated DNN ARchitecture Assessment technique (DARA), capable of detecting PTM naming anomalies. We envision future works on leveraging meta-features of PTMs to improve model reuse and trustworthiness.

Recently, Agentic AI has become an increasingly popular research field. However, we argue that current agent research practices lack standardization and scientific rigor, making it hard to conduct fair comparisons among methods. As a result, it is still unclear how different design choices in agent frameworks affect effectiveness, and measuring their progress remains challenging. In this work, we conduct a systematic empirical study on GAIA benchmark and BrowseComp to examine the impact of popular design choices in key agent components in a fair and rigorous manner. We find that the lack of a standard evaluation protocol makes previous works, even open-sourced ones, non-reproducible, with significant variance between random runs. Therefore, we introduce a more robust evaluation protocol to stabilize comparisons. Our study reveals which components and designs are crucial for effective agents, while others are redundant, despite seeming logical. Based on our findings, we build and open-source OAgents, a new foundation agent framework that achieves state-of-the-art performance among open-source projects. OAgents offers a modular design for various agent components, promoting future research in Agentic AI.

Model-sharing platforms, such as Hugging Face, ModelScope, and OpenCSG, have become central to modern machine learning development, enabling developers to share, load, and fine-tune pre-trained models with minimal effort. However, the flexibility of these ecosystems introduces a critical security concern: the execution of untrusted code during model loading (i.e., via trust_remote_code or trust_repo). In this work, we conduct the first large-scale empirical study of custom model loading practices across five major model-sharing platforms to assess their prevalence, associated risks, and developer perceptions. We first quantify the frequency with which models require custom code to function and identify those that execute arbitrary Python files during loading. We then apply three complementary static analysis tools: Bandit, CodeQL, and Semgrep, to detect security smells and potential vulnerabilities, categorizing our findings by CWE identifiers to provide a standardized risk taxonomy. We also use YARA to identify malicious patterns and payload signatures. In parallel, we systematically analyze the documentation, API design, and safety mechanisms of each platform to understand their mitigation strategies and enforcement levels. Finally, we conduct a qualitative analysis of over 600 developer discussions from GitHub, Hugging Face, and PyTorch Hub forums, as well as Stack Overflow, to capture community concerns and misconceptions regarding security and usability. Our findings reveal widespread reliance on unsafe defaults, uneven security enforcement across platforms, and persistent confusion among developers about the implications of executing remote code. We conclude with actionable recommendations for designing safer model-sharing infrastructures and striking a balance between usability and security in future AI ecosystems.

In rapid release development processes, patches that fix critical issues, or implement high-value features are often promoted directly from the development channel to a stabilization channel, potentially skipping one or more stabilization channels. This practice is called patch uplift. Patch uplift is risky, because patches that are rushed through the stabilization phase can end up introducing regressions in the code. This paper examines patch uplift operations at Mozilla, with the aim to identify the characteristics of uplifted patches that introduce regressions. Through statistical and manual analyses, we quantitatively and qualitatively investigate the reasons behind patch uplift decisions and the characteristics of uplifted patches that introduced regressions. Additionally, we interviewed three Mozilla release managers to understand organizational factors that affect patch uplift decisions and outcomes. Results show that most patches are uplifted because of a wrong functionality or a crash. Uplifted patches that lead to faults tend to have larger patch size, and most of the faults are due to semantic or memory errors in the patches. Also, release managers are more inclined to accept patch uplift requests that concern certain specific components, and-or that are submitted by certain specific developers.

Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.

Background: The C and C++ languages hold significant importance in Software Engineering research because of their widespread use in practice. Numerous studies have utilized Machine Learning (ML) and Deep Learning (DL) techniques to detect software vulnerabilities (SVs) in the source code written in these languages. However, the application of these techniques in function-level SV assessment has been largely unexplored. SV assessment is increasingly crucial as it provides detailed information on the exploitability, impacts, and severity of security defects, thereby aiding in their prioritization and remediation. Aims: We conduct the first empirical study to investigate and compare the performance of ML and DL models, many of which have been used for SV detection, for function-level SV assessment in C/C++. Method: Using 9,993 vulnerable C/C++ functions, we evaluated the performance of six multi-class ML models and five multi-class DL models for the SV assessment at the function level based on the Common Vulnerability Scoring System (CVSS). We further explore multi-task learning, which can leverage common vulnerable code to predict all SV assessment outputs simultaneously in a single model, and compare the effectiveness and efficiency of this model type with those of the original multi-class models. Results: We show that ML has matching or even better performance compared to the multi-class DL models for function-level SV assessment with significantly less training time. Employing multi-task learning allows the DL models to perform significantly better, with an average of 8-22% increase in Matthews Correlation Coefficient (MCC). Conclusions: We distill the practices of using data-driven techniques for function-level SV assessment in C/C++, including the use of multi-task DL to balance efficiency and effectiveness. This can establish a strong foundation for future work in this area.

Developers are sharing pre-trained Machine Learning (ML) models through a variety of model sharing platforms, such as Hugging Face, in an effort to make ML development more collaborative. To share the models, they must first be serialized. While there are many methods of serialization in Python, most of them are unsafe. To tame this insecurity, Hugging Face released safetensors as a way to mitigate the threats posed by unsafe serialization formats. In this context, this paper investigates developer's shifts towards using safetensors on Hugging Face in an effort to understand security practices in the ML development community, as well as how developers react to new methods of serialization. Our results find that more developers are adopting safetensors, and many safetensor adoptions were made by automated conversions of existing models by Hugging Face's conversion tool. We also found, however, that a majority of developers ignore the conversion tool's pull requests, and that while many developers are facing issues with using safetensors, they are eager to learn about and adapt the format.

Software logging is essential for maintaining and debugging complex systems, yet it remains unclear how AI coding agents handle this non-functional requirement. While prior work characterizes human logging practices, the behaviors of AI coding agents and the efficacy of natural language instructions in governing them are unexplored. To address this gap, we conduct an empirical study of 4,550 agentic pull requests across 81 open-source repositories. We compare agent logging patterns against human baselines and analyze the impact of explicit logging instructions. We find that agents change logging less often than humans in 58.4% of repositories, though they exhibit higher log density when they do. Furthermore, explicit logging instructions are rare (4.7%) and ineffective, as agents fail to comply with constructive requests 67% of the time. Finally, we observe that humans perform 72.5% of post-generation log repairs, acting as "silent janitors" who fix logging and observability issues without explicit review feedback. These findings indicate a dual failure in natural language instruction (i.e., scarcity of logging instructions and low agent compliance), suggesting that deterministic guardrails might be necessary to ensure consistent logging practices.

Knowledge Graph-based Retrieval-Augmented Generation (KG-RAG) is an increasingly explored approach for combining the reasoning capabilities of large language models with the structured evidence of knowledge graphs. However, current evaluation practices fall short: existing benchmarks often include questions that can be directly answered using existing triples in KG, making it unclear whether models perform reasoning or simply retrieve answers directly. Moreover, inconsistent evaluation metrics and lenient answer matching criteria further obscure meaningful comparisons. In this work, we introduce a general method for constructing benchmarks, together with an evaluation protocol, to systematically assess KG-RAG methods under knowledge incompleteness. Our empirical results show that current KG-RAG methods have limited reasoning ability under missing knowledge, often rely on internal memorization, and exhibit varying degrees of generalization depending on their design.

Synthetic data generated by large language models has become integral to modern NLP training pipelines, from bootstrapping reasoning capabilities to augmenting instruction-following datasets. While recent work demonstrates successful applications maintaining high external data ratios, systematic understanding of how synthetic data proportion affects model behavior across different scales remains limited. This paper presents a controlled empirical study examining model performance, calibration, and output characteristics when trained on varying synthetic-to-external data ratios. Using the Pythia model suite (410M-12B parameters) across five diverse tasks, we evaluate models after one to three training iterations with synthetic data proportions ranging from 0-50\%. Our key findings include: models maintain stable performance with up to 20\% synthetic data, but degradation accelerates beyond 30\%; larger models (6.9B-12B) show greater robustness to synthetic data than smaller models (410M-1.4B); calibration degradation precedes accuracy loss, providing an early warning signal; and task characteristics matter, with reasoning tasks degrading faster than retrieval tasks under synthetic data training. Importantly, we find that current best practices, such as those employed in STaR and Self-Instruct systems that maintain greater than 80\% external data, operate well within safe regimes identified by our experiments. We provide practical guidance for practitioners on synthetic data budgets based on model scale and task requirements, alongside detailed comparison with concurrent work including Shumailov et al.'s model collapse findings.

Large Language Models (LLMs) have become powerful tools for automated code generation. However, these models often overlook critical security practices, which can result in the generation of insecure code that contains vulnerabilities-weaknesses or flaws in the code that attackers can exploit to compromise a system. However, there has been limited exploration of strategies to guide LLMs in generating secure code and a lack of in-depth analysis of the effectiveness of LLMs in repairing code containing vulnerabilities. In this paper, we present a comprehensive evaluation of state-of-the-art LLMs by examining their inherent tendencies to produce insecure code, their capability to generate secure code when guided by self-generated vulnerability hints, and their effectiveness in repairing vulnerabilities when provided with different levels of feedback. Our study covers both proprietary and open-weight models across various scales and leverages established benchmarks to assess a wide range of vulnerability types. Through quantitative and qualitative analyses, we reveal that although LLMs are prone to generating insecure code, advanced models can benefit from vulnerability hints and fine-grained feedback to avoid or fix vulnerabilities. We also provide actionable suggestions to developers to reduce vulnerabilities when using LLMs for code generation.

Agentic coding tools receive goals written in natural language as input, break them down into specific tasks, and write or execute the actual code with minimal human intervention. Central to this process are agent context files ("READMEs for agents") that provide persistent, project-level instructions. In this paper, we conduct the first large-scale empirical study of 2,303 agent context files from 1,925 repositories to characterize their structure, maintenance, and content. We find that these files are not static documentation but complex, difficult-to-read artifacts that evolve like configuration code, maintained through frequent, small additions. Our content analysis of 16 instruction types shows that developers prioritize functional context, such as build and run commands (62.3%), implementation details (69.9%), and architecture (67.7%). We also identify a significant gap: non-functional requirements like security (14.5%) and performance (14.5%) are rarely specified. These findings indicate that while developers use context files to make agents functional, they provide few guardrails to ensure that agent-written code is secure or performant, highlighting the need for improved tooling and practices.

Performance optimization is a critical yet challenging aspect of software development, often requiring a deep understanding of system behavior, algorithmic tradeoffs, and careful code modifications. Although recent advances in AI coding agents have accelerated code generation and bug fixing, little is known about how these agents perform on real-world performance optimization tasks. We present the first empirical study comparing agent- and human-authored performance optimization commits, analyzing 324 agent-generated and 83 human-authored PRs from the AIDev dataset across adoption, maintainability, optimization patterns, and validation practices. We find that AI-authored performance PRs are less likely to include explicit performance validation than human-authored PRs (45.7\% vs. 63.6\%, p=0.007). In addition, AI-authored PRs largely use the same optimization patterns as humans. We further discuss limitations and opportunities for advancing agentic code optimization.

In this position paper, we observe that empirical evaluation in Generative AI is at a crisis point since traditional ML evaluation and benchmarking strategies are insufficient to meet the needs of evaluating modern GenAI models and systems. There are many reasons for this, including the fact that these models typically have nearly unbounded input and output spaces, typically do not have a well defined ground truth target, and typically exhibit strong feedback loops and prediction dependence based on context of previous model outputs. On top of these critical issues, we argue that the problems of {\em leakage} and {\em contamination} are in fact the most important and difficult issues to address for GenAI evaluations. Interestingly, the field of AI Competitions has developed effective measures and practices to combat leakage for the purpose of counteracting cheating by bad actors within a competition setting. This makes AI Competitions an especially valuable (but underutilized) resource. Now is time for the field to view AI Competitions as the gold standard for empirical rigor in GenAI evaluation, and to harness and harvest their results with according value.

Large Language Model (LLM) libraries have emerged as the foundational infrastructure powering today's AI revolution, serving as the backbone for LLM deployment, inference optimization, fine-tuning, and production serving across diverse applications. Despite their critical role in the LLM ecosystem, these libraries face frequent quality issues and bugs that threaten the reliability of AI systems built upon them. To address this knowledge gap, we present the first comprehensive empirical investigation into bug characteristics and testing practices in modern LLM libraries. We examine 313 bug-fixing commits extracted across two widely-adopted LLM libraries: HuggingFace Transformers and vLLM.Through rigorous manual analysis, we establish comprehensive taxonomies categorizing bug symptoms into 5 types and root causes into 14 distinct categories.Our primary discovery shows that API misuse has emerged as the predominant root cause (32.17%-48.19%), representing a notable transition from algorithm-focused defects in conventional deep learning frameworks toward interface-oriented problems. Additionally, we examine 7,748 test functions to identify 7 distinct test oracle categories employed in current testing approaches, with predefined expected outputs (such as specific tensors and text strings) being the most common strategy. Our assessment of existing testing effectiveness demonstrates that the majority of bugs escape detection due to inadequate test cases (41.73%), lack of test drivers (32.37%), and weak test oracles (25.90%). Drawing from these findings, we offer some recommendations for enhancing LLM library quality assurance.

Large Language Models, particularly decoder-only generative models such as GPT, are increasingly used to automate Software Engineering tasks. These models are primarily guided through natural language prompts, making prompt engineering a critical factor in system performance and behavior. Despite their growing role in SE research, prompt-related decisions are rarely documented in a systematic or transparent manner, hindering reproducibility and comparability across studies. To address this gap, we conducted a two-phase empirical study. First, we analyzed nearly 300 papers published at the top-3 SE conferences since 2022 to assess how prompt design, testing, and optimization are currently reported. Second, we surveyed 105 program committee members from these conferences to capture their expectations for prompt reporting in LLM-driven research. Based on the findings, we derived a structured guideline that distinguishes essential, desirable, and exceptional reporting elements. Our results reveal significant misalignment between current practices and reviewer expectations, particularly regarding version disclosure, prompt justification, and threats to validity. We present our guideline as a step toward improving transparency, reproducibility, and methodological rigor in LLM-based SE research.

Agent skills are the de facto mechanism for extending LLM agents with reusable guidance. A skill can shape the agent's task execution, including planning, tool use, problem-solving, and validation. Prior work reported mixed results of agent skills: some skills improve task success rates, while others have no effect, increase token use and execution time, and even reduce success rates. This paper presents a comprehensive analysis of skill-induced agent failures by attributing task failures and cost regressions to specific loaded skills. We introduce a differential analysis framework that attributes a failure or regression to a skill by comparing a target skill-guided run against a no-skill or semantically matched skill reference run that solves the same task, or solves it more cheaply. We instantiate this framework on SkillsBench and SWE-Skills-Bench, yielding 307 skill-induced failures, including 125 functional failures and 182 efficiency regressions. We also build SkillTriage, a taxonomy-guided attribution tool that normalizes paired cases, extracts differential evidence, and produces triage reports. Our major findings include: (1) Skill induced functional failures are rarely caused by obviously irrelevant skills; instead, seemingly relevant skills often make the agent incorrectly implement or omit task-required implementation elements. (2) Skill-induced efficiency regressions are not explained by prompt length alone. (3) The largest sources within Excessive Procedure are excessive verification and heavy implementation pipelines, contributing 67 and 30 cases, respectively. This shows that skills often turn validation checklists and construction recipes into mandatory work. Based on our findings, we propose research topics and tooling improvements for safer and more cost-aware skill reuse.

AI-driven applications have become woven into students' academic and creative workflows, influencing how they learn, write, and produce ideas. Gaining a nuanced understanding of these usage patterns is essential, yet conventional survey and interview methods remain limited by recall bias, self-presentation effects, and the underreporting of habitual behaviors. While ethnographic methods offer richer contextual insights, they often face challenges of scale and reproducibility. To bridge this gap, we introduce a privacy-conscious approach that repurposes VPN-based network traffic analysis as a scalable ethnographic technique for examining students' real-world engagement with AI tools. By capturing anonymized metadata rather than content, this method enables fine-grained behavioral tracing while safeguarding personal information, thereby complementing self-report data. A three-week field deployment with university students reveals fragmented, short-duration interactions across multiple tools and devices, with intense bursts of activity coinciding with exam periods-patterns mirroring institutional rhythms of academic life. We conclude by discussing methodological, ethical, and empirical implications, positioning network traffic analysis as a promising avenue for large-scale digital ethnography on technology-in-practice.

Autonomous AI agents are transforming software development and redefining how developers collaborate with AI. Prior research shows that the adoption and use of AI-powered tools differ between core and peripheral developers. However, it remains unclear how this dynamic unfolds in the emerging era of autonomous coding agents. In this paper, we present the first empirical study of 9,427 agentic PRs, examining how core and peripheral developers use, review, modify, and verify agent-generated contributions prior to acceptance. Through a mix of qualitative and quantitative analysis, we make four key contributions. First, a subset of peripheral developers use agents more often, delegating tasks evenly across bug fixing, feature addition, documentation, and testing. In contrast, core developers focus more on documentation and testing, yet their agentic PRs are frequently merged into the main/master branch. Second, core developers engage slightly more in review discussions than peripheral developers, and both groups focus on evolvability issues. Third, agentic PRs are less likely to be modified, but when they are, both groups commonly perform refactoring. Finally, peripheral developers are more likely to merge without running CI checks, whereas core developers more consistently require passing verification before acceptance. Our analysis offers a comprehensive view of how developer experience shapes integration offer insights for both peripheral and core developers on how to effectively collaborate with coding agents.

Artificial Intelligence systems, which benefit from the availability of large-scale datasets and increasing computational power, have become effective solutions to various critical tasks, such as natural language understanding, speech recognition, and image processing. The advancement of these AI systems is inseparable from open-source software (OSS). This paper presents an empirical study that investigates the issues in the repositories of open-source AI repositories to assist developers in understanding problems during the process of employing AI systems. We collect 576 repositories from the PapersWithCode platform. Among these repositories, we find 24,953 issues by utilizing GitHub REST APIs. Our empirical study includes three phases. First, we manually analyze these issues to categorize the problems that developers are likely to encounter in open-source AI repositories. Specifically, we provide a taxonomy of 13 categories related to AI systems. The two most common issues are runtime errors (23.18%) and unclear instructions (19.53%). Second, we see that 67.5% of issues are closed. We also find that half of these issues resolve within four days. Moreover, issue management features, e.g., label and assign, are not widely adopted in open-source AI repositories. In particular, only 7.81% and 5.9% of repositories label issues and assign these issues to assignees, respectively. Finally, we empirically show that employing GitHub issue management features and writing issues with detailed descriptions facilitate the resolution of issues. Based on our findings, we make recommendations for developers to help better manage the issues of open-source AI repositories and improve their quality.

Vision language models (VLMs) show promise in medical diagnosis, but their performance across demographic subgroups when using in-context learning (ICL) remains poorly understood. We examine how the demographic composition of demonstration examples affects VLM performance in two medical imaging tasks: skin lesion malignancy prediction and pneumothorax detection from chest radiographs. Our analysis reveals that ICL influences model predictions through multiple mechanisms: (1) ICL allows VLMs to learn subgroup-specific disease base rates from prompts and (2) ICL leads VLMs to make predictions that perform differently across demographic groups, even after controlling for subgroup-specific disease base rates. Our empirical results inform best-practices for prompting current VLMs (specifically examining demographic subgroup performance, and matching base rates of labels to target distribution at a bulk level and within subgroups), while also suggesting next steps for improving our theoretical understanding of these models.

In many binary classification applications such as disease diagnosis and spam detection, practitioners often face great needs to control type I errors (i.e., the conditional probability of misclassifying a class 0 observation as class 1) so that it remains below a desired threshold. To address this need, the Neyman-Pearson (NP) classification paradigm is a natural choice; it minimizes type II error (i.e., the conditional probability of misclassifying a class 1 observation as class 0) while enforcing an upper bound, α, on the type I error. Although the NP paradigm has a century-long history in hypothesis testing, it has not been well recognized and implemented in classification schemes. Common practices that directly limit the empirical type I error to no more than α do not satisfy the type I error control objective because the resulting classifiers are still likely to have type I errors much larger than α. As a result, the NP paradigm has not been properly implemented for many classification scenarios in practice. In this work, we develop the first umbrella algorithm that implements the NP paradigm for all scoring-type classification methods, including popular methods such as logistic regression, support vector machines and random forests. Powered by this umbrella algorithm, we propose a novel graphical tool for NP classification methods: NP receiver operating characteristic (NP-ROC) bands, motivated by the popular receiver operating characteristic (ROC) curves. NP-ROC bands will help choose α in a data adaptive way and compare different NP classifiers. We demonstrate the use and properties of the NP umbrella algorithm and NP-ROC bands, available in the R package nproc, through simulation and real data case studies.

In the context of pretraining of Large Language Models (LLMs), synthetic data has emerged as an alternative for generating high-quality pretraining data at scale. This is particularly beneficial in low-resource language settings where the benefits of recent LLMs have been unevenly distributed across languages. In this work, we present a systematic study on the generation and evaluation of synthetic multilingual pretraining data for Indic languages, where we construct a large-scale synthetic dataset BhashaKritika, comprising 540B tokens using 5 different techniques for 10 languages. We explore the impact of grounding generation in documents, personas, and topics. We analyze how language choice, both in the prompt instructions and document grounding, affects data quality, and we compare translations of English content with native generation in Indic languages. To support scalable and language-sensitive evaluation, we introduce a modular quality evaluation pipeline that integrates script and language detection, metadata consistency checks, n-gram repetition analysis, and perplexity-based filtering using KenLM models. Our framework enables robust quality control across diverse scripts and linguistic contexts. Empirical results through model runs reveal key trade-offs in generation strategies and highlight best practices for constructing effective multilingual corpora.

Large language model (LLM)-driven agents are emerging as a powerful new paradigm for solving complex problems. Despite the empirical success of these practices, a theoretical framework to understand and unify their macroscopic dynamics remains lacking. This Letter proposes a method based on the least action principle to estimate the underlying generative directionality of LLMs embedded within agents. By experimentally measuring the transition probabilities between LLM-generated states, we statistically discover a detailed balance in LLM-generated transitions, indicating that LLM generation may not be achieved by generally learning rule sets and strategies, but rather by implicitly learning a class of underlying potential functions that may transcend different LLM architectures and prompt templates. To our knowledge, this is the first discovery of a macroscopic physical law in LLM generative dynamics that does not depend on specific model details. This work is an attempt to establish a macroscopic dynamics theory of complex AI systems, aiming to elevate the study of AI agents from a collection of engineering practices to a science built on effective measurements that are predictable and quantifiable.

Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.

Evaluating LLMs is challenging, as benchmark scores often fail to capture models' real-world usefulness. Instead, users often rely on ``vibe-testing'': informal experience-based evaluation, such as comparing models on coding tasks related to their own workflow. While prevalent, vibe-testing is often too ad hoc and unstructured to analyze or reproduce at scale. In this work, we study how vibe-testing works in practice and then formalize it to support systematic analysis. We first analyze two empirical resources: (1) a survey of user evaluation practices, and (2) a collection of in-the-wild model comparison reports from blogs and social media. Based on these resources, we formalize vibe-testing as a two-part process: users personalize both what they test and how they judge responses. We then introduce a proof-of-concept evaluation pipeline that follows this formulation by generating personalized prompts and comparing model outputs using user-aware subjective criteria. In experiments on coding benchmarks, we find that combining personalized prompts and user-aware evaluation can change which model is preferred, reflecting the role of vibe-testing in practice. These findings suggest that formalized vibe-testing can serve as a useful approach for bridging benchmark scores and real-world experience.

Scaling up model and data size has been quite successful for the evolution of LLMs. However, the scaling law for the diffusion based text-to-image (T2I) models is not fully explored. It is also unclear how to efficiently scale the model for better performance at reduced cost. The different training settings and expensive training cost make a fair model comparison extremely difficult. In this work, we empirically study the scaling properties of diffusion based T2I models by performing extensive and rigours ablations on scaling both denoising backbones and training set, including training scaled UNet and Transformer variants ranging from 0.4B to 4B parameters on datasets upto 600M images. For model scaling, we find the location and amount of cross attention distinguishes the performance of existing UNet designs. And increasing the transformer blocks is more parameter-efficient for improving text-image alignment than increasing channel numbers. We then identify an efficient UNet variant, which is 45% smaller and 28% faster than SDXL's UNet. On the data scaling side, we show the quality and diversity of the training set matters more than simply dataset size. Increasing caption density and diversity improves text-image alignment performance and the learning efficiency. Finally, we provide scaling functions to predict the text-image alignment performance as functions of the scale of model size, compute and dataset size.

Recent progress in vision and language foundation models has significantly advanced multimodal understanding, reasoning, and generation, inspiring a surge of interest in extending such capabilities to embodied settings through vision-language-action (VLA) models. Yet, most VLA models are still trained with supervised fine-tuning (SFT), which struggles to generalize under distribution shifts due to error accumulation. Reinforcement learning (RL) offers a promising alternative by directly optimizing task performance through interaction, but existing attempts remain fragmented and lack a unified platform for fair and systematic comparison across model architectures and algorithmic designs. To address this gap, we introduce RLinf-VLA, a unified and efficient framework for scalable RL training of VLA models. The system adopts a highly flexible resource allocation design that addresses the challenge of integrating rendering, training, and inference in RL+VLA training. In particular, for GPU-parallelized simulators, RLinf-VLA implements a novel hybrid fine-grained pipeline allocation mode, achieving a 1.61x-1.88x speedup in training. Through a unified interface, RLinf-VLA seamlessly supports diverse VLA architectures (e.g., OpenVLA, OpenVLA-OFT), multiple RL algorithms (e.g., PPO, GRPO), and various simulators (e.g., ManiSkill, LIBERO). In simulation, a unified model achieves 98.11\% across 130 LIBERO tasks and 97.66\% across 25 ManiSkill tasks. Beyond empirical performance, our study distills a set of best practices for applying RL to VLA training and sheds light on emerging patterns in this integration. Furthermore, we present preliminary deployment on a real-world Franka robot, where RL-trained policies exhibit stronger generalization than those trained with SFT. We envision RLinf-VLA as a foundation to accelerate and standardize research on embodied intelligence.

The proliferation of open Pre-trained Language Models (PTLMs) on model registry platforms like Hugging Face (HF) presents both opportunities and challenges for companies building products around them. Similar to traditional software dependencies, PTLMs continue to evolve after a release. However, the current state of release practices of PTLMs on model registry platforms are plagued by a variety of inconsistencies, such as ambiguous naming conventions and inaccessible model training documentation. Given the knowledge gap on current PTLM release practices, our empirical study uses a mixed-methods approach to analyze the releases of 52,227 PTLMs on the most well-known model registry, HF. Our results reveal 148 different naming practices for PTLM releases, with 40.87% of changes to model weight files not represented in the adopted name-based versioning practice or their documentation. In addition, we identified that the 52,227 PTLMs are derived from only 299 different base models (the modified original models used to create 52,227 PTLMs), with Fine-tuning and Quantization being the most prevalent modification methods applied to these base models. Significant gaps in release transparency, in terms of training dataset specifications and model card availability, still exist, highlighting the need for standardized documentation. While we identified a model naming practice explicitly differentiating between major and minor PTLM releases, we did not find any significant difference in the types of changes that went into either type of releases, suggesting that major/minor version numbers for PTLMs often are chosen arbitrarily. Our findings provide valuable insights to improve PTLM release practices, nudging the field towards more formal semantic versioning practices.

Package confusion attacks such as typosquatting threaten software supply chains. Attackers make packages with names that syntactically or semantically resemble legitimate ones, tricking engineers into installing malware. While prior work has developed defenses against package confusions in some software package registries, notably NPM, PyPI, and RubyGems, gaps remain: high false-positive rates; generalization to more software package ecosystems; and insights from real-world deployment. In this work, we introduce ConfuGuard, a solution designed to address the challenges posed by package confusion threats. We begin by presenting the first empirical analysis of benign signals derived from prior package confusion data, uncovering their threat patterns, engineering practices, and measurable attributes. We observed that 13.3% of real package confusion attacks are initially stealthy, so we take that into consideration and refined the definitions. Building on state-of-the-art approaches, we extend support from three to six software package registries, and leverage package metadata to distinguish benign packages. Our approach significantly reduces 64% false-positive (from 77% to 13%), with acceptable additional overhead to filter out benign packages by analyzing the package metadata. ConfuGuard is in production at our industry partner, whose analysts have already confirmed 301 packages detected by ConfuGuard as real attacks. We share lessons learned from production and provide insights to researchers.

Efficiently training a multi-task neural solver for various combinatorial optimization problems (COPs) has been less studied so far. In this paper, we propose a general and efficient training paradigm based on multi-armed bandits to deliver a unified combinarotial multi-task neural solver. To this end, we resort to the theoretical loss decomposition for multiple tasks under an encoder-decoder framework, which enables more efficient training via proper bandit task-sampling algorithms through an intra-task influence matrix. Our method achieves much higher overall performance with either limited training budgets or the same training epochs, compared to standard training schedules, which can be promising for advising efficient training of other multi-task large models. Additionally, the influence matrix can provide empirical evidence of some common practices in the area of learning to optimize, which in turn supports the validity of our approach.

Machine learning models are often personalized with categorical attributes that are protected, sensitive, self-reported, or costly to acquire. In this work, we show models that are personalized with group attributes can reduce performance at a group level. We propose formal conditions to ensure the "fair use" of group attributes in prediction tasks by training one additional model -- i.e., collective preference guarantees to ensure that each group who provides personal data will receive a tailored gain in performance in return. We present sufficient conditions to ensure fair use in empirical risk minimization and characterize failure modes that lead to fair use violations due to standard practices in model development and deployment. We present a comprehensive empirical study of fair use in clinical prediction tasks. Our results demonstrate the prevalence of fair use violations in practice and illustrate simple interventions to mitigate their harm.

As modern software systems continue to grow in complexity, triage has become a fundamental process in system operations and maintenance. Triage aims to efficiently prioritize, assign, and assess issues to ensure the reliability of complex environments. The vast amount of heterogeneous data generated by software systems has made effective triage indispensable for maintaining reliability, facilitating maintainability, and enabling rapid issue response. Motivated by these challenges, researchers have devoted extensive effort to advancing triage automation and have achieved significant progress over the past two decades. This survey provides a comprehensive review of 234 papers from 2004 to the present, offering an in-depth examination of the fundamental concepts, system architecture, and problem statement. By comparing the distinct goals of academic and industrial research and by analyzing empirical studies of industrial practices, we identify the major obstacles that limit the practical deployment of triage systems. To assist practitioners in method selection and performance evaluation, we summarize widely adopted open-source datasets and evaluation metrics, providing a unified perspective on the measurement of triage effectiveness. Finally, we outline potential future directions and emerging opportunities to foster a closer integration between academic innovation and industrial application. All reviewed papers and projects are available at https://github.com/AIOps-Lab-NKU/TriageSurvey.

In order to improve reproducibility, deep reinforcement learning (RL) has been adopting better scientific practices such as standardized evaluation metrics and reporting. However, the process of hyperparameter optimization still varies widely across papers, which makes it challenging to compare RL algorithms fairly. In this paper, we show that hyperparameter choices in RL can significantly affect the agent's final performance and sample efficiency, and that the hyperparameter landscape can strongly depend on the tuning seed which may lead to overfitting. We therefore propose adopting established best practices from AutoML, such as the separation of tuning and testing seeds, as well as principled hyperparameter optimization (HPO) across a broad search space. We support this by comparing multiple state-of-the-art HPO tools on a range of RL algorithms and environments to their hand-tuned counterparts, demonstrating that HPO approaches often have higher performance and lower compute overhead. As a result of our findings, we recommend a set of best practices for the RL community, which should result in stronger empirical results with fewer computational costs, better reproducibility, and thus faster progress. In order to encourage the adoption of these practices, we provide plug-and-play implementations of the tuning algorithms used in this paper at https://github.com/facebookresearch/how-to-autorl.

This paper aims to understand the impacts of various data combinations (e.g., web text, wikipedia, github, books) on the training of large language models using SlimPajama. SlimPajama is a rigorously deduplicated, multi-source dataset, which has been refined and further deduplicated to 627B tokens from the extensive 1.2T tokens RedPajama dataset contributed by Together. We've termed our research as SlimPajama-DC, an empirical analysis designed to uncover fundamental characteristics and best practices associated with employing SlimPajama in the training of large language models. During our research with SlimPajama, two pivotal observations emerged: (1) Global deduplication vs. local deduplication. We analyze and discuss how global (across different sources of datasets) and local (within the single source of dataset) deduplications affect the performance of trained models. (2) Proportions of high-quality/highly-deduplicated multi-source datasets in the combination. To study this, we construct six configurations of SlimPajama dataset and train individual ones using 1.3B Cerebras-GPT model with Alibi and SwiGLU. Our best configuration outperforms the 1.3B model trained on RedPajama using the same number of training tokens by a significant margin. All our 1.3B models are trained on Cerebras 16times CS-2 cluster with a total of 80 PFLOP/s in bf16 mixed precision. We further extend our discoveries (such as increasing data diversity is crucial after global deduplication) on a 7B model with large batch-size training. Our models and the separate SlimPajama-DC datasets are available at: https://huggingface.co/MBZUAI-LLM and https://huggingface.co/datasets/cerebras/SlimPajama-627B.

Natural agents can effectively learn from multiple data sources that differ in size, quality, and types of measurements. We study this heterogeneity in the context of offline reinforcement learning (RL) by introducing a new, practically motivated semi-supervised setting. Here, an agent has access to two sets of trajectories: labelled trajectories containing state, action and reward triplets at every timestep, along with unlabelled trajectories that contain only state and reward information. For this setting, we develop and study a simple meta-algorithmic pipeline that learns an inverse dynamics model on the labelled data to obtain proxy-labels for the unlabelled data, followed by the use of any offline RL algorithm on the true and proxy-labelled trajectories. Empirically, we find this simple pipeline to be highly successful -- on several D4RL benchmarks~fu2020d4rl, certain offline RL algorithms can match the performance of variants trained on a fully labelled dataset even when we label only 10\% of trajectories which are highly suboptimal. To strengthen our understanding, we perform a large-scale controlled empirical study investigating the interplay of data-centric properties of the labelled and unlabelled datasets, with algorithmic design choices (e.g., choice of inverse dynamics, offline RL algorithm) to identify general trends and best practices for training RL agents on semi-supervised offline datasets.

Large language models (LLMs) targeting different deployment scales and sizes are currently produced by training each variant from scratch; this is extremely compute-intensive. In this paper, we investigate if pruning an existing LLM and then re-training it with a fraction (<3%) of the original training data can be a suitable alternative to repeated, full retraining. To this end, we develop a set of practical and effective compression best practices for LLMs that combine depth, width, attention and MLP pruning with knowledge distillation-based retraining; we arrive at these best practices through a detailed empirical exploration of pruning strategies for each axis, methods to combine axes, distillation strategies, and search techniques for arriving at optimal compressed architectures. We use this guide to compress the Nemotron-4 family of LLMs by a factor of 2-4x, and compare their performance to similarly-sized models on a variety of language modeling tasks. Deriving 8B and 4B models from an already pretrained 15B model using our approach requires up to 40x fewer training tokens per model compared to training from scratch; this results in compute cost savings of 1.8x for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. We have open-sourced Minitron model weights on Huggingface, with corresponding supplementary material including example code available on GitHub.

Advances in machine learning are closely tied to the creation of datasets. While data documentation is widely recognized as essential to the reliability, reproducibility, and transparency of ML, we lack a systematic empirical understanding of current dataset documentation practices. To shed light on this question, here we take Hugging Face -- one of the largest platforms for sharing and collaborating on ML models and datasets -- as a prominent case study. By analyzing all 7,433 dataset documentation on Hugging Face, our investigation provides an overview of the Hugging Face dataset ecosystem and insights into dataset documentation practices, yielding 5 main findings: (1) The dataset card completion rate shows marked heterogeneity correlated with dataset popularity. (2) A granular examination of each section within the dataset card reveals that the practitioners seem to prioritize Dataset Description and Dataset Structure sections, while the Considerations for Using the Data section receives the lowest proportion of content. (3) By analyzing the subsections within each section and utilizing topic modeling to identify key topics, we uncover what is discussed in each section, and underscore significant themes encompassing both technical and social impacts, as well as limitations within the Considerations for Using the Data section. (4) Our findings also highlight the need for improved accessibility and reproducibility of datasets in the Usage sections. (5) In addition, our human annotation evaluation emphasizes the pivotal role of comprehensive dataset content in shaping individuals' perceptions of a dataset card's overall quality. Overall, our study offers a unique perspective on analyzing dataset documentation through large-scale data science analysis and underlines the need for more thorough dataset documentation in machine learning research.

Globally, artificial intelligence (AI) implementation is growing, holding the capability to fundamentally alter organisational processes and decision making. Simultaneously, this brings a multitude of emergent risks to organisations, exposing vulnerabilities in their extant risk management frameworks. This necessitates a greater understanding of how organisations can position themselves in response. This issue is particularly pertinent within the financial sector with relatively mature AI applications matched with severe societal repercussions of potential risk events. Despite this, academic risk management literature is trailing behind the speed of AI implementation. Adopting a management perspective, this study aims to contribute to the understanding of AI risk management in organisations through an exploratory empirical investigation into these practices. In-depth insights are gained through interviews with nine practitioners from different organisations within the UK financial sector. Through examining areas of organisational convergence and divergence, the findings of this study unearth levels of risk management framework readiness and prevailing approaches to risk management at both a processual and organisational level. Whilst enhancing the developing literature concerning AI risk management within organisations, the study simultaneously offers a practical contribution, providing key areas of guidance for practitioners in the operational development of AI risk management frameworks.

In the absence of data protection measures, software applications lead to privacy breaches, posing threats to end-users and software organisations. Privacy Enhancing Technologies (PETs) are technical measures that protect personal data, thus minimising such privacy breaches. However, for software applications to deliver data protection using PETs, software developers should actively and correctly incorporate PETs into the software they develop. Therefore, to uncover ways to encourage and support developers to embed PETs into software, this Systematic Literature Review (SLR) analyses 39 empirical studies on developers' privacy practices. It reports the usage of six PETs in software application scenarios. Then, it discusses challenges developers face when integrating PETs into software, ranging from intrinsic challenges, such as the unawareness of PETs, to extrinsic challenges, such as the increased development cost. Next, the SLR presents the existing solutions to address these challenges, along with the limitations of the solutions. Further, it outlines future research avenues to better understand PETs from a developer perspective and minimise the challenges developers face when incorporating PETs into software.

Despite the growing consensus that stakeholders affected by AI systems should participate in their design, enormous variation and implicit disagreements exist among current approaches. For researchers and practitioners who are interested in taking a participatory approach to AI design and development, it remains challenging to assess the extent to which any participatory approach grants substantive agency to stakeholders. This article thus aims to ground what we dub the "participatory turn" in AI design by synthesizing existing theoretical literature on participation and through empirical investigation and critique of its current practices. Specifically, we derive a conceptual framework through synthesis of literature across technology design, political theory, and the social sciences that researchers and practitioners can leverage to evaluate approaches to participation in AI design. Additionally, we articulate empirical findings concerning the current state of participatory practice in AI design based on an analysis of recently published research and semi-structured interviews with 12 AI researchers and practitioners. We use these empirical findings to understand the current state of participatory practice and subsequently provide guidance to better align participatory goals and methods in a way that accounts for practical constraints.

Hyperparameter tuning can dramatically impact training stability and final performance of large-scale models. Recent works on neural network parameterisations, such as μP, have enabled transfer of optimal global hyperparameters across model sizes. These works propose an empirical practice of search for optimal global base hyperparameters at a small model size, and transfer to a large size. We extend these works in two key ways. To handle scaling along most important scaling axes, we propose the Complete^{(d)} Parameterisation that unifies scaling in width and depth -- using an adaptation of CompleteP -- as well as in batch-size and training duration. Secondly, with our parameterisation, we investigate per-module hyperparameter optimisation and transfer. We characterise the empirical challenges of navigating the high-dimensional hyperparameter landscape, and propose practical guidelines for tackling this optimisation problem. We demonstrate that, with the right parameterisation, hyperparameter transfer holds even in the per-module hyperparameter regime. Our study covers an extensive range of optimisation hyperparameters of modern models: learning rates, AdamW parameters, weight decay, initialisation scales, and residual block multipliers. Our experiments demonstrate significant training speed improvements in Large Language Models with the transferred per-module hyperparameters.

DINO and DINOv2 are two model families being widely used to learn representations from unlabeled imagery data at large scales. Their learned representations often enable state-of-the-art performance for downstream tasks, such as image classification and segmentation. However, they employ many empirically motivated design choices and their training pipelines are highly complex and unstable -- many hyperparameters need to be carefully tuned to ensure that the representations do not collapse -- which poses considerable difficulty to improving them or adapting them to new domains. In this work, we posit that we can remove most such-motivated idiosyncrasies in the pre-training pipelines, and only need to add an explicit coding rate term in the loss function to avoid collapse of the representations. As a result, we obtain highly simplified variants of the DINO and DINOv2 which we call SimDINO and SimDINOv2, respectively. Remarkably, these simplified models are more robust to different design choices, such as network architecture and hyperparameters, and they learn even higher-quality representations, measured by performance on downstream tasks, offering a Pareto improvement over the corresponding DINO and DINOv2 models. This work highlights the potential of using simplifying design principles to improve the empirical practice of deep learning.

This paper studies how to keep a vision backbone effective while removing token mixers in its basic building blocks. Token mixers, as self-attention for vision transformers (ViTs), are intended to perform information communication between different spatial tokens but suffer from considerable computational cost and latency. However, directly removing them will lead to an incomplete model structure prior, and thus brings a significant accuracy drop. To this end, we first develop an RepIdentityFormer base on the re-parameterizing idea, to study the token mixer free model architecture. And we then explore the improved learning paradigm to break the limitation of simple token mixer free backbone, and summarize the empirical practice into 5 guidelines. Equipped with the proposed optimization strategy, we are able to build an extremely simple vision backbone with encouraging performance, while enjoying the high efficiency during inference. Extensive experiments and ablative analysis also demonstrate that the inductive bias of network architecture, can be incorporated into simple network structure with appropriate optimization strategy. We hope this work can serve as a starting point for the exploration of optimization-driven efficient network design. Project page: https://techmonsterwang.github.io/RIFormer/.

This work studies the challenge of aligning large language models (LLMs) with offline preference data. We focus on alignment by Reinforcement Learning from Human Feedback (RLHF) in particular. While popular preference optimization methods exhibit good empirical performance in practice, they are not theoretically guaranteed to converge to the optimal policy and can provably fail when the data coverage is sparse by classical offline reinforcement learning (RL) results. On the other hand, a recent line of work has focused on theoretically motivated preference optimization methods with provable guarantees, but these are not computationally efficient for large-scale applications like LLM alignment. To bridge this gap, we propose SPAC, a new offline preference optimization method with self-play, inspired by the on-average pessimism technique from the offline RL literature, to be the first provable and scalable approach to LLM alignment. We both provide theoretical analysis for its convergence under single-policy concentrability for the general function approximation setting and demonstrate its competitive empirical performance for LLM alignment on a 7B Mistral model with Open LLM Leaderboard evaluations.

Classroom dialogue plays a crucial role in fostering student engagement and deeper learning. However, analysing dialogue sequences has traditionally relied on either theoretical frameworks or empirical descriptions of practice, with limited integration between the two. This study addresses this gap by developing a comprehensive rule base of dialogue sequences and an Artificial Intelligence (AI) agent that combines expert-informed rule-based systems with a large language model (LLM). The agent applies expert knowledge while adapting to the complexities of natural language, enabling accurate and flexible categorisation of classroom dialogue sequences. By synthesising findings from over 30 studies, we established a comprehensive framework for dialogue analysis. The agent was validated against human expert coding, achieving high levels of precision and reliability. The results demonstrate that the agent provides theory-grounded and adaptive functions, tremendously enhancing the efficiency and scalability of classroom dialogue analysis, offering significant potential in improving classroom teaching practices and supporting teacher professional development.

The pervasive role played by software in virtually all industries has fostered ever-increasing development of applied research in software engineering. In this chapter, we contribute our experience in using the V-Model as a framework for teaching how to conduct applied research in empirical software engineering. The foundational idea of using the V-Model is presented, and guidance for using it to frame the research is provided. Furthermore, we show how the framework has been instantiated throughout nearly two decades of PhD theses done at the University of Kaiserslautern (RPTU Kaiserslautern) in partnership with Fraunhofer IESE, including the most frequent usage patterns, how the different empirical methods fit into the framework, and the lessons we have learned from this experience.

The field of Deep Learning (DL) has undergone explosive growth during the last decade, with a substantial impact on Natural Language Processing (NLP) as well. Yet, compared to more established disciplines, a lack of common experimental standards remains an open challenge to the field at large. Starting from fundamental scientific principles, we distill ongoing discussions on experimental standards in NLP into a single, widely-applicable methodology. Following these best practices is crucial to strengthen experimental evidence, improve reproducibility and support scientific progress. These standards are further collected in a public repository to help them transparently adapt to future needs.

We consider the classic supervised learning problem, where a continuous non-negative random label Y (i.e. a random duration) is to be predicted based upon observing a random vector X valued in R^d with dgeq 1 by means of a regression rule with minimum least square error. In various applications, ranging from industrial quality control to public health through credit risk analysis for instance, training observations can be right censored, meaning that, rather than on independent copies of (X,Y), statistical learning relies on a collection of ngeq 1 independent realizations of the triplet (X, ; min{Y,; C},; δ), where C is a nonnegative r.v. with unknown distribution, modeling censorship and δ=I{Yleq C} indicates whether the duration is right censored or not. As ignoring censorship in the risk computation may clearly lead to a severe underestimation of the target duration and jeopardize prediction, we propose to consider a plug-in estimate of the true risk based on a Kaplan-Meier estimator of the conditional survival function of the censorship C given X, referred to as Kaplan-Meier risk, in order to perform empirical risk minimization. It is established, under mild conditions, that the learning rate of minimizers of this biased/weighted empirical risk functional is of order O_{P}(log(n)/n) when ignoring model bias issues inherent to plug-in estimation, as can be attained in absence of censorship. Beyond theoretical results, numerical experiments are presented in order to illustrate the relevance of the approach developed.

Healthcare automation is shaped by local procedures and organizational constraints, so agent capabilities rarely transfer unchanged across settings. Agent skills, self-contained directories that package reusable procedures for AI agents, are emerging as a procedural layer for adapting healthcare agents across diverse healthcare settings. We present the first empirical analysis of healthcare agent skills, drawing on 557 healthcare-related skills filtered from 58,159 public skills on ClawHub and annotated along ten dimensions covering function, deployment context, autonomy, and safety. We find that public healthcare skills emphasize patient-facing workflow automation and monitoring rather than the diagnostic and treatment-oriented tasks foregrounded in healthcare-agent research; coverage of the healthcare lifecycle and specialized clinical inputs remains uneven; and general technical risk does not reliably capture clinical risk. These findings position healthcare skills as a procedural layer not yet addressed by current benchmarks and risk frameworks.

Synthetic data generators, when trained using privacy-preserving techniques like differential privacy, promise to produce synthetic data with formal privacy guarantees, facilitating the sharing of sensitive data. However, it is crucial to empirically assess the privacy risks associated with the generated synthetic data before deploying generative technologies. This paper outlines the key concepts and assumptions underlying empirical privacy evaluation in machine learning-based generative and predictive models. Then, this paper explores the practical challenges for privacy evaluations of generative models for use cases with millions of training records, such as data from statistical agencies and healthcare providers. Our findings indicate that methods designed to verify the correct operation of the training algorithm are effective for large datasets, but they often assume an adversary that is unrealistic in many scenarios. Based on the findings, we highlight a crucial trade-off between the computational feasibility of the evaluation and the level of realism of the assumed threat model. Finally, we conclude with ideas and suggestions for future research.

Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-32 LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of 0.280. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain (p < 10^{-8}). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.

Context: The SZZ algorithm is the de facto standard for labeling bug fixing commits and finding inducing changes for defect prediction data. Recent research uncovered potential problems in different parts of the SZZ algorithm. Most defect prediction data sets provide only static code metrics as features, while research indicates that other features are also important. Objective: We provide an empirical analysis of the defect labels created with the SZZ algorithm and the impact of commonly used features on results. Method: We used a combination of manual validation and adopted or improved heuristics for the collection of defect data. We conducted an empirical study on 398 releases of 38 Apache projects. Results: We found that only half of the bug fixing commits determined by SZZ are actually bug fixing. If a six-month time frame is used in combination with SZZ to determine which bugs affect a release, one file is incorrectly labeled as defective for every file that is correctly labeled as defective. In addition, two defective files are missed. We also explored the impact of the relatively small set of features that are available in most defect prediction data sets, as there are multiple publications that indicate that, e.g., churn related features are important for defect prediction. We found that the difference of using more features is not significant. Conclusion: Problems with inaccurate defect labels are a severe threat to the validity of the state of the art of defect prediction. Small feature sets seem to be a less severe threat.

Case-oriented physics education research - which seeks to refine and develop theory by linking that theory to cases - incorporates distinct practices for selecting data for analysis, generalizing results, and making causal claims. Unanswered questions about these practices may constrain researchers more familiar with the recurrence-oriented research paradigm - which seeks to inform instructional predictions by discerning reproducible, representative patterns and relationships - from participating in or critically engaging with case-oriented research. We use results from interviews with physics education researchers, a synthesis of the literature on research methodologies, and published examples of case-oriented and recurrence-oriented research to answer "hard-hitting questions" that researchers may pose. In doing so, we aim to substantiate our position that both case-oriented and recurrence- oriented PER are rigorous but that the rigor is of a different nature in each paradigm.

Background. The development of empirical studies in software engineering mainly relies on the data available on code hosting platforms, being GitHub the most representative. Nevertheless, in the last years, the emergence of Machine Learning (ML) has led to the development of platforms specifically designed for developing ML-based projects, being Hugging Face Hub (HFH) the most popular one. With over 250k repositories, and growing fast, HFH is becoming a promising ecosystem of ML artifacts and therefore a potential source of data for empirical studies. However, so far there have been no studies evaluating the potential of HFH for such studies. Objective. In this proposal for a registered report, we aim at performing an exploratory study of the current state of HFH in order to investigate its suitability to be used as a source platform for empirical studies. Method. We conduct a qualitative and quantitative analysis of HFH for empirical studies. The former will be performed by comparing the features of HFH with those of other code hosting platforms, such as GitHub and GitLab. The latter will be performed by analyzing the data available in HFH.

Modern foundation models rely heavily on using scaling laws to guide crucial training decisions. Researchers often extrapolate the optimal architecture and hyper parameters settings from smaller training runs by describing the relationship between, loss, or task performance, and scale. All components of this process vary, from the specific equation being fit, to the training setup, to the optimization method. Each of these factors may affect the fitted law, and therefore, the conclusions of a given study. We discuss discrepancies in the conclusions that several prior works reach, on questions such as the optimal token to parameter ratio. We augment this discussion with our own analysis of the critical impact that changes in specific details may effect in a scaling study, and the resulting altered conclusions. Additionally, we survey over 50 papers that study scaling trends: while 45 of these papers quantify these trends using a power law, most under-report crucial details needed to reproduce their findings. To mitigate this, we we propose a checklist for authors to consider while contributing to scaling law research.

Authorship ethics is a central topic of discussion in research ethics fora. There are various guidelines for authorship (i.e., naming and order). It is not easy to decide the authorship in the presence of varying authorship guidelines. This paper gives an overview of research on authorship practices and issues. It presents a review of 16 empirical research papers published between 2014 -- 2020. The objective is to learn how various research disciplines handle authorship. What are the authorship practices in various research disciplines, and what are the issues associated with these practices?

Large language models (LLMs) have enabled agent-based systems that aim to automate scientific research workflows. Most existing approaches focus on fully autonomous discovery, where AI systems generate research ideas, conduct analyses, and produce manuscripts with minimal human involvement. However, empirical research in economics and the social sciences poses additional constraints: research questions must be grounded in available datasets, identification strategies require careful design, and human judgment remains essential for evaluating economic significance. We introduce HLER (Human-in-the-Loop Economic Research), a multi-agent architecture that supports empirical research automation while preserving critical human oversight. The system orchestrates specialized agents for data auditing, data profiling, hypothesis generation, econometric analysis, manuscript drafting, and automated review. A key design principle is dataset-aware hypothesis generation, where candidate research questions are constrained by dataset structure, variable availability, and distributional diagnostics, reducing infeasible or hallucinated hypotheses. HLER further implements a two-loop architecture: a question quality loop that screens and selects feasible hypotheses, and a research revision loop where automated review triggers re-analysis and manuscript revision. Human decision gates are embedded at key stages, allowing researchers to guide the automated pipeline. Experiments on three empirical datasets show that dataset-aware hypothesis generation produces feasible research questions in 87% of cases (versus 41% under unconstrained generation), while complete empirical manuscripts can be produced at an average API cost of 0.8-1.5 per run. These results suggest that Human-AI collaborative pipelines may provide a practical path toward scalable empirical research.

Physics education researchers (PER) often analyze student data with single-level regression models (e.g., linear and logistic regression). However, education datasets can have hierarchical structures, such as students nested within courses, that single-level models fail to account for. The improper use of single-level models to analyze hierarchical datasets can lead to biased findings. Hierarchical models (a.k.a., multi-level models) account for this hierarchical nested structure in the data. In this publication, we outline the theoretical differences between how single-level and multi-level models handle hierarchical datasets. We then present analysis of a dataset from 112 introductory physics courses using both multiple linear regression and hierarchical linear modeling to illustrate the potential impact of using an inappropriate analytical method on PER findings and implications. Research can leverage multi-institutional datasets to improve the field's understanding of how to support student success in physics. There is no post hoc fix, however, if researchers use inappropriate single-level models to analyze multi-level datasets. To continue developing reliable and generalizable knowledge, PER should use hierarchical models when analyzing hierarchical datasets. The supplemental materials include a sample dataset, R code to model the building and analysis presented in the paper, and an HTML output from the R code.

This paper introduces LegalRikai: Open Benchmark, a new benchmark comprising four complex tasks that emulate Japanese corporate legal practices. The benchmark was created by legal professionals under the supervision of an attorney. This benchmark has 100 samples that require long-form, structured outputs, and we evaluated them against multiple practical criteria. We conducted both human and automated evaluations using leading LLMs, including GPT-5, Gemini 2.5 Pro, and Claude Opus 4.1. Our human evaluation revealed that abstract instructions prompted unnecessary modifications, highlighting model weaknesses in document-level editing that were missed by conventional short-text tasks. Furthermore, our analysis reveals that automated evaluation aligns well with human judgment on criteria with clear linguistic grounding, and assessing structural consistency remains a challenge. The result demonstrates the utility of automated evaluation as a screening tool when expert availability is limited. We propose a dataset evaluation framework to promote more practice-oriented research in the legal domain.

This position paper provides a critical but constructive discussion of current practices in benchmarking and evaluative practices in the field of formal reasoning and automated theorem proving. We take the position that open code, open data, and benchmarks that are complete and error-free will accelerate progress in this field. We identify practices that create barriers to contributing to this field and suggest ways to remove them. We also discuss some of the practices that might produce misleading evaluative information. We aim to create discussions that bring together people from various groups contributing to automated theorem proving, autoformalization, and informal reasoning.

Software repositories is one of the sources of data in Empirical Software Engineering, primarily in the Mining Software Repositories field, aimed at extracting knowledge from the dynamics and practice of software projects. With the emergence of social coding platforms such as GitHub, researchers have now access to millions of software repositories to use as source data for their studies. With this massive amount of data, sampling techniques are needed to create more manageable datasets. The creation of these datasets is a crucial step, and researchers have to carefully select the repositories to create representative samples according to a set of variables of interest. However, current sampling methods are often based on random selection or rely on variables which may not be related to the research study (e.g., popularity or activity). In this paper, we present a methodology for creating representative samples of software repositories, where such representativeness is properly aligned with both the characteristics of the population of repositories and the requirements of the empirical study. We illustrate our approach with use cases based on Hugging Face repositories.

Making an open-weight language model generate text is not the same as making it finish coding work. This retrospective empirical systems experience report follows my engineering work from July to October 2026 across coding clients, model-specific serving recipes, vLLM, SGLang, llama.cpp, and MLX. Models behind a nominally compatible endpoint still differed in tool-call formatting, turn control, context behavior, cache reuse, and reliability. A served model, coherent text, executed tool calls, and code that passes independent behavioral checks proved to be distinct levels of progress. The evidence includes 4,165 coding-agent gateway turns from my own sessions, 623 debug jobs, and three separate delegation studies with hidden behavioral checks and a fixed Claude Opus 5.5 parent. Saved cold-prefix requests had a median time to first token of about 69.4 s, versus 2.8 s for high-reuse requests (observational). In a 45-trial study, direct coding passed 15/15 and delegation to a quantized Qwen3-Coder 30B-A3B worker passed 12/15 and 14/15. On a second 15-task suite under a fixed artifact contract, the completed replacement cohort had 45/45 for direct coding and 12/45 for a Nemotron Cascade 2 worker (11 worker substitutions). Sixteen sessions stopped by a subscription limit before any model call were replaced; counting them as failures gives 37/45 and 8/45. A Qwen3-Coder worker passed 0/45. In both studies with a direct arm, direct coding had lower API list-price-equivalent frontier cost per verified success. A post-hoc diagnostic found that 19 of 36 artifacts rejected only for renamed parameters passed the hidden checks unchanged; they are not counted as successes. The contribution is an evidence-qualified account of serving, agent protocol, session continuity, artifact validation, and verification in practice. The studies are not pooled and do not rank models or runtimes.

As the parameters of Large Language Models (LLMs) have scaled to hundreds of billions, the demand for efficient training methods -- balancing faster computation and reduced memory usage without sacrificing accuracy -- has become more critical than ever. In recent years, various mixed precision strategies, which involve different precision levels for optimization components, have been proposed to increase training speed with minimal accuracy degradation. However, these strategies often require manual adjustments and lack theoretical justification. In this work, we leverage stochastic rounding (SR) to address numerical errors of training with low-precision representation. We provide theoretical analyses of implicit regularization and convergence under the Adam optimizer when SR is utilized. With the insights from these analyses, we extend previous BF16 + SR strategy to be used in distributed settings, enhancing the stability and performance for large scale training. Empirical results from pre-training models with up to 6.7B parameters, for the first time, demonstrate that our BF16 with SR strategy outperforms (BF16, FP32) mixed precision strategies, achieving better validation perplexity, up to 1.54times higher throughput, and 30% less memory usage.

Machine learning models, such as neural networks, decision trees, random forests, and gradient boosting machines, accept a feature vector, and provide a prediction. These models learn in a supervised fashion where we provide feature vectors mapped to the expected output. It is common practice to engineer new features from the provided feature set. Such engineered features will either augment or replace portions of the existing feature vector. These engineered features are essentially calculated fields based on the values of the other features. Engineering such features is primarily a manual, time-consuming task. Additionally, each type of model will respond differently to different kinds of engineered features. This paper reports empirical research to demonstrate what kinds of engineered features are best suited to various machine learning model types. We provide this recommendation by generating several datasets that we designed to benefit from a particular type of engineered feature. The experiment demonstrates to what degree the machine learning model can synthesize the needed feature on its own. If a model can synthesize a planned feature, it is not necessary to provide that feature. The research demonstrated that the studied models do indeed perform differently with various types of engineered features.

Medical large language models (LLMs) research often makes bold claims, from encoding clinical knowledge to reasoning like a physician. These claims are usually backed by evaluation on competitive benchmarks; a tradition inherited from mainstream machine learning. But how do we separate real progress from a leaderboard flex? Medical LLM benchmarks, much like those in other fields, are arbitrarily constructed using medical licensing exam questions. For these benchmarks to truly measure progress, they must accurately capture the real-world tasks they aim to represent. In this position paper, we argue that medical LLM benchmarks should (and indeed can) be empirically evaluated for their construct validity. In the psychological testing literature, "construct validity" refers to the ability of a test to measure an underlying "construct", that is the actual conceptual target of evaluation. By drawing an analogy between LLM benchmarks and psychological tests, we explain how frameworks from this field can provide empirical foundations for validating benchmarks. To put these ideas into practice, we use real-world clinical data in proof-of-concept experiments to evaluate popular medical LLM benchmarks and report significant gaps in their construct validity. Finally, we outline a vision for a new ecosystem of medical LLM evaluation centered around the creation of valid benchmarks.

A code generation model generates code by taking a prompt from a code comment, existing code, or a combination of both. Although code generation models (e.g. GitHub Copilot) are increasingly being adopted in practice, it is unclear whether they can successfully be used for unit test generation without fine-tuning. We investigated how well three generative models (Codex, GPT-3.5-Turbo, and StarCoder) can generate test cases to fill this gap. We used two benchmarks (HumanEval and Evosuite SF110) to investigate the context generation's effect in the unit test generation process. We evaluated the models based on compilation rates, test correctness, coverage, and test smells. We found that the Codex model achieved above 80% coverage for the HumanEval dataset, but no model had more than 2% coverage for the EvoSuite SF110 benchmark. The generated tests also suffered from test smells, such as Duplicated Asserts and Empty Tests.

In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated carefully. The standard practice is to use this data directly on the model that will be deployed, for example by running GRPO on the deployment student. We argue that this is often an inefficient allocation because it overlooks a reward-density principle: sparse sequence-level reward should train models where exploration is productive, while dense token-level teacher reward should be used where the aim is to compress behavior into a smaller model. In this view, GRPO-style sparse RL and OPD-style dense teacher supervision are not separate recipes; they are different reward-density regimes. The allocation rule is simple: use scarce labeled training data upstream on the strongest model that can turn it into reward-shaped behavior, then transfer that behavior downstream as dense supervision. We evaluate this rule on verifiable math with Qwen3 and Llama models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student, while transfer from the same teacher before RL underperforms. The bridge is important: a forward-KL warmup on teacher rollouts followed by OPD on student rollouts is consistently strongest on MATH before any post-bridge student-side sparse RL, and also gives the best pre-Stage~3 AIME endpoints for the canonical 8B/14B teachers. The bridge also makes later student-side sparse RL effective: GRPO that is weak on a cold student lifts MATH from 75.4% to 78.5% after the bridge and outperforms a matched replay control by 2.8 points. The operational principal is to avoid using scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge.

Forming a reliable judgement of a machine learning (ML) model's appropriateness for an application ecosystem is critical for its responsible use, and requires considering a broad range of factors including harms, benefits, and responsibilities. In practice, however, evaluations of ML models frequently focus on only a narrow range of decontextualized predictive behaviours. We examine the evaluation gaps between the idealized breadth of evaluation concerns and the observed narrow focus of actual evaluations. Through an empirical study of papers from recent high-profile conferences in the Computer Vision and Natural Language Processing communities, we demonstrate a general focus on a handful of evaluation methods. By considering the metrics and test data distributions used in these methods, we draw attention to which properties of models are centered in the field, revealing the properties that are frequently neglected or sidelined during evaluation. By studying these properties, we demonstrate the machine learning discipline's implicit assumption of a range of commitments which have normative impacts; these include commitments to consequentialism, abstractability from context, the quantifiability of impacts, the limited role of model inputs in evaluation, and the equivalence of different failure modes. Shedding light on these assumptions enables us to question their appropriateness for ML system contexts, pointing the way towards more contextualized evaluation methodologies for robustly examining the trustworthiness of ML models

This paper studies the theoretical framework of the alignment process of generative models with Reinforcement Learning from Human Feedback (RLHF). We consider a standard mathematical formulation, the reverse-KL regularized contextual bandit for RLHF. Despite its widespread practical application, a rigorous theoretical analysis of this formulation remains open. We investigate its behavior in three distinct settings -- offline, online, and hybrid -- and propose efficient algorithms with finite-sample theoretical guarantees. Moving towards practical applications, our framework, with a robust approximation of the information-theoretical policy improvement oracle, naturally gives rise to several novel RLHF algorithms. This includes an iterative version of the Direct Preference Optimization (DPO) algorithm for online settings, and a multi-step rejection sampling strategy for offline scenarios. Our empirical evaluations on real-world alignment experiment of large language model demonstrate that these proposed methods significantly surpass existing strong baselines, such as DPO and Rejection Sampling Optimization (RSO), showcasing the connections between solid theoretical foundations and their powerful practical implementations.

Language models often exhibit undesirable behavior, e.g., generating toxic or gender-biased text. In the case of neural language models, an encoding of the undesirable behavior is often present in the model's representations. Thus, one natural (and common) approach to prevent the model from exhibiting undesirable behavior is to steer the model's representations in a manner that reduces the probability of it generating undesirable text. This paper investigates the formal and empirical properties of steering functions, i.e., transformation of the neural language model's representations that alter its behavior. First, we derive two optimal, in the least-squares sense, affine steering functions under different constraints. Our theory provides justification for existing approaches and offers a novel, improved steering approach. Second, we offer a series of experiments that demonstrate the empirical effectiveness of the methods in mitigating bias and reducing toxic generation.

Large language models write correct code for isolated problems but remain far weaker at autonomous machine-learning development, where an agent must revise data pipelines, models, and validation over hours of feedback, and on most competitions still finishes below strong human competitors. Outcome-based benchmarks record this gap but not its cause, because they grade the final submission and discard the development process behind it. We introduce TraceML, which pairs human and agent work on the same competitions under one version-level schema: 4,465 human Kaggle trajectories across 134 competitions, seven of which are also worked by two agent scaffolds, giving 430 paired human and 207 agent trajectories. Every code version carries its score, its timestamp, and labels for the action taken, its intent, the edit size, and the score effect. Read this way, the gap becomes concrete. Experts alternate data work, validation, model changes, and ensembling, and return to approaches they had set aside. Each agent scaffold instead collapses into a narrow loop: Codex spends its steps re-weighting ensembles and tuning submissions, MLEvolve mutates its model in place, and neither pivots at the human rate nor reopens abandoned work. A short planning prompt distilled from human practice moves the behaviors it names toward the human profile and lifts scores, but the effort profile stays agent-shaped: instruction closes only the part of the gap that reduces to instructions. We release the corpus, the schema, the labelers, and the extraction pipeline at https://huggingface.co/datasets/jerryyan/TraceML.

Privacy estimation techniques for differentially private (DP) algorithms are useful for comparing against analytical bounds, or to empirically measure privacy loss in settings where known analytical bounds are not tight. However, existing privacy auditing techniques usually make strong assumptions on the adversary (e.g., knowledge of intermediate model iterates or the training data distribution), are tailored to specific tasks, model architectures, or DP algorithm, and/or require retraining the model many times (typically on the order of thousands). These shortcomings make deploying such techniques at scale difficult in practice, especially in federated settings where model training can take days or weeks. In this work, we present a novel ``one-shot'' approach that can systematically address these challenges, allowing efficient auditing or estimation of the privacy loss of a model during the same, single training run used to fit model parameters, and without requiring any a priori knowledge about the model architecture, task, or DP training algorithm. We show that our method provides provably correct estimates for the privacy loss under the Gaussian mechanism, and we demonstrate its performance on well-established FL benchmark datasets under several adversarial threat models.

Thinking Tokens (TT) have been proposed as an unsupervised method to facilitate reasoning in language models. However, despite their conceptual appeal, our findings show that TTs marginally improves performance and consistently underperforms compared to Chain-of-Thought (CoT) reasoning across multiple benchmarks. We hypothesize that this underperformance stems from the reliance on a single embedding for TTs, which results in inconsistent learning signals and introduces noisy gradients. This paper provides a comprehensive empirical analysis to validate this hypothesis and discusses the implications for future research on unsupervised reasoning in LLMs.

The right batch size is important when training language models at scale: a large batch size is necessary for fast training, but a batch size that is too large will harm token efficiency. To navigate this tradeoff, McCandlish et al. (2018) suggest that a critical batch size (CBS), below which training will not substantially degrade loss, can be estimated based on the gradient noise scale during training. While their method has been adopted in practice, e.g., when training GPT-3, strong assumptions are required to justify gradient noise as a proxy for the CBS, which makes it unclear whether their approach should be trusted in practice, limiting its applicability. In this paper, we introduce a simple, empirical approach to directly measure the CBS and show how the CBS evolves over training. Applying our approach to the OLMo models, we find that CBS is near 0 at initialization, increases rapidly at first, and then plateaus as training progresses. Furthermore, we find that this trend holds across different model sizes (1B and 7B), suggesting CBS from small training runs can inform larger-scale training runs. Our findings about how the CBS changes over training motivate batch size warmup as a natural way to reliably train language models at large batch size: start the batch size small and increase it as the CBS grows. To validate this claim, we use batch size warmup to train OLMo 1B to slightly better loss than the original training run with 43% fewer gradient steps. This shows how our framework can be applied to reliably train language models at larger batch sizes, increasing data parallelism without compromising performance.

Recent advances in Multilingual Large Language Models (MLLMs) have significantly enhanced cross-lingual conversational capabilities, yet modeling culturally nuanced and context-dependent communication remains a critical bottleneck. Specifically, existing state-of-the-art models exhibit a severe pragmatic gap when handling structural variations, regional idioms, and honorific consistencies in low-resource contexts like Bangla. To address this limitation, we introduce a novel, culturally aligned instruction-tuning dataset for BangLa Application and DialoguE generation - BLADE and benchmarking framework comprising 4,196 meticulously curated interaction pairs. We leverage this resource to systematically fine-tune and evaluate leading open-weight architectures, including DeepSeek-8B and LLaMA-3.2-3B, utilizing parameter-efficient fine-tuning via LoRA adapters in a 4-bit NormalFloat (NF4) quantization framework. Our empirical evaluations demonstrate that models fine-tuned on our dataset yield substantial improvements in structural fidelity and honorific alignment, providing a rigorous benchmark for bridging pragmatic disparities in low-resource multilingual text generation. Code and dataset: https://github.com/ashuvo25/Bangla_Application_LLM/tree/main

Code agents spend much of their effort simply locating the right code inside a repository. Two approaches dominate current practice. In Semantic Search, the agent retrieves code blocks from a vector index built from the repository in advance. In Deep Agentic Search (also known as grep-search by subagent), a planning agent delegates the exploration to a separate subagent that works in an isolated context window and returns only a condensed result. The second design, which is considered good context engineering practice, exists to protect the main agent from context pollution (also known as context rot), the loss of accuracy that occurs as unrelated material accumulates in the context window. Recent code agents (such as Claude Code, Codex, Antigravity, etc) have adopted it quickly, but there is little evidence on whether it produces better answers. We compare the two approaches on SWE-QA, a benchmark for repository-level code question answering. Semantic search answered 65.2% of questions correctly against 46.2% for deep agentic search, and it produced each correct answer at less than half the cost. To explain the gap, we then coded every failed run into a taxonomy of failure modes. The taxonomy shows that deep agentic search did not remove failures but introduced a new class of them: the single largest share of its failures, 41.8%, occurred at the hand-off between the planner and its sub-agent, and these were usually silent, ending in a fluent and confident answer that was wrong. Deep agentic search addresses a real problem and is now the preferred design in many code agents. However, our results show that the protection it offers may not be free, and that for read-only questions over a repository that can be indexed, retrieval was the stronger and cheaper option.

Data-driven social science research is inherently slow, relying on iterative cycles of observation, hypothesis generation, and experimental validation. While recent data-driven methods promise to accelerate parts of this process, they largely fail to support end-to-end scientific discovery. To address this gap, we introduce EXPERIGEN, an agentic framework that operationalizes end-to-end discovery through a Bayesian optimization inspired two-phase search, in which a Generator proposes candidate hypotheses and an Experimenter evaluates them empirically. Across multiple domains, EXPERIGEN consistently discovers 2-4x more statistically significant hypotheses that are 7-17 percent more predictive than prior approaches, and naturally extends to complex data regimes including multimodal and relational datasets. Beyond statistical performance, hypotheses must be novel, empirically grounded, and actionable to drive real scientific progress. To evaluate these qualities, we conduct an expert review of machine-generated hypotheses, collecting feedback from senior faculty. Among 25 reviewed hypotheses, 88 percent were rated moderately or strongly novel, 70 percent were deemed impactful and worth pursuing, and most demonstrated rigor comparable to senior graduate-level research. Finally, recognizing that ultimate validation requires real-world evidence, we conduct the first A/B test of LLM-generated hypotheses, observing statistically significant results with p less than 1e-6 and a large effect size of 344 percent.

Chain-of-thought (CoT), tree-of-thought (ToT), and related techniques work surprisingly well in practice for some complex reasoning tasks with Large Language Models (LLMs), but why? This work seeks the underlying reasons by conducting experimental case studies and linking the performance benefits to well-established sample and computational complexity principles in machine learning. We experimented with 6 reasoning tasks, ranging from grade school math, air travel planning, ..., to Blocksworld. The results suggest that (i) both CoT and ToT benefit significantly from task decomposition, which breaks a complex reasoning task into a sequence of steps with low sample complexity and explicitly outlines the reasoning structure, and (ii) for computationally hard reasoning tasks, the more sophisticated tree structure of ToT outperforms the linear structure of CoT. These findings provide useful guidelines for the use of LLM in solving reasoning tasks in practice.

Fairness has been a critical issue that affects the adoption of deep learning models in real practice. To improve model fairness, many existing methods have been proposed and evaluated to be effective in their own contexts. However, there is still no systematic evaluation among them for a comprehensive comparison under the same context, which makes it hard to understand the performance distinction among them, hindering the research progress and practical adoption of them. To fill this gap, this paper endeavours to conduct the first large-scale empirical study to comprehensively compare the performance of existing state-of-the-art fairness improving techniques. Specifically, we target the widely-used application scenario of image classification, and utilized three different datasets and five commonly-used performance metrics to assess in total 13 methods from diverse categories. Our findings reveal substantial variations in the performance of each method across different datasets and sensitive attributes, indicating over-fitting on specific datasets by many existing methods. Furthermore, different fairness evaluation metrics, due to their distinct focuses, yield significantly different assessment results. Overall, we observe that pre-processing methods and in-processing methods outperform post-processing methods, with pre-processing methods exhibiting the best performance. Our empirical study offers comprehensive recommendations for enhancing fairness in deep learning models. We approach the problem from multiple dimensions, aiming to provide a uniform evaluation platform and inspire researchers to explore more effective fairness solutions via a set of implications.

Aiming to train and deploy predictive models, organizations collect large amounts of detailed client data, risking the exposure of private information in the event of a breach. To mitigate this, policymakers increasingly demand compliance with the data minimization (DM) principle, restricting data collection to only that data which is relevant and necessary for the task. Despite regulatory pressure, the problem of deploying machine learning models that obey DM has so far received little attention. In this work, we address this challenge in a comprehensive manner. We propose a novel vertical DM (vDM) workflow based on data generalization, which by design ensures that no full-resolution client data is collected during training and deployment of models, benefiting client privacy by reducing the attack surface in case of a breach. We formalize and study the corresponding problem of finding generalizations that both maximize data utility and minimize empirical privacy risk, which we quantify by introducing a diverse set of policy-aligned adversarial scenarios. Finally, we propose a range of baseline vDM algorithms, as well as Privacy-aware Tree (PAT), an especially effective vDM algorithm that outperforms all baselines across several settings. We plan to release our code as a publicly available library, helping advance the standardization of DM for machine learning. Overall, we believe our work can help lay the foundation for further exploration and adoption of DM principles in real-world applications.

Large Language Models (LLMs) are becoming integral to modern software development workflows, assisting developers with code generation, API explanation, and iterative problem-solving through natural language conversations. Despite widespread adoption, there is limited understanding of how developers interact with LLMs in practice and how these conversational dynamics influence task outcomes, code quality, and software engineering workflows. To address this, we leverage CodeChat, a large dataset comprising 82,845 real-world developer-LLM conversations, containing 368,506 code snippets generated across over 20 programming languages, derived from the WildChat dataset. We find that LLM responses are substantially longer than developer prompts, with a median token-length ratio of 14:1. Multi-turn conversations account for 68% of the dataset and often evolve due to shifting requirements, incomplete prompts, or clarification requests. Topic analysis identifies web design (9.6% of conversations) and neural network training (8.7% of conversations) as the most frequent LLM-assisted tasks. Evaluation across five languages (i.e., Python, JavaScript, C++, Java, and C#) reveals prevalent and language-specific issues in LLM-generated code: generated Python and JavaScript code often include undefined variables (83.4% and 75.3% of code snippets, respectively); Java code lacks required comments (75.9%); C++ code frequently omits headers (41.1%) and C# code shows unresolved namespaces (49.2%). During a conversation, syntax and import errors persist across turns; however, documentation quality in Java improves by up to 14.7%, and import handling in Python improves by 3.7% over 5 turns. Prompts that point out mistakes in code generated in prior turns and explicitly request a fix are most effective for resolving errors.

In the current era of deep learning and especially generative models, there is significant investment in training very large generative models. Thus far, such models have been "black boxes" that are difficult to understand in the sense that they have opaque internal mechanisms, leading to difficulties in interpretability, reliability, and control. Naturally, this lack of understanding has led to both hype and fear. This book is an attempt to "open the black box" and understand the mechanisms of large deep networks, through the perspective of representation learning, which is a major factor - arguably the single most important one - in the empirical power of deep learning models. A brief outline of this book is as follows. Chapter 1 will summarize the threads that underlie the whole text. Chapters 2, 3, 4, 5, and 6 will explain the design principles of modern neural network architectures through optimization and information theory, reducing the process of architecture development (long having been described as a sort of "alchemy") to undergraduate-level linear algebra and calculus exercises once the underlying principles are introduced. Chapters 7 and 8 will discuss applications of these principles to solve problems in more paradigmatic ways, obtaining new methods and models which are efficient, interpretable, and controllable by design, and yet no less - sometimes even more - powerful than the black-box models they resemble. Chapter 9 will discuss potential future directions for deep learning, the role of representation learning, as well as some open problems.

Pre-trained models (PTMs) are machine learning models that have been trained in advance, often on large-scale data, and can be reused for new tasks, thereby reducing the need for costly training from scratch. Their widespread adoption introduces a new class of software dependency, which we term Software Dependencies 2.0, extending beyond conventional libraries to learned behaviors embodied in trained models and their associated artifacts. The integration of PTMs as software dependencies in real projects remains unclear, potentially threatening maintainability and reliability of modern software systems that increasingly rely on them. Objective: In this study, we investigate Software Dependencies 2.0 in open-source software (OSS) projects by examining the reuse of PTMs, with a focus on how developers manage and integrate these models. Specifically, we seek to understand: (1) how OSS projects structure and document their PTM dependencies; (2) what stages and organizational patterns emerge in the reuse pipelines of PTMs within these projects; and (3) the interactions among PTMs and other learned components across pipeline stages. We conduct a mixed-methods analysis of a statistically significant random sample of 401 GitHub repositories from the PeaTMOSS dataset (28,575 repositories reusing PTMs from Hugging Face and PyTorch Hub). We quantitatively examine PTM reuse by identifying patterns and qualitatively investigate how developers integrate and manage these models in practice.

Trusted Execution Environments (TEEs), such as Intel SGX and ARM TrustZone, provide isolated regions of CPU and memory for secure computation and are increasingly used to protect sensitive data and code across diverse application domains. However, little is known about how developers actually use TEEs in practice. This paper presents the first large-scale empirical study of real-world TEE applications. We collected and analyzed 241 open-source projects from GitHub that utilize the two most widely-adopted TEEs, Intel SGX and ARM TrustZone. By combining manual inspection with customized static analysis scripts, we examined their adoption contexts, usage patterns, and development practices across three phases. First, we categorized the projects into 8 application domains and identified trends in TEE adoption over time. We found that the dominant use case is IoT device security (30%), which contrasts sharply with prior academic focus on blockchain and cryptographic systems (7%), while AI model protection (12%) is rapidly emerging as a growing domain. Second, we analyzed how TEEs are integrated into software and observed that 32.4% of the projects reimplement cryptographic functionalities instead of using official SDK APIs, suggesting that current SDKs may have limited usability and portability to meet developers' practical needs. Third, we examined security practices through manual inspection and found that 25.3% (61 of 241) of the projects exhibit insecure coding behaviors when using TEEs, such as hardcoded secrets and missing input validation, which undermine their intended security guarantees. Our findings have important implications for improving the usability of TEE SDKs and supporting developers in trusted software development.

LLMs have become the mainstream approaches to code generation. Existing LLMs mainly employ autoregressive generation, i.e. generating code token-by-token from left to right. However, the underlying autoregressive generation has two limitations in code generation. First, autoregressive LLMs only generate a token at each step, showing low efficiency in practice. Second, programming is a non-sequential process involving back-and-forth editing, while autoregressive LLMs only employ the left-to-right generation order. These two intrinsic limitations hinder the further development of LLMs in code generation. Recently, diffusion LLMs have emerged as a promising alternative. Diffusion LLMs address the above limitations with two advances, including multi-token prediction (i.e. generating multiple tokens at each step) and flexible generation order (i.e. flexibly determining which positions to generate tokens). However, there is no systematic study exploring diffusion LLMs in code generation. To bridge the knowledge gap, we present the first empirical study of diffusion LLMs for code generation. Our study involves 9 representative diffusion LLMs and conduct experiments on 4 widely used benchmarks. Based on the results, we summarize the following findings. (1) Existing diffusion LLMs are competitive with autoregressive LLMs with similar sizes. (2) Diffusion LLMs have a stronger length extrapolation ability than autoregressive LLMs and perform better in long code understanding. (3) We explore factors impacting the effectiveness and efficiency of diffusion LLMs, and provide practical guidance. (4) We discuss several promising further directions to improve diffusion LLMs on code generation. We open-source all source code, data, and results to facilitate the following research. The code is publicly available at https://github.com/zhangyitonggg/dllm4code.

Agentic coding tools, such as OpenAI Codex, Claude Code, and Cursor, are transforming the software engineering landscape. These AI-powered systems function as autonomous teammates capable of planning and executing complex development tasks. Agents have become active participants in refactoring, a cornerstone of sustainable software development aimed at improving internal code quality without altering observable behavior. Despite their increasing adoption, there is a critical lack of empirical understanding regarding how agentic refactoring is utilized in practice, how it compares to human-driven refactoring, and what impact it has on code quality. To address this empirical gap, we present a large-scale study of AI agent-generated refactorings in real-world open-source Java projects, analyzing 15,451 refactoring instances across 12,256 pull requests and 14,988 commits derived from the AIDev dataset. Our empirical analysis shows that refactoring is a common and intentional activity in this development paradigm, with agents explicitly targeting refactoring in 26.1% of commits. Analysis of refactoring types reveals that agentic efforts are dominated by low-level, consistency-oriented edits, such as Change Variable Type (11.8%), Rename Parameter (10.4%), and Rename Variable (8.5%), reflecting a preference for localized improvements over the high-level design changes common in human refactoring. Additionally, the motivations behind agentic refactoring focus overwhelmingly on internal quality concerns, with maintainability (52.5%) and readability (28.1%). Furthermore, quantitative evaluation of code quality metrics shows that agentic refactoring yields small but statistically significant improvements in structural metrics, particularly for medium-level changes, reducing class size and complexity (e.g., Class LOC median Δ = -15.25).

Autonomous coding agents are increasingly deployed as AI teammates in modern software engineering, independently authoring pull requests (PRs) that modify production code at scale. This study aims to systematically characterize how autonomous coding agents contribute to software security in practice, how these security-related contributions are reviewed and accepted, and which observable signals are associated with PR rejection. We conduct a large-scale empirical analysis of agent-authored PRs using the AIDev dataset, comprising of over 33,000 curated PRs from popular GitHub repositories. Security-relevant PRs are identified using a keyword filtering strategy, followed by manual validation, resulting in 1,293 confirmed security-related agentic-PRs. We then analyze prevalence, acceptance outcomes, and review latency across autonomous agents, programming ecosystems, and types of code changes. Moreover, we apply qualitative open coding to identify recurring security-related actions and underlying intents, and examine review metadata to identify early signals associated with PR rejection. Security-related Agentic-PRs constitute a meaningful share of agent activity (approximately 4\%). Rather than focusing solely on narrow vulnerability fixes, agents most frequently perform supportive security hardening activities, including testing, documentation, configuration, and improved error handling. Compared to non-security PRs, security-related Agentic-PRs exhibit lower merge rates and longer review latency, reflecting heightened human scrutiny, with variation across agents and programming ecosystems. PR rejection is more strongly associated with PR complexity and verbosity than with explicit security topics.

We study auto research as a closed empirical loop driven by external measurement. Each submitted trial carries a hypothesis, an executable code edit, an evaluator-owned outcome, and feedback that shapes the next proposal. The output is not a generated paper or a single model checkpoint, but an auditable trajectory of proposals, code diffs, experiments, scores, and failure labels. We instantiate this loop with specialist agents that partition recipe surfaces and share measured lineage across trials. The central empirical finding is that lineage feedback lets agents turn evaluator outcomes, including crashes, budget overruns, size failures, and accuracy-gate misses, into later program-level recipe edits rather than one-shot suggestions. Across 1,197 headline-run trials plus 600 Parameter Golf control trials after one-time setup and launch, humans did not choose proposals, edit recipes, override scores, or repair failed trials during the search. In the three headline runs, the same submitted-trial loop reduces Parameter Golf validation bpb by 0.81%, raises NanoChat-D12 CORE by 38.7%, and reduces CIFAR-10 Airbench96 wallclock by 4.59%, with each task measured by its own external evaluator and legality checks. The trace includes a strict architecture-domain audit of 157 headline-run submissions and program rewrites such as a NanoChat attention-kernel path change. Within this scope the loop autonomously writes code, submits experiments, absorbs feedback, applies and combines known techniques inside each environment, and improves public starting recipes.

Inspired by the principle of deliberate practice in human learning, we propose Deliberate Practice for Synthetic Data Generation (DP), a novel framework that improves sample efficiency through dynamic synthetic data generation. Prior work has shown that scaling synthetic data is inherently challenging, as naively adding new data leads to diminishing returns. To address this, pruning has been identified as a key mechanism for improving scaling, enabling models to focus on the most informative synthetic samples. Rather than generating a large dataset and pruning it afterward, DP efficiently approximates the direct generation of informative samples. We theoretically show how training on challenging, informative examples improves scaling laws and empirically validate that DP achieves better scaling performance with significantly fewer training samples and iterations. On ImageNet-100, DP generates 3.4x fewer samples and requires six times fewer iterations, while on ImageNet-1k, it generates 8x fewer samples with a 30 percent reduction in iterations, all while achieving superior performance compared to prior work.

Large language models (LLMs) are increasingly being integrated into software development processes. The ability to generate code and submit pull requests with minimal human intervention, through the use of autonomous AI agents, is poised to become a standard practice. However, little is known about the practical usefulness of these pull requests and the extent to which their contributions are accepted in real-world projects. In this paper, we empirically study 567 GitHub pull requests (PRs) generated using Claude Code, an agentic coding tool, across 157 diverse open-source projects. Our analysis reveals that developers tend to rely on agents for tasks such as refactoring, documentation, and testing. The results indicate that 83.8% of these agent-assisted PRs are eventually accepted and merged by project maintainers, with 54.9% of the merged PRs are integrated without further modification. The remaining 45.1% require additional changes benefit from human revisions, especially for bug fixes, documentation, and adherence to project-specific standards. These findings suggest that while agent-assisted PRs are largely acceptable, they still benefit from human oversight and refinement.

Benchmarks are necessary for healthcare evaluation, but are not sufficient for predicting deployment performance. Our position is that the evaluation--deployment gap arises not because of poorly designed benchmarks, but from implicit assumptions about how users interact with models that cannot be surfaced from benchmarks alone. To make this precise, we propose a classification of assumptions into two categories: task, which can be tested from conversation data alone, and outcome, which requires outcome data and behavioral studies for testing. Critically, outcome assumptions depend on human behavior, something that even well-designed benchmarks cannot directly observe. To demonstrate the operationality of this framework, we retrospectively analyze a healthcare RCT as a case study and find that the gap naturally separates into task and outcome gaps of roughly equal size. To address this, we make two contributions: first, we propose BenchmarkCards, an artifact that documents assumptions, and second, we propose staged evaluation, a procedure that systematically tests assumptions and evaluates performance.

The release of ChatGPT in 2022 triggered a rapid surge in generative artificial intelligence mobile apps (i.e., Gen-AI apps). Despite widespread adoption, little is known about how end users perceive and evaluate these Gen-AI functionalities in practice. In this work, we conduct a user-centered analysis of 676,066 reviews from 173 Gen-AI apps on the Google Play Store. We introduce a four-phase methodology, SARA (Selection, Acquisition, Refinement, and Analysis), that enables the systematic extraction of user insights using prompt-based LLM techniques. First, we demonstrate the reliability of LLMs in topic extraction, achieving 91% accuracy through five-shot prompting and non-informative review filtering. Then, we apply this method to the informative reviews, identify the top 10 user-discussed topics (e.g., AI Performance, Content Quality, and Content Policy & Censorship) and analyze the key challenges and emerging opportunities. Finally, we examine how these topics evolve over time, offering insight into shifting user expectations and engagement patterns with Gen-AI apps. Based on our findings and observations, we present actionable implications for developers and researchers.

Automated segmentation is a fundamental medical image analysis task, which enjoys significant advances due to the advent of deep learning. While foundation models have been useful in natural language processing and some vision tasks for some time, the foundation model developed with image segmentation in mind - Segment Anything Model (SAM) - has been developed only recently and has shown similar promise. However, there are still no systematic analyses or "best-practice" guidelines for optimal fine-tuning of SAM for medical image segmentation. This work summarizes existing fine-tuning strategies with various backbone architectures, model components, and fine-tuning algorithms across 18 combinations, and evaluates them on 17 datasets covering all common radiology modalities. Our study reveals that (1) fine-tuning SAM leads to slightly better performance than previous segmentation methods, (2) fine-tuning strategies that use parameter-efficient learning in both the encoder and decoder are superior to other strategies, (3) network architecture has a small impact on final performance, (4) further training SAM with self-supervised learning can improve final model performance. We also demonstrate the ineffectiveness of some methods popular in the literature and further expand our experiments into few-shot and prompt-based settings. Lastly, we released our code and MRI-specific fine-tuned weights, which consistently obtained superior performance over the original SAM, at https://github.com/mazurowski-lab/finetune-SAM.

Security patches for open-source software constitute a foundational resource for vulnerability remediation research and practice. However, analyzing and applying multiple patches remains challenging, especially when trying to determine at what point in a patch sequence a vulnerability is fully remediated. This paper presents a systematic analysis of multi-patch vulnerability fixes, focusing on their root causes, characteristics, and methods for verifying remediation status throughout the fixing process. Through a manual examination of 1,646 multi-patch fix records, we develop a taxonomy with three primary categories and six subcategories based on their causes. We then compare the distinctive characteristics of multi-patch fixes with those of single-patch fixes and analyze feature variations across categories. In addition, we assess representative vulnerability detection methods for validating complete remediation during multi-patch fixing. Our findings provide new insights into multi-patch fixes and lay a foundation for future research in this field.

AI coding agents are now submitting pull requests (PRs) to software projects, acting not just as assistants but as autonomous contributors. As these agentic contributions are rapidly increasing across real repositories, little is known about how they behave in practice and why many of them fail to be merged. In this paper, we conduct a large-scale study of 33k agent-authored PRs made by five coding agents across GitHub. (RQ1) We first quantitatively characterize merged and not-merged PRs along four broad dimensions: 1) merge outcomes across task types, 2) code changes, 3) CI build results, and 4) review dynamics. We observe that tasks related to documentation, CI, and build update achieve the highest merge success, whereas performance and bug-fix tasks perform the worst. Not-merged PRs tend to involve larger code changes, touch more files, and often do not pass the project's CI/CD pipeline validation. (RQ2) To further investigate why some agentic PRs are not merged, we qualitatively analyze 600 PRs to derive a hierarchical taxonomy of rejection patterns. This analysis complements the quantitative findings in RQ1 by uncovering rejection reasons not captured by quantitative metrics, including lack of meaningful reviewer engagement, duplicate PRs, unwanted feature implementations, and agent misalignment. Together, our findings highlight key socio-technical and human-AI collaboration factors that are critical to improving the success of future agentic workflows.

We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of practices that utilize knowledge about evaluation tasks at training time. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data prior to evaluation. We then show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models, with broad implications for benchmarking and the study of emergent capabilities.

Many cognitive approaches to well-being, such as recognizing and reframing unhelpful thoughts, have received considerable empirical support over the past decades, yet still lack truly widespread adoption in self-help format. A barrier to that adoption is a lack of adequately specific and diverse dedicated practice material. This work examines whether current language models can be leveraged to both produce a virtually unlimited quantity of practice material illustrating standard unhelpful thought patterns matching specific given contexts, and generate suitable positive reframing proposals. We propose PATTERNREFRAME, a novel dataset of about 10k examples of thoughts containing unhelpful thought patterns conditioned on a given persona, accompanied by about 27k positive reframes. By using this dataset to train and/or evaluate current models, we show that existing models can already be powerful tools to help generate an abundance of tailored practice material and hypotheses, with no or minimal additional model training required.

A common task in legal Information Retrieval (IR) is to find relevant legal sources from case-law collections. While legal practice often requires pinpoint citations (pincites) to specific case paragraphs, most existing public legal IR datasets lack paragraph-level citation annotations. Yet, publicly available datasets with such information contain data leakage in the query text and exclude paragraphs that are neither citing nor cited from the corpora, creating an unrealistic and oversimplified retrieval setting, potentially leading to inflated performance. To address these limitations, we contribute a large-scale legal IR dataset constructed from Court of Justice of the European Union (CJEU) judgments. The dataset contains: (i) masked case/paragraph queries, with removed citation information; (ii) a corpus that includes all paragraphs; and (iii) case- and paragraph-level ground-truth citations, with partial human expert validation. Our dataset supports both the development and rigorous evaluation of legal IR methods, at multiple query-document levels (case-to-case, paragraph-to-case, and paragraph-to-paragraph retrieval). Link to dataset: https://huggingface.co/datasets/theresiavr/legalpincite

As Engineering Education Research (EER) develops as a discipline it is necessary for EER scholars to contribute to the development of learning theory rather than simply being informed by it. It has been suggested that to do this effectively will require partnerships between Engineering scholars and psychologists, education researchers, including other social scientists. The formation of such partnerships is particularly important when considering the introduction of business-related skills into engineering curriculum designed to prepare 21st Century Engineering Students for workplace challenges. In order to encourage scholars beyond Engineering to engage with EER, it is necessary to provide an introduction to the complexities of EER. With this aim in mind, this paper provides an outline review of what is considered rigorous research from an EER perspective as well as highlighting some of the core methodological traditions of EER. The paper aims to facilitate further discussion between EER scholars and researchers from other disciplines, ultimately leading to future collaboration on innovative and rigorous EER.

Assessing journal legitimacy during literature reviews, publication venue selection, and citation verification requires consulting information scattered across multiple incompatible data-sets. This paper introduces Aletheia-Probe, an open-source tool that systematically aggregates curated databases and pattern analysis from multiple authoritative sources to provide transparent, confidence-scored journal assessments. The tool explicitly reports which sources were consulted, what each found, and where evidence conflicts. The tool integrates into research workflows through command-line and programmatic interfaces. It reduces manual assessment overhead while explicitly flagging uncertain cases. We present the tool's architecture, core design principles, and practical integration approach. Comprehensive empirical validation will be presented in forthcoming work.

Finding relevant literature underpins the practice of evidence-based medicine. From 2014 to 2016, TREC conducted a clinical decision support track, wherein participants were tasked with finding articles relevant to clinical questions posed by physicians. In total, 87 teams have participated over the past three years, generating 395 runs. During this period, each team has trialled a variety of methods. While there was significant overlap in the methods employed by different teams, the results were varied. Due to the diversity of the platforms used, the results arising from the different techniques are not directly comparable, reducing the ability to build on previous work. By using a stable platform, we have been able to compare different document and query processing techniques, allowing us to experiment with different search parameters. We have used our system to reproduce leading teams runs, and compare the results obtained. By benchmarking our indexing and search techniques, we can statistically test a variety of hypotheses, paving the way for further research.

Meta-learning researchers face two fundamental issues in their empirical work: prototyping and reproducibility. Researchers are prone to make mistakes when prototyping new algorithms and tasks because modern meta-learning methods rely on unconventional functionalities of machine learning frameworks. In turn, reproducing existing results becomes a tedious endeavour -- a situation exacerbated by the lack of standardized implementations and benchmarks. As a result, researchers spend inordinate amounts of time on implementing software rather than understanding and developing new ideas. This manuscript introduces learn2learn, a library for meta-learning research focused on solving those prototyping and reproducibility issues. learn2learn provides low-level routines common across a wide-range of meta-learning techniques (e.g. meta-descent, meta-reinforcement learning, few-shot learning), and builds standardized interfaces to algorithms and benchmarks on top of them. In releasing learn2learn under a free and open source license, we hope to foster a community around standardized software for meta-learning research.

A research idea may be novel, coherent, and scientifically plausible, yet its proposed method may remain insufficiently specified for faithful implementation. We study the codification readiness of implementation-facing research-method specifications, defined by whether they provide sufficient methodological information for a competent implementer or coding agent to construct the intended method without unsupported assumptions. We construct evidence-grounded specifications and their supported resolutions from papers, codebases, issue threads, and reproduction artifacts. We introduce IdeaAMBIG, a benchmark of 660 evidence-grounded instances: 163 real-world gaps from reproducibility reports and GitHub issues, and 497 controlled synthetic gaps injected into codification-ready references. IdeaAMBIG evaluates three capabilities: codification-readiness assessment, defect localization, and clarification action generation. Defect localization receives only the specification, whereas clarification additionally receives the annotated defect. Across 13 LLMs, the best model achieves 9.6% Macro Defect Recovery Rate on real-world instances but 80.6% Macro Clarification Action Success Rate when given the defect. In an oracle study, supplying the gold resolution raises the downstream codification-ready rate from 14% to 98%. Across all evaluated models, defect localization is the main bottleneck, with stronger clarification given the defect.

We introduce a boosting algorithm to pre-process data for fairness. Starting from an initial fair but inaccurate distribution, our approach shifts towards better data fitting while still ensuring a minimal fairness guarantee. To do so, it learns the sufficient statistics of an exponential family with boosting-compliant convergence. Importantly, we are able to theoretically prove that the learned distribution will have a representation rate and statistical rate data fairness guarantee. Unlike recent optimization based pre-processing methods, our approach can be easily adapted for continuous domain features. Furthermore, when the weak learners are specified to be decision trees, the sufficient statistics of the learned distribution can be examined to provide clues on sources of (un)fairness. Empirical results are present to display the quality of result on real-world data.

Autonomous research systems are increasingly capable of executing long research workflows, yet automation alone does not ensure that the resulting process remains scientifically grounded. We introduce AutoResearch, a two-stage system that connects Idea Generation with Idea Execution to address both how research ideas are formed and how they are reliably established through experimentation. In Idea Generation, AutoResearch continuously integrates emerging research signals with accumulated domain knowledge, identifies transferable mechanistic insights, and uses multi-model generation and cross-review to produce grounded, testable research plans. In Idea Execution, coordinated agents decompose these plans into experiments, iteratively implement and diagnose them, and employ independent evidence-based review before accepting research conclusions. Across representative settings in cross-modal retrieval, systems optimization, and benchmark-driven machine learning, AutoResearch turns generated ideas into measurable progress, detects and corrects unreliable experimental results, and makes evidence-conditioned decisions to continue, revise, or terminate research directions. For example, on RSICD benchmark, an AutoResearch-generated idea improves mean Recall from 32.84 to 34.69, while recording only 5 audit-confirmed issue events compared with 11-27 for other autonomous research systems. These results demonstrate a research process in which meaningful insight is grounded before experimentation and conclusions are grounded before acceptance: Insight In, Hallucination Out.

Researchers spend a great deal of time reading research papers. Keshav (2012) provides a three-pass method to researchers to improve their reading skills. This article extends Keshav's method for reading a research compendium. Research compendia are an increasingly used form of publication, which packages not only the research paper's text and figures, but also all data and software for better reproducibility. We introduce the existing conventions for research compendia and suggest how to utilise their shared properties in a structured reading process. Unlike the original, this article is not build upon a long history but intends to provide guidance at the outset of an emerging practice.

Autonomous research systems increasingly make the scientific workflow executable: agents can propose ideas, run code, inspect results, and draft papers. But executable workflows do not by themselves produce research judgment. We analyze where current systems lose trial experience: weak evidence becomes prose, pilot signals become broad claims, memory remains textual, and recurring process failures do not change later behavior. We introduce Sibyl-AutoResearch, a self-evolving AutoResearch framework built around Scientific Trial-and-Error Harnesses. A harness lets agents run bounded trials, preserve positive and negative outcomes, and route lessons into later planning, validation, claim scope, scheduling, critique, writing, and harness repair. We formalize this through two auditable conversion units: trial-to-behavior conversion, which links trial signals to later research actions, and trial-to-harness-behavior conversion, which links recurring process failures to system updates. We implement the framework in SIBYL, a file-backed autonomous research system that exposes the state, roles, memory, gates, and artifact traces needed to inspect these conversion paths. A retrospective audit identifies eight high-confidence conversion events, with a median latency of one iteration and a maximum latency of three iterations. A recovered-failure registry further shows how five naturally occurring failure classes, including duplicate results, stale numbers, and unsupported statistics, were blocked, downgraded, or routed into later repair. These traces do not establish a comparative performance claim; they show that the proposed conversion units are recoverable from realistic autonomous-research workspaces. The SIBYL framework and system are available at https://github.com/Sibyl-Research-Team/AutoResearch-SibylSystem.

In Brazil, all legal professionals must demonstrate their knowledge of the law and its application by passing the OAB exams, the national bar exams. The OAB exams therefore provide an excellent benchmark for the performance of legal information systems since passing the exam would arguably signal that the system has acquired capacity of legal reasoning comparable to that of a human lawyer. This article describes the construction of a new data set and some preliminary experiments on it, treating the problem of finding the justification for the answers to questions. The results provide a baseline performance measure against which to evaluate future improvements. We discuss the reasons to the poor performance and propose next steps.