AI Code Generation

Reinforcement learning (RL) with verifiable rewards has proven effective at post-training LLMs for coding, yet deploying separate task-specific specialists incurs costs that scale with the number of tasks, motivating a unified multi-task RL (MTRL) approach. However, existing MTRL methods treat all coding tasks uniformly, relying on fixed data curricula under a shared optimization strategy, ultimately limiting the effectiveness of multi-task training. To address these limitations, we propose ASTOR, a multi-tASk code reinforcement learning framework via uTility-driven coORdination. Centered on task utility, a signal capturing each task learning potential and cross-task synergy, ASTOR comprises two coupled modules: 1) Hierarchical Utility-Routed Data Scheduling module hierarchically allocates training budget and prioritizes informative prompts, steering training toward the most valuable data and 2) Adaptive Utility-Calibrated Policy Optimization module dynamically scales per-task KL regularization, matching update constraints to each tasks current training state. Experiments on two widely-used LLMs across four representative coding tasks demonstrate that ASTOR consistently improves a single model across all tasks, outperforming the best task-specific specialist by 9.0%-9.5% and surpassing the strongest MTRL baseline by 7.5%-12.8%.

Large language model (LLM)-based coding assistants have made substantial progress, yet most systems remain reactive, requiring developers to explicitly formulate their needs. Proactive coding assistants aim to infer latent developer intent from integrated development environment (IDE) interactions and repository context, thereby reducing interaction overhead and supporting more seamless assistance. However, research in this direction is limited by the scarcity of large-scale real-world developer behavior data. Existing studies therefore often rely on LLM-simulated IDE traces, whose fidelity to real development behavior remains unclear. In this paper, we investigate this simulation-to-reality gap through a large-scale empirical study. We collect real IDE interaction traces from 1{,}246 experienced industry developers over three consecutive days using a custom Visual Studio Code extension, and construct paired LLM-simulated traces for controlled comparison. Our analysis shows that simulated traces differ substantially from real traces in behavioral diversity, temporal structure, and exploratory patterns. Based on the collected data, we introduce \textbf{ProCodeBench}, a real-world benchmark for proactive intent prediction. Experiments with representative LLMs, retrieval-augmented methods, and agentic baselines show that current approaches remain far from reliable under real IDE traces, suggesting that simulation-based evaluation can overestimate real-world performance. Finally, our training study shows that simulated data cannot replace real data, but can complement it when used before real-world fine-tuning. These findings highlight the importance of real developer behavior data for evaluating and training proactive coding assistants.

Multi-agent LLM systems for code generation face a fundamental routing problem: the optimal orchestration topology depends on the structural complexity of the code under modification, yet existing systems select topologies without consulting the codebase. We present Retrieval-Guided Adaptive Orchestration (RGAO), an architecture that closes this loop by extracting a structural complexity vector from a hierarchical code index before selecting the orchestration topology. RGAO operates within Code-Agent, a multi-agent framework whose sub-agents are governed by formal contracts with six-dimensional budget vectors. Our headline contribution is the composition of two previously separate lines of work -- complexity-conditioned LLM routing and formal resource algebras -- yielding a property neither admits alone: provable budget conservation under retrieval-conditioned dynamic topology selection. Concretely we contribute: (1) a complexity-conditioned topology router that reduces proxy-measured misrouting from 30.1% to 8.2%; (2) a budget algebra with a structural-induction conservation theorem; and (3) a hierarchical code retrieval engine. Empirical evaluation demonstrates sub-millisecond DAG construction and linear tree-index scalability.

Code translation is one of the core capabilities of LLMs. However, evaluating the correctness of translations remains difficult, as commonly used metrics such as BLEU measure only syntactic similarity, disregarding program semantics. We propose a novel evaluation methodology for code translation tasks, emphasizing semantic equivalence over surface-level string similarity. Our approach applies established compiler testing methodology to a new domain, allowing the assessment of an LLM fine-tuned for binary lifting tasks (i.e. decompiling binaries to higher-level representations). We introduce a semantic correctness score, defined as the proportion of translations that produce correct execution outcomes, and demonstrate its application by evaluating LLM-based and heuristic decompilers. Our findings show that LLM-based approaches significantly outperform heuristic ones, while BLEU scores show negligible correlation with semantic correctness (r = -0.127 to 0.354), demonstrating that syntactic metrics fail to predict functional accuracy.

AI-assisted development tools enable rapid prototyping of services but often lack awareness of architectural constraints, infrastructure dependencies, and organizational standards required in production environments. Consequently, generated artifacts may exhibit brittle behavior and limited deployability. We propose a retrieval-augmented scaffolding approach that combines platform-based code generation with agentic clarification loops to expose and resolve architectural constraint ambiguities. By combining template retrieval with structured interaction, the method embeds production-relevant considerations during service scaffolding. Evaluation indicates improved architectural consistency and deployability compared to general-purpose AI code generation workflows, suggesting that constraint-aware retrieval is essential for aligning AI-assisted service development with production software engineering practices.

Large language models (LLMs) show strong potential for neural architecture generation, yet existing approaches produce complete model implementations from scratch -- computationally expensive and yielding verbose code. We propose Delta-Code Generation, where fine-tuned LLMs generate compact unified diffs (deltas) to refine baseline architectures rather than synthesizing entire models. Our pipeline iteratively fine-tunes the LLM via LoRA on curated architectures from the LEMUR dataset, with MinHash-Jaccard novelty filtering for structural diversity. We evaluate three 7B-class LLMs -- DeepSeek-Coder-7B, Qwen2.5-Coder-7B, and Mistral-7B -- across six datasets (CIFAR-10, CIFAR-100, MNIST, SVHN, ImageNette, CelebA) using a 22-cycle protocol (1,100 candidates per LLM). All three substantially surpass the full-generation baseline (50.6% valid rate, 42.3% mean first-epoch accuracy): DeepSeek-Coder reaches 75.3% valid rate and 65.8% mean accuracy; Qwen2.5-Coder 72.1%/64.6%; Mistral 66.6%/66.1%. On CIFAR-10, best first-epoch accuracies reach 85.5% (Mistral), 85.2% (DeepSeek), 80.6% (Qwen) -- well above 63.98% full generation and 71.5% for the concurrent approach of Gu et al. Output lengths are 30-50 lines versus 200+ for full generation (75-85% reduction). A 50-epoch study confirms the 1-epoch proxy preserves rankings (Mistral: Spearman $ρ$ = 0.926). Delta-based generation is a token-efficient, multi-domain, LLM-agnostic alternative to full-model synthesis for LLM-driven NAS.

Large language models (LLMs) frequently generate defective outputs in code generation tasks, ranging from logical bugs to security vulnerabilities. While these generation failures are often treated as model-level limitations, empirical evidence increasingly traces their root causes to imperfections within the training corpora. Yet, the specific mechanisms linking training data quality issues to generated code quality issues remain largely unmapped. This paper presents a systematic literature review of 114 primary studies to investigate how training data quality issues propagate into code generation. We establish a unified taxonomy that categorizes generated code quality issues across nine dimensions and training data quality issues into code and non-code attributes. Based on this taxonomy, we formalize a causal framework detailing 18 typical propagation mapping mechanisms. Furthermore, we synthesize state-of-the-art detection and mitigation techniques across the data, model, and generation lifecycles. The reviewed literature reveals a clear methodological shift: quality assurance is transitioning from reactive, heuristic-based post-generation filtering toward proactive, data-centric governance and closed-loop repair. Finally, we identify open challenges and outline research directions for developing reliable LLMs for code through integrated data curation and continuous evaluation. Our repository is available at https://github.com/SYSUSELab/From-Data-to-Code.

Large language models (LLMs) are widely recognised for their applications in natural language generation and are increasingly used for code generation tasks. However, concerns about bias in their generated outputs remain significant. This paper focuses on GPT-4o and Gemini, mainstream tools for code generation, and proposes a framework for evaluating bias in LLM-generated code, specifically examining the influence of protected attributes, prompts and web-search capability. We use two metrics: the code bias score (CBS) and the attribute change ratio (ACR), to quantify the prevalence of bias and the degree of influence of different attributes, respectively. In addition, we investigate four lightweight mitigation strategies: Few-Shot, Chain-of-Thought, Few-Shot Chain-of-Thought, and Multi-agent, aimed at mitigating bias in generated code. Our findings reveal that bias remains prevalent across different protected attributes and datasets even after applying mitigation strategies, highlighting the need for more effective approaches to reduce bias in AI-driven code generation systems.

The promise of Large Language Models in automated software engineering is often measured by functional correctness, overlooking the critical issue of long term maintainability. This paper presents a systematic audit of technical debt in AI-generated software, revealing that AI does not eliminate flaws but rather introduces a distinct machine signature of defects. Our multi-scale analysis, spanning single-file algorithmic tasks and complex, agent generated systems, identifies a fundamental Reasoning-Complexity Trade-off: as models become more capable, they generate increasingly bloated and coupled code. This architectural decay is so pronounced that we establish a Volume-Quality Inverse Law, where code volume is a near perfect predictor of structural degradation. Crucially, we demonstrate that neither functional correctness nor detailed prompting mitigates this decay. These findings challenge the current paradigm of prompt-driven generation, reframing the central problem of AI-based software engineering from one of code generation to one of architectural complexity management. We conclude that future progress depends on equipping agents with explicit architectural foresight to ensure the software they build is not just functional, but also maintainable.

State-of-the-art Large Language Models (LLMs) excel in code generation at the function level. However, the output quality significantly declines when scaling to repository-level systems. Current workflows relying only on natural language prompts suffer from inherent ambiguity and a lack of verifiability. To address this, we propose structured spec-driven engineering (SSDE), a paradigm that leverages structured artifacts to guide LLM generation. We argue that structured specifications as LLM inputs make high-quality, repository-level code generation a tangible goal, while at the same time offering superior verifiability, leading to significant potential for improvement. We first investigate the feasibility of this vision through a pilot study generating Model-View-Controller (MVC) business logic for three software systems using five LLMs, and then highlight the potential, challenges, and future roadmap for SSDE.

Molecular Vibe Coding, a paradigm where chemists interact with LLMs to generate executable programs for molecular tasks, has emerged as a flexible alternative to chemical agents with predefined tools, enabling chemists to express arbitrarily complex, customized workflows. Unlike general coding tasks, molecular coding imposes a distinctive challenge that LLMs should jointly equip programming, molecular understanding, and domain-specific reasoning capabilities. However, existing benchmarks remain disconnected. General code generation benchmarks such as HumanEval and SWE-bench require no chemistry knowledge, while chemistry-focused benchmarks such as S^2-Bench and ChemCoTBench evaluate knowledge recall or property prediction rather than executable code generation. To bridge this gap, we introduce MolViBench, the first benchmark tailored for Molecular Vibe Coding. MolViBench comprises 358 curated tasks across five cognitive levels, ranging from single-API recall to end-to-end virtual screening pipeline design, spanning 12 real-world drug discovery workflows. To rigorously assess generated code, we also propose a multi-layered evaluation framework that combines type-aware output comparison and AST-based API-semantic fallback analysis, which jointly measures executability and chemical correctness. We systematically evaluate 9 frontier coding LLMs and compare three real-world Molecular Vibe Coding paradigms, providing a practical and fine-grained testbed for diagnosing LLMs' coding capabilities in AI-accelerated molecular discovery.

Geometric Algebra (GA) presents challenges to learners due to its highly abstract mathematical structure and complex operational rules, as translating algebraic manipulations into concrete geometric interpretations is a non-intuitive process when developing related code. Currently, some existing GA software packages rely on manually written scripts for code generation and visualization, but their high learning curve hinders widespread adoption. Meanwhile, methods based on Large Language Models (LLMs) often produce logical errors when generating specific GA scripts, such as GAALOPScript, resulting in generally low accuracy. To address these issues, this study proposes GA-VisAgent -- a multi-agent interactive learning application for GA code generation and visualization -- building upon a Geometric algebra large language model (GAGPT). Integrating task planning mechanisms with ReAct reasoning strategies, GA-VisAgent can decompose complex operations into five standardized subtasks, including core operations like geometric products, rotations, and reflections. It supports natural language and mathematical formulas as input to automatically generate executable code, accompanied by interactive visualizations to aid user comprehension. Experimental results show that GA-VisAgent achieved a 90% code generation success rate across 40 typical Conformal GA tasks, representing a 70% improvement over GPT-4o. This application introduces an extensible new paradigm for teaching GA and developing visualization tools for related mathematical concepts. The online service for this project will be available at http://gagis.cn/gacrac.

Understanding how developers interact with AI coding assistants requires more than chat logs or git histories in isolation; it requires reconstructing the full context: which prompt led to which edit, what the developer tried and discarded, and how their strategy evolved over time. We present RECAP (Replay and Examine Captured AI Programming), an open-source platform that (1) passively records AI chat sessions and fine-grained code edits inside VS Code without disrupting the developer's workflow, (2) merges them into a unified timeline for interactive session replay, and (3) exposes an extensible analysis layer, with example modules for behavioral classification and AI reliance measurement. Deployed in a university software engineering course, RECAP captured 2,034 prompts and 8,239 code edits from 41 students across a multi-week project. We demonstrate how the platform's linked data and replay capabilities enable analyses of developer-AI interaction patterns that no single data source could support.

Hallucinations, outputs that sound plausible but are factually incorrect, remain an open challenge for deployed LLMs. In code generation, models frequently hallucinate non-existent software packages, recommending imports and installation commands for fictional libraries. This creates a critical supply-chain vulnerability: an attacker can proactively register such packages on public registries with malicious payloads that are subsequently installed and executed by developers or autonomous agents, a class of package confusion attack known as slopsquatting. Once a model is deployed, mitigating this failure mode is difficult: full retraining is costly, and existing approaches either cause severe degradation of model utility or rely on a pre-specified forget-set, an assumption that does not apply to the unbounded space of hallucinations. To address this problem, we present Adaptive Unlearning (AU), a post-deployment framework that surgically suppresses hallucinations while preserving general model utility. AU introduces a hybrid token-level objective that simultaneously reinforces valid outputs and suppresses hallucinated ones. Combined with an adaptive discovery loop that continuously surfaces new hallucination-inducing contexts without human supervision, AU enables generalization to unseen prompts and hallucinations. We demonstrate that AU reduces package hallucination rates by 81%, corresponding to a substantial reduction in slopsquatting attack surface, while maintaining performance on standard coding benchmarks. Our analysis shows that distributional changes are concentrated on package-related generations, leaving general coding behavior largely unaffected and confirming that AU's effect is isolated to the targeted distribution. AU operates entirely on model-generated data, requires no human annotation, and generalizes across domains.

Generating diverse, readable statistical charts from tabular data remains challenging for LLMs, as many failures become apparent after rendering and are not detectable from data or code alone. Existing chart datasets also rarely provide fully aligned artifacts, such as executable code, dataset context, and question-answer pairs. We present a structured LLM-based workflow that decomposes chart generation into dataset screening, plot proposal, code synthesis, rendering, validation-driven refinement, description generation, and question-answer generation. By incorporating rendered-output validation, the workflow addresses visualization-specific failure modes such as readability and semantic mismatch. It treats chart generation as an inspectable process rather than a one-shot prompt-to-code task, retaining each chart with its code, dataset context, description, and question-answer pairs. Applied to UCI datasets, the workflow produces 1,500 charts from 74 datasets, spanning 24 chart families and paired with 30,003 question-answer pairs. We evaluate 16 multimodal LLMs (MLLMs) on these chart-question pairs. The results show that chart-syntax questions are nearly saturated, while value extraction, comparison, and reasoning remain more challenging, illustrating the workflow's utility for diagnostic studies of chart-grounded multimodal reasoning.

Reinforcement learning (RL) from unit-test feedback has become a standard post-training recipe for improving large language models (LLMs) on code generation. However, the pass-all-tests binary reward can be sparse, yielding no learning signal on challenging problems where none of the sampled solutions passes all tests. A common remedy is to use the test-case pass rate as a surrogate reward. In this work, we study pass-rate rewards in critic-free RL for code generation (e.g., GRPO and RLOO) and report a consistent pattern across base models and algorithms: despite alleviating reward sparsity, pass-rate rewards do not reliably improve final performance over binary rewards in rigorous controlled experiments. To understand this discrepancy, we analyze reward density and the resulting gradient directions. We find that pass-rate rewards are denser, but the induced gradient updates do not consistently move probability mass toward full-pass solutions. This arises because test-case pass rate is a miscalibrated surrogate for progress toward full correctness, and partial-pass solutions within the same group can induce conflicting gradient directions that cancel out. Overall, our results suggest that, in critic-free RL, pass-rate rewards are insufficient to improve code generation and motivate reward designs that better align optimization with the goal of full correctness.

Code generation, which aims to automatically generate source code from given programming requirements, has the potential to substantially improve software development efficiency. With the rapid advancement of large language models (LLMs), LLM-based code generation has attracted widespread attention from both academia and industry. However, as programming requirements become increasingly complex, existing LLMs still exhibit notable performance limitations. To address this challenge, recent studies have proposed training-based curriculum reinforcement learning (CRL) strategies to improve LLM code generation performance. Despite their effectiveness, existing CRL approaches suffer from several limitations, including misaligned requirement difficulty perception, the absence of requirement difficulty optimization, and suboptimal curriculum sampling strategies. In CRL-based code generation, programming requirements serve as the sole input to the model, making their quality and difficulty critical to training effectiveness. Motivated by insights from software requirements engineering, we propose RECRL, a novel requirement-aware curriculum reinforcement learning framework for enhancing LLM-based code generation. RECRL automatically perceives model-specific requirement difficulty, optimizes challenging requirements to improve training data utilization, and employs an adaptive curriculum sampling strategy to construct training batches with smoothly varying difficulty. Extensive experiments on five state-of-the-art LLMs across five widely-used code generation benchmarks by comparing with five state-of-the-art baselines, demonstrate the significant effectiveness of RECRL. For example, RECRL achieves an average Pass@1 improvement of 1.23%-5.62% over all state-of-the-art baselines.

Large Language Models (LLMs) are increasingly deployed to generate code for human-centered applications where demographic fairness is critical. However, existing evaluations focus almost exclusively on functional correctness, leaving social bias in LLM-generated code largely unexamined. Extending our prior work on Solar, we conduct a comprehensive empirical study using SocialBias-Bench, a benchmark of 343 real-world coding tasks spanning seven demographic dimensions. We evaluate four prominent LLMs and find severe bias across all models, with Code Bias Scores reaching up to 60.58%. We further show that standard prompt-level interventions, such as Chain-of-Thought reasoning and fairness persona assignment, inadvertently amplify bias rather than reduce it. We then investigate whether structured multi-agent software process frameworks can improve fairness, finding that structured pipelines reduce bias when early roles correctly scope what the code should and should not consider. However, adding explicit fairness instructions to all agent roles produces worse outcomes than providing none, suggesting that diffused responsibility goes unaddressed. To address these limitations, we propose the Fairness Monitor Agent (FMA), a modular component that plugs into any existing code generation pipeline without modifying it. FMA analyzes the task description to determine which attributes should be considered or restricted, then detects and corrects violations through an iterative review process, without requiring an executable test suite. Evaluated on all 343 tasks, FMA reduces bias by 65.1% compared to a developer agent alone and improves functional correctness from 75.80% to 83.97%, outperforming all other studied approaches.

This report documents the preparedness assessment of Code World Model (CWM), a model for code generation and reasoning about code from Meta. We conducted pre-release testing across domains identified in our Frontier AI Framework as potentially presenting catastrophic risks, and also evaluated the model's misaligned propensities. Our assessment found that CWM does not pose additional frontier risks beyond those present in the current AI ecosystem. We therefore release it as an open-weight model.

Multimodal large language models (MLLMs) are increasingly used to translate visual artifacts into code, from UI mockups into HTML to scientific plots into Python scripts. A circuit diagram can be viewed as a visual domain-specific language for hardware: it encodes timing, topology, and bit level semantics that are invisible to casual inspection yet safety critical once fabricated in silicon. Translating such diagrams into register-transfer-level(RTL) code therefore represents an extreme reliability test for vision-to-code generation. We reveal a phenomenon we call Mirage: replacing a circuit diagram with a blank image leaves Pass@k unchanged or even higher, because models bypass the visual input and instead exploit identifier semantics in the module header to retrieve canonical RTL templates. This constitutes a new, highly covert class of defect in AI-assisted code generation that directly undermines MLLMs' trustworthiness. To quantify the effect, we construct C2VEVAL and evaluate eight MLLMs under a paired Normal/Anony protocol in which Anony mode anonymizes all identifiers in both the diagram and the module header; Anony-mode scores drop sharply across all models, confirming that high Normal-mode accuracy is largely a Mirage. We then propose VeriGround (4B), trained with identifier anonymization, refusal augmentation, and D-ORPO (Decision-Focused ORPO) preference alignment that up-weights pivotal generate-or-refuse tokens. VeriGround achieves Functional Pass@1 of 46.11%/42.51%(Normal/Anony) with a False Refusal Rate of only 1.20%/0.00%, while maintaining >92% Refusal Rate on blank images. With only 4B parameters, VeriGround performs on par with GPT-5.4 under Normal and significantly outperforms all baselines under Anony, confirming genuine visual grounding.

Understanding how Large Language Models (LLMs) make token-level decisions during code generation remains a major challenge for both researchers and practitioners. While recent tools provide insights into model internals or generation outcomes, they often lack decoding-time signals, fine-grained uncertainty measures, and interactive mechanisms for exploring alternative generation paths. We present TokenScope, an interactive interpretability and analysis tool for decoder-based LLMs that exposes token-level metrics, attention patterns, and structural information during generation. TokenScope supports interactive token replacement, counterfactual branching, and code-aware aggregation via abstract syntax trees. By unifying decoding-time signals with structural program analysis, TokenScope enables systematic investigation of LLM behaviour during code generation.

Large Language Models (LLMs) are increasingly used for code editing, yet the prevalent full-code generation paradigm suffers from severe efficiency bottlenecks, posing challenges for interactive coding assistants that demand low latency and cost. Despite the predominant focus on scaling model capabilities, the edit format itself has been largely overlooked in model training. In this paper, we begin with a systematic study of conventional diff formats and reveal that fragile offsets and fragmented hunks make generation highly unnatural for LLMs. To address it, we introduce BlockDiff and FuncDiff, two structure-aware diff formats that represent changes as block-level rewrites of syntactically coherent units such as control structures and functions. Furthermore, we propose AdaEdit, a general adaptive edit strategy that trains LLMs to dynamically choose the most token-efficient format between a given diff format and full code. Extensive experiments demonstrate that AdaEdit paired with structure-aware diff formats consistently matches the accuracy of full-code generation, while reducing both latency and cost by over 30% on long-code editing tasks.

LLMs have achieved strong results on both function-level code synthesis and repository-level code modification, yet a capability that falls between these two extremes -- compositional code creation, i.e., building a complete, internally structured class from a specification -- remains underserved. Current evaluations are either confined to isolated functions or rely on manually curated class-level tasks that are expensive to scale and increasingly susceptible to data contamination. We introduce ClassEval-Pro, a benchmark of 300 class-level tasks spanning 11 domains, constructed through an automated three-stage pipeline that combines complexity enhancement, cross-domain class composition, and integration of real-world GitHub code contributed after January 2025. Every task is validated by an LLM Judge Ensemble and must pass test suites with over 90% line coverage. We evaluate five frontier LLMs under five generation strategies. The best model achieves only 45.6% class-level Pass@1, with a 17.7-point gap between the strongest and weakest models, confirming the benchmark's discriminative power. Strategy choice strongly interacts with model capability: structured approaches such as bottom-up improve weaker models by up to 9.4 percentage points, while compositional generation collapses to as low as 1.3%. Error analysis over 500 manually annotated failures reveals that logic errors (56.2%) and dependency errors (38.0%) dominate, identifying cross-method coordination as the core bottleneck.

Creating interactive STEM courseware traditionally requires HTML/CSS/JavaScript expertise, leaving barriers for educators. While generative AI can produce HTML codes, existing tools generate static presentations rather than interactive simulations, struggle with long documents, and lack pedagogical accuracy mechanisms. Furthermore, full regeneration for modifications requires 200--600 seconds, disrupting creative flow. We present MAIC-UI, a zero-code authoring system that enables educators to create and rapidly edit interactive courseware from textbooks, PPTs, and PDFs. MAIC-UI employs: (1) structured knowledge analysis with multi-modal understanding to ensure pedagogical rigor; (2) a two-stage generate-verify-optimize pipeline separating content alignment from visual refinement; and (3) Click-to-Locate editing with Unified Diff-based incremental generation achieving sub-10-second iteration cycles. A controlled lab study with 40 participants shows MAIC-UI reduces editing iterations (4.9 vs. 7.0) and significantly improves learnability and controllability compared to direct Text-to-HTML generation. A three-month classroom deployment with 53 high school students demonstrates that MAIC-UI fosters learning agency and reduces outcome disparities -- the pilot class achieved 9.21-point gains in STEM subjects compared to -2.32 points in control classes. Our code is available at https://github.com/THU-MAIC/MAIC-UI.

AI coding agents powered by large language models can read codebases and produce functional code, but they routinely violate team-specific product decisions that are invisible in the source code alone. We introduce a controlled benchmark measuring decision compliance, the rate at which an AI coding agent follows established product, design, and engineering decisions, across 8 realistic software engineering tasks containing 41 weighted decision points. We compare a baseline configuration (Claude Code with codebase access only) against an augmented configuration that adds Brief, a product-context retrieval system providing spec generation, mid-build consultation, and retrieval of recorded decisions, persona pain points, customer signals, and competitive intelligence. On identical prompts and the same repository, the augmented configuration achieves 95% decision compliance versus 46% for the baseline, a 49 percentage point improvement. Per-decision analysis reveals that the baseline achieves 100% compliance on decisions visible in the codebase and 0-33% on decisions requiring product context, suggesting that product-context retrieval is a key driver of the improvement. We release the benchmark repository, all 16 pull requests, and scoring harness for independent reproduction.

Adaptive programming practice often relies on fixed libraries of worked examples and practice problems, which require substantial authoring effort and may not correspond well to the logical errors and partial solutions students produce while writing code. As a result, students may receive learning content that does not directly address the concepts they are working to understand, while instructors must either invest additional effort in expanding content libraries or accept a coarse level of personalization. We present an approach for knowledge-component (KC) guided educational content generation using pattern-based KCs extracted from student code. Given a problem statement and student submissions, our pipeline extracts recurring structural KC patterns from students' code through AST-based analysis and uses them to condition a generative model. In this study, we apply this approach to worked example generation, and compare baseline and KC-conditioned outputs through expert evaluation. Results suggest that KC-conditioned generation improves topical focus and relevance to students' underlying logical errors, providing evidence that KC-based steering of generative models can support personalized learning at scale.

Large language models are widely used for code generation, yet they rely on an implicit assumption that the task descriptions are sufficiently detailed and well-formed. However, in practice, users may provide defective descriptions, which can have a strong effect on code correctness. To address this issue, we develop SpecValidator, a lightweight classifier based on a small model that has been parameter-efficiently finetuned, to automatically detect task description defects. We evaluate SpecValidator on three types of defects, Lexical Vagueness, Under-Specification and Syntax-Formatting on 3 benchmarks with task descriptions of varying structure and complexity. Our results show that SpecValidator achieves defect detection of F1 = 0.804 and MCC = 0.745, significantly outperforming GPT-5-mini (F1 = 0.469 and MCC = 0.281) and Claude Sonnet 4 (F1 = 0.518 and MCC = 0.359). Perhaps more importantly, our analysis indicates that SpecValidator can generalize to unseen issues and detect unknown Under-Specification defects in the original (real) descriptions of the benchmarks used. Our results also show that the robustness of LLMs in task description defects depends primarily on the type of defect and the characteristics of the task description, rather than the capacity of the model, with Under-Specification defects being the most severe. We further found that benchmarks with richer contextual grounding, such as LiveCodeBench, exhibit substantially greater resilience, highlighting the importance of structured task descriptions for reliable LLM-based code generation.

Large language models (LLMs) perform strongly on general-purpose code generation, yet their applicability to enterprise domain-specific languages (DSLs) remains underexplored, especially for repository-scale change generation spanning multiple files and folder structures from a single natural-language (NL) instruction. We report an industrial case study at BMW that adapts code-oriented LLMs to generate and modify project-root DSL artifacts for an Xtext-based DSL that drives downstream Java/TypeScript code generation. We develop an end-to-end pipeline for dataset construction, multi-file task representation, model adaptation, and evaluation. We encode DSL folder hierarchies as structured, path-preserving JSON, allowing single-response generation at repository scale and learning cross-file dependencies. We evaluate two instruction-tuned code LLMs (Qwen2.5-Coder and DeepSeek-Coder, 7B) under three configurations: baseline prompting, one-shot in-context learning, and parameter-efficient fine-tuning (QLoRA). Beyond standard similarity metrics, we introduce task-specific measures that assess edit correctness and repository structural fidelity. Fine-tuning yields the most significant gains across models and metrics, achieving high exact-match accuracy, substantial edit similarity, and structural fidelity of 1.00 on our held-out set for multi-file outputs. At the same time, one-shot in-context learning provides smaller but consistent improvements over baseline prompting. We further validate practical utility via an expert developer survey and an execution-based check using the existing code generator.

Chart-to-code generation converts a chart image into an executable plotting script, enabling faithful reproduction and editable visualizations. Existing methods are largely Python-centric, limiting practical use and overlooking a critical source of supervision: the same chart can be expressed by semantically equivalent scripts in different plotting languages. To fill this gap, we introduce Chart2NCode, a dataset of 176K charts paired with aligned scripts in Python, R, and LaTeX that render visually equivalent outputs, constructed via a metadata-to-template pipeline with rendering verification and human quality checks. Building on a LLaVA-style architecture, we further propose CharLuMA, a parameter-efficient adaptation module that augments the multimodal projector with a language-conditioned mixture of low-rank subspaces, allowing the model to share core chart understanding while specializing code generation to the target language through lightweight routing. Extensive experiments show consistent gains in executability and visual fidelity across all languages, outperforming strong open-source baselines and remaining competitive with proprietary systems. Further analyses reveal that balanced multi-language supervision benefits all languages and that the adapter allocates a compact shared core plus language-specific capacity. Codes and data are available at https://github.com/Zhihan72/CharLuMA.

Large Language Models (LLMs) have achieved strong performance on general code generation, but their effectiveness drops sharply in enterprise settings where software development relies on internal private libraries absent from public pre-training corpora. Existing Retrieval-Augmented Generation (RAG) methods provide a training-free solution by retrieving static API documentation, but our analysis shows that documentation mainly helps models identify what APIs to use and remains insufficient for teaching how to use them correctly. Even with oracle API-document retrieval, LLMs still make recurring errors at the API, cross-API, and task levels, including API misuse or hallucination, flawed API composition, and incorrect solution strategies. To address this limitation, we propose MEMCoder, a training-free self-evolving memory framework for private-library code generation. MEMCoder augments existing RAG pipelines with a Multi-level Evolving Memory that continuously accumulates and reuses execution-derived Usage Guidelines at the API, cross-API, and task levels. During generation, MEMCoder retrieves both static API documentation and relevant historical memories to guide code generation; after execution, it analyzes feedback to refine memory through a closed loop of generation, execution, reflection, and update. Extensive experiments on NdonnxEval and NumbaEval show that MEMCoder consistently enhances different RAG backbones across LLMs of different scales, yielding an average absolute pass@1 improvement of 18.41 percentage points. Moreover, MEMCoder outperforms existing self-evolving memory methods and validates the effectiveness of organizing execution feedback into multi-level usage memories.

Software documentation frequently becomes outdated or fails to exist entirely, yet developers need focused views of their codebase to understand complex systems. While automated reverse engineering tools can generate UML diagrams from code, they produce overwhelming detail without considering developer intent. We introduce query-driven UML diagram generation, where LLMs create diagrams that directly answer natural language questions about code. Unlike existing methods, our approach produces semantically focused diagrams containing only relevant elements with contextual descriptions. We fine-tune Qwen2.5-Coder-14B on a curated dataset of code files, developer queries, and corresponding diagram representations in a structured JSON format, evaluating with both automatic detection of structural defects and human assessment of semantic relevance. Results demonstrate that fine-tuning on a modest amount of manually corrected data yields dramatic improvements: our best model achieves the highest F1 scores while reducing defect rates below state-of-the-art LLMs, generating diagrams that are both structurally sound and semantically faithful to developer queries. Thus, we establish the feasibility of using LLMs for scalable contextual, on-demand documentation generation. We make our code and dataset publicly available at https://github.com/i-need-a-pencil/query2diagram.

Reward hacking in code generation, where models exploit evaluation loopholes to obtain high reward without correctly solving the intended task, poses a critical challenge for Reinforcement Learning (RL) and the deployment of reasoning models. Existing studies often rely on explicitly prompted hacking trajectories, but it remains unclear whether monitors trained on such data can detect reward hacks that arise without direct hacking instructions during RL training. In this work, we introduce Trace-and-Amplify, a framework for scalable curation of reward-hacking trajectories that arise during RL training without explicit hacking instructions. The framework uses unit-test tracers to identify hacking solutions when they occur and retains such trajectories for monitor training and evaluation. Through controlled comparisons between monitors trained on prompt-elicited hacking trajectories and training-time reward-hacking trajectories collected by Trace-and-Amplify, we find that \textbf{(1) prompt-elicited-data-trained monitors often fail to generalize to trajectories curated by our framework}, and \textbf{(2) monitors trained on our Trace-and-Amplify trajectories demonstrate stronger generalizability to unseen hacking types}. Our results indicate that prompted reward hacking data may not fully reflect training-time reward-hacking behaviors, and that relying solely on these data can lead to misleading conclusions. Codebase is available at https://github.com/LichenLillc/CoTMonitoring.git

Large language models (LLMs) have demonstrated strong performance on a wide range of software engineering tasks, including code generation and analysis. However, most prior work relies on cloud-based models or specialized hardware, limiting practical applicability in privacy-sensitive or resource-constrained environments. In this paper, we present a systematic empirical evaluation of two locally deployed LLMs, LLaMA 3.2 and Mistral, for real-world Python bug detection using the BugsInPy benchmark. We evaluate 349 bugs across 17 projects using a zero-shot prompting approach at the function level and an automated keyword-based evaluation framework. Our results show that locally executed models achieve accuracy between 43% and 45%, while producing a large proportion of partially correct responses that identify problematic code regions without pinpointing the exact fix. Performance varies significantly across projects, highlighting the importance of codebase characteristics. The results demonstrate that local models can identify a meaningful share of bugs, though precise localization remains difficult for locally executed LLMs, particularly when handling complex and context dependent bugs in realistic development scenarios.

Research shows that dialogue, the interactive process through which participants articulate their thinking, plays a central role in constructing shared understanding, coordinating action, and shaping learning outcomes in teams. Analysing dialogue content has been central to advancing team learning theory and informing the design of computer-supported collaborative learning environments, yet this progress has depended on labour-intensive qualitative coding. LLMs offer new possibilities for automating and enhancing the dialogue layer within emerging multimodal learning analytics approaches, with recent studies showing that they can approximate human coding through few-shot prompting. However, prior work has focused on replicating human coding accuracy for research purposes, rather than addressing a more educationally consequential question: how can we design prompts that allow an LLM to label team dialogue accurately and fast enough to be useful in real settings, such as in-person healthcare simulations, where results must be returned quickly and computational cost and sustainability also matter? This paper investigates how prompt design and batching strategies can be optimised to balance coding accuracy, processing time, and environmental impact in team-based healthcare simulation debriefing. Using a dataset of 11,647 utterances coded across 6 dialogue constructs, we compared 4 prompt designs across varying batch sizes, evaluating coding performance, processing time, and energy consumption, as well as the trade-offs between these metrics. Results indicate that increasing batch size improves speed and reduces energy use, but negatively impacts coding performance. Beyond demonstrating the feasibility of LLM-based qualitative analysis, this study offers practical guidance for scaling dialogue analytics in contexts where timeliness, privacy, and sustainability are critical.

Lean 4 autoformalization has become increasingly popular in recent years, with frontier language models and open-weight autoformalizers now producing valid formalizations of mathematical theorems. However, these evaluations often rely on single canonical phrasings of theorems and rarely probe whether outputs are robust to natural variation in inputs, while prior work has shown that semantically equivalent paraphrases often induce divergent formal outputs. We study the structure of these divergences in Lean 4 by applying deterministic paraphrase rules to datasets of undergraduate and Olympiad-level math problems. Across four frontier models and three open-weight autoformalizers, we find that paraphrase sensitivity is dominated by failures at the code-generation layer, and that these failures are typed differently by dataset. Furthermore, these patterns generalize to open-weight models, showing that state-of-the-art autoformalizers still struggle to generate valid Lean code. Our results provide a failure-mode taxonomy for autoformalization and motivate training-time interventions targeted at specific compilation failures.

Existing multi-agent Large Language Model (LLM) frameworks for code generation typically use execution feedback and improve iteratively using Input/Output (I/O) test cases. However, this does not work for scientific workflows, where I/O test cases do not exist, and generating them requires solving the very problem at hand. To address this, we introduce MOSAIC, a training-free multi-agent framework for scientific code generation without I/O supervision. Instead of execution feedback, MOSAIC employs a student-teacher knowledge distillation framework that grounds generation through domain-specific examples and structured problem decomposition. To further mitigate hallucinations across chained subproblems, we introduce a Consolidated Context Window (CCW) for maintaining consistent reasoning across agents. Experiments on the SciCode benchmark show that MOSAIC improves accuracy, executability, and numerical precision over existing approaches while relying on lightweight models.

Automatically generating formal ontologies from unstructured natural language remains a central challenge in knowledge engineering. While large language models (LLMs) show promise, it remains unclear which architectural design choices drive generation quality and why current approaches fail. We present a controlled experimental study using domain-specific insurance contracts to investigate these questions. We first establish a single-agent LLM baseline, identifying key failure modes such as poor Ontology Design Pattern compliance, structural redundancy, and ineffective iterative repair. We then introduce a multi-agent architecture that decomposes ontology construction into four artifact-driven roles: Domain Expert, Manager, Coder, and Quality Assurer. We evaluate performance across architectural quality (via a panel of heterogeneous LLM judges) and functional usability (via competency question driven SPARQL evaluation with complementary retrieval augmented generation based assessment). Results show that the multi-agent approach significantly improves structural quality and modestly enhances queryability, with gains driven primarily by front-loaded planning. These findings highlight planning-first, artifact-driven generation as a promising and more auditable path toward scalable automated ontology engineering.

Large Language Models (LLMs) show promise in automated software engineering, yet their guarantee of correctness is frequently undermined by erroneous or hallucinated code. To enforce model honesty, formal verification requires LLMs to synthesize implementation logic alongside formal specifications that are subsequently proven correct by a mathematical verifier. However, the transition from informal natural language to precise formal specification remains an arduous task. Our work addresses this by providing the NaturalLanguage2VerifiedCode (NL2VC)-60 dataset: a collection of 60 complex algorithmic problems. We evaluate 11 randomly selected problem sets across seven open-weight LLMs using a tiered prompting strategy: contextless prompts, signature prompts providing structural anchors, and self-healing prompts utilizing iterative feedback from the Dafny verifier. To address vacuous verification, where models satisfy verifiers with trivial specifications, we integrate the uDebug platform to ensure functional validation. Our results show that while contextless prompting leads to near-universal failure, structural signatures and iterative self-healing facilitate a dramatic performance turnaround. Specifically, Gemma 4-31B achieved a 90.91% verification success rate, while GPT-OSS 120B rose from zero to 81.82% success with signature-guided feedback. These findings indicate that formal verification is now attainable for open-weight LLMs, which serve as effective apprentices for synthesizing complex annotations and facilitating high-assurance software development.

Chart-to-code generation demands strict visual precision and syntactic correctness from Vision-Language Models (VLMs). However, existing approaches are fundamentally constrained by data-centric limitations: despite the availability of growing chart-to-code datasets, simply scaling homogeneous chart-code pairs conflates visual perception with program logic, preventing models from fully leveraging the richness of multimodal supervision. We present CharTide, a novel data-centric framework that systematically redesigns both training and alignment data for chart-to-code generation. First, we construct a 2M-sample dataset via a Tri-Perspective Tuning strategy, explicitly decoupling training into visual perception, pure-text code logic, and modality fusion streams, enabling a 7B model to surpass specialized baselines using only supervised data. Second, we reformulate alignment as a data verification problem rather than a heuristic scoring task. To this end, we introduce an Inquiry-Driven RL framework grounded in the principle of information invariance: a downstream model should yield consistent answers to identical visual queries across both original and generated charts. Moving beyond rigid rule matching or VLM scoring, we employ a frozen Inspector to objectively verify generated charts through atomic QA tasks, providing verifiable reward signals based on answer accuracy. Experiments on ChartMimic, Plot2Code, and ChartX show that CharTide-7B/8B significantly outperforms open-source baselines, surpasses GPT-4o, and is competitive with GPT-5.

The capabilities of AI-assisted coding are progressing at breakneck speed. Chat-based vibe coding has evolved into fully fledged AI-assisted, agentic software development using agent scaffolds where the human developer creates a plan that agentic AIs implement. One current trend is utilizing documents beyond this plan document, such as project and method-scoped documents. Here we propose GROUNDING$.$md, a community-governed, field-scoped epistemic grounding document, using mass spectrometry-based proteomics as an example. This explicit field-scoped grounding document encodes Hard Constraints (non-negotiable validity invariants empirically required for scientific correctness) and Convention Parameters (community-agreed defaults) that override all other contexts to enforce validity, regardless of what the user prompts. In practice, this will empower a non-domain expert to generate code, tools, and software that have best practices baked in at the ground level, providing confidence to the software developer but also to those reviewing or using the final product. Undoubtedly it is easier to have agentic AIs adhere to guidelines than humans, and this opportunity allows for organizations to develop epistemic grounding documents in such a way as to keep domain experts in the loop in a future of democratized generation of bespoke software solutions.

Prior work evaluates code generation bias primarily through simple conditional statements, which represent only a narrow slice of real-world programming and reveal solely overt, explicitly encoded bias. We demonstrate that this approach dramatically underestimates bias in practice by examining a more realistic task: generating machine learning (ML) pipelines. Testing both code-specialized and general-instruction large language models, we find that generated pipelines exhibit significant bias during feature selection. Sensitive attributes appear in 87.7% of cases on average, despite models demonstrably excluding irrelevant features (e.g., including "race" while dropping "favorite color" for credit scoring). This bias is substantially more prevalent than that captured by conditional statements, where sensitive attributes appear in only 59.2% of cases. These findings are robust across prompt mitigation strategies, varying numbers of attributes, and different pipeline difficulty levels. Our results challenge simple conditionals as valid proxies for bias evaluation and suggest current benchmarks underestimate bias risk in practical deployments.

Multi-agent systems are frequently employed for autonomous code generation, demonstrating strong utility in complex algorithmic problem-solving. Recent studies tackle the difficulty of producing functionally correct programs by leveraging simulation-guided planning and debugging, wherein language models step through execution traces to validate logic. Nevertheless, these methods rely heavily on human-authored public test cases to anchor the simulation and debugging cycles. Hand-crafting exhaustive input-output pairs creates a significant, labor-intensive bottleneck within the software development lifecycle. Since ground-truth examples are seldom accessible before actual implementation in real-world scenarios, this reliance limits existing approaches primarily to curated competitive programming datasets. Additionally, we demonstrate that depending on these public tests creates an "overconfidence gap," leading frameworks to overfit to basic examples and underperform on hidden test suites. Conversely, we note that external input samples are not an absolute requirement for successful code generation. We show that large language models possess the capability to autonomously construct valid inputs and simulate execution flows for self-correction. Building on this, we introduce DryRUN, a framework that removes the necessity for ground-truth data by enabling the LLM to iteratively plan, synthesize its own test inputs, and run simulated executions, thereby mitigating algorithmic overconfidence. Assessments using the LiveCodeBench v6 dataset (post-March 2025) reveal that DryRUN achieves comparable performance to CodeSIM, a state-of-the-art, test-dependent baseline. Notably, it does so entirely without public tests or external execution signals, all while decreasing overall output token usage.

Multi-domain detection of the machine-generated code snippets in various programming languages is a challenging task. SemEval-2026 Task~13 copes with this challenge in various angles, as a binary detection problem as well as attribution of the source. Specifically, its subtasks also cover generator LLM family detection, as well as a hybrid code co-generated by humans and machines, or adversarially modified codes hiding its origin. Our submitted systems adjusted the existing mdok approach (focused on machine-generated text detection) to these specific kinds of problems by exploring various base models, more suitable for code understanding. The results indicate that the submitted systems are competitive in all three subtasks. However, the margins from the top-performing systems are significant, and thus further improvements are possible.

Small language models (1-3B) are practical to run locally, but individually limited on harder code generation tasks. We ask whether composing them into pipelines can recover some of that lost capability. We study code generation pipelines built from 1-3B models with execution feedback, and use a NEAT-inspired evolutionary search to test whether more complex pipeline structure helps beyond a simple refinement loop. We evaluate on HumanEval (164 problems) and sanitized MBPP (427 problems), all with local inference on a single laptop. Self-refinement with execution feedback improves code generation by more than 4 standard deviations on both benchmarks. The gains are narrow in mechanism: refinement fixes many runtime errors (especially NameError and SyntaxError), but rarely fixes logic errors such as AssertionError. Within our tested general-purpose model pool, generator identity mattered less than refiner capability: a 1.5B generator paired with a 3B refiner matched a 3B model doing both roles. Early stopping is essential; without it, every iteration is net-negative. The code-specialized models outperform every general-purpose pipeline configuration, suggesting model specialization matters more than pipeline architecture. Preliminary text-only pipeline experiments without execution feedback did not show gains at this scale. In our constrained search space, evolutionary search mostly rediscovered the same simple generate-execute-refine loop we found manually, with no clearly significant gain from added topology. Single-evaluation fitness inflates results by 5-7 percent, selecting lucky genomes over good ones. On these benchmarks at 1-3B scale, execution feedback mattered more than added pipeline complexity in determining whether composition helped.

The formal verification of operating system kernels requires precise specifications that capture the intended behavior of system calls. Writing these specifications manually demands deep domain expertise, motivating the use of large language models (LLMs) to automate the process. However, in OSV-Bench, a benchmark of 245 specification generation tasks derived from the Hyperkernel OS kernel, the best reported Pass@1 is 55.10%. We propose a domain knowledge prompting method (BODHI), which augments the standard few-shot prompt with a structured C-to-Python translation guide covering 15 categories of domain-specific translation patterns. Inspired by Structured Chain-of-Thought (SCoT) prompting, the guide organizes translation by separation of concerns, addressing pre-condition extraction and post-condition generation as distinct categories. Evaluated on nine models from six providers (Anthropic, Mistral, Amazon, DeepSeek, Meta, Alibaba), covering dense, mixture-of-experts and reasoning architectures, BODHI improves every model tested, with gains ranging from +11% to +32%. The best configuration (Claude Opus 4.6 + BODHI) reaches 96.73% Pass@1. BODHI reduces both syntax and semantic errors, with the strongest effect on models that have sufficient instruction-following capability to utilize structured reference material. These results demonstrate that domain knowledge injection is a model-agnostic technique that substantially bridges the gap between general-purpose code generation and formal specification synthesis.

AI coding agents are being adopted at scale, yet we lack empirical evidence on how people actually use them and how much of their output is useful in practice. We present SWE-chat, the first large-scale dataset of real coding agent sessions collected from open-source developers in the wild. The dataset currently contains 6,000 sessions, comprising more than 63,000 user prompts and 355,000 agent tool calls. SWE-chat is a living dataset; our collection pipeline automatically and continually discovers and processes sessions from public repositories. Leveraging SWE-chat, we provide an initial empirical characterization of real-world coding agent usage and failure modes. We find that coding patterns are bimodal: in 41% of sessions, agents author virtually all committed code ("vibe coding"), while in 23%, humans write all code themselves. Despite rapidly improving capabilities, coding agents remain inefficient in natural settings. Just 44% of all agent-produced code survives into user commits, and agent-written code introduces more security vulnerabilities than code authored by humans. Furthermore, users push back against agent outputs -- through corrections, failure reports, and interruptions -- in 44% of all turns. By capturing complete interaction traces with human vs. agent code authorship attribution, SWE-chat provides an empirical foundation for moving beyond curated benchmarks towards an evidence-based understanding of how AI agents perform in real developer workflows.

Large language models can generate plausible game code, but turning this capability into \emph{iterative creative improvement} remains difficult. In practice, single-shot generation often produces brittle runtime behavior, weak accumulation of experience across versions, and creativity scores that are too subjective to serve as reliable optimization signals. A further limitation is that mechanics are frequently treated only as post-hoc descriptions, rather than as explicit objects that can be planned, tracked, preserved, and evaluated during generation. This report presents \textbf{CreativeGame}, a multi-agent system for iterative HTML5 game generation that addresses these issues through four coupled ideas: a proxy reward centered on programmatic signals rather than pure LLM judgment; lineage-scoped memory for cross-version experience accumulation; runtime validation integrated into both repair and reward; and a mechanic-guided planning loop in which retrieved mechanic knowledge is converted into an explicit mechanic plan before code generation begins. The goal is not merely to produce a playable artifact in one step, but to support interpretable version-to-version evolution. The current system contains 71 stored lineages, 88 saved nodes, and a 774-entry global mechanic archive, implemented in 6{,}181 lines of Python together with inspection and visualization tooling. The system is therefore substantial enough to support architectural analysis, reward inspection, and real lineage-level case studies rather than only prompt-level demos. A real 4-generation lineage shows that mechanic-level innovation can emerge in later versions and can be inspected directly through version-to-version records. The central contribution is therefore not only game generation, but a concrete pipeline for observing progressive evolution through explicit mechanic change.

AI coding assistants increasingly generate code alongside tests. How developers structure test code, whether inline with the implementation or in separate blocks, has traditionally been a matter of testing philosophy. We investigate whether this choice affects AI code generation quality. We conduct a large-scale empirical study (830+ generated files, 12 models, 3 providers) using SEGA, a three-dimensional evaluation framework measuring Determinism, Preservation, and Correctness. Comparing inline test syntax (Python doctests) against separated test syntax (Rust #[test] blocks) on a d-ary heap implementation, we find that: (1) inline tests yield near-perfect preservation (100%) and correctness (92-100%) across all models; (2) separated tests expose stark model-tier gaps (0-100% correctness) and independence between preservation and correctness; (3) model behavior evolves across generations, and notably one model breaks the test suppression pattern of its three predecessors; (4) mechanistic analysis on 7 open-source architectures (6 transformers and a gated-linear Recurrent Neural Network (RNN)) reveals inline test markers receive 2.8-4.4$\times$ stronger attention in 5/7 models, with causal validation via knockout and steering experiments on the 4 code-specialized transformers and RWKV-6; the co-location mechanism extends to a non-transformer architecture, suggesting the design recommendation is robust to future architectural shifts. In the Foundation Model era, test syntax structure is a software design concern: co-locating tests with implementation code produces measurably better AI-generated code. This arxiv long version includes appendices that further qualify the effect as bounded by both model capability and programming language.

Reward models (RMs) are central to aligning large language models (LLMs) with human preferences, powering RLHF and advanced decoding strategies. While most prior work focuses on single-step generation, real-world applications increasingly adopt multi-stage LLM pipelines, where effective reward guidance remains underexplored. We investigate this through code generation for combinatorial optimization, constructing a pipeline that integrates reward models into both formulation and solution stages. We identify a critical challenge: inconsistency between reward model predictions and actual pipeline execution outcomes. To address this, we propose the Pipeline-Adapted Reward Model (PARM), which leverages pipeline-specific data and direct preference optimization to align rewards with downstream feedback. We instantiate PARM as a two-stage pipeline (formulation -> code generation) and evaluate it on four public optimization benchmarks, measuring execution rate and solving accuracy against baselines and sampling methods. A supplementary cross-domain experiment on GSM8K assesses transferability. Results demonstrate that PARM consistently improves pipeline output quality and stability, providing new insights into reward modeling for multi-stage LLM reasoning.

Recently, code-oriented large language models (LLMs) have demonstrated strong capabilities in translating natural language into executable code. Text-to-SQL is a significant application of this ability, enabling non-technical users to interact with relational databases using natural language. However, state-of-the-art models continue to struggle with highly complex logic, particularly deeply nested statements involving multiple joins and conditions, as well as with real-world database schemas that are noisy or poorly structured. In this paper, we investigate whether curriculum learning can improve the performance of code-based LLMs on Text-to-SQL tasks. Employing benchmarks including Spider and BIRD, we fine-tune models under different curriculum strategies. Our experiments show that naive curriculum, simply ordering training samples by complexity in a single epoch, fails to surpass standard fine-tuning due to catastrophic forgetting. To overcome this, we propose a Modular Adapter Composition (MAC) strategy. By sequentially training tier-specific adapters on incremental complexity levels (Easy to Extra-Hard), we create a scaffolded learning environment that improves performance on complex queries. Our approach not only produces measurable performance gains on the Spider and BIRD benchmarks but also provides a flexible, "Lego-like" architecture, allowing models to be composed and deployed based on specific schema difficulty requirements. These findings demonstrate that structured, modular learning is a superior alternative to monolithic fine-tuning for mastering the syntax and logic of complex code generation.

Transpilation, or code translation, aims to convert source code from one programming language (PL) to another. It is beneficial for many downstream applications, from modernizing large legacy codebases to augmenting data for low-resource PLs. Recent large language model (LLM)-based approaches have demonstrated immense potential for code translation. Among these approaches, training-based methods are particularly important because LLMs currently do not effectively adapt to domain-specific settings that suffer from a lack of knowledge without targeted training. This limitation is evident in transpilation tasks involving low-resource PLs. However, existing training-based approaches rely on a pairwise transpilation paradigm, making it impractical to support a diverse range of PLs. This limitation is particularly prominent for low-resource PLs due to a scarcity of training data. Furthermore, these methods suffer from suboptimal reinforcement learning (RL) reward formulations. To address these limitations, we propose CodePivot, a training framework that leverages Python as an intermediate representation (IR), augmented by a novel RL reward mechanism, Aggressive-Partial-Functional reward, to bootstrap the model's multilingual transpilation ability without requiring parallel corpora. Experiments involving 10 PLs show that the resulting 7B model, trained on Python-to-Others tasks, consistently improves performance across both general and low-resource PL-related transpilation tasks. It outperforms substantially larger mainstream models with hundreds of billions more parameters, such as Deepseek-R1 and Qwen3-235B-A22B-Instruct-2507, on Python-to-Others tasks and Others-to-All tasks, respectively. In addition, it outperforms its counterpart trained directly on Any-to-Any tasks on general transpilation tasks. The code and data are available at https://github.com/lishangyu-hkust/CodePivot.

Code secrets are sensitive assets for software developers, and their leakage poses significant cybersecurity risks. While the rapid development of AI code assistants powered by Code Large Language Models (CLLMs), CLLMs are shown to inadvertently leak such secrets due to a notorious memorization phenomenon. This study first reveals that Byte-Pair Encoding (BPE) tokenization leads to unexpected behavior of secret memorization, which we term as \textit{gibberish bias}. Specifically, we identified that some secrets are among the easiest for CLLMs to memorize. These secrets yield high character-level entropy, but low token-level entropy. Then, this paper supports the biased claim with numerical data. We identified that the roots of the bias are the token distribution shift between the CLLM training data and the secret data. We further discuss how gibberish bias manifests under the ``larger vocabulary'' trend. To conclude the paper, we discuss potential mitigation strategies and the broader implications on current tokenizer design.

Recent advances in large language models have highlighted their potential to automate computational research, particularly reproducing experimental results. However, existing approaches still use fixed sequential agent pipelines with weak global coordination, which limits their robustness and overall performance. In this work, we propose Hierarchical Research Agent System (HiRAS), a hierarchical multi-agent framework for end-to-end experiment reproduction that employs supervisory manager agents to coordinate specialised agents across fine-grained stages. We also identify limitations in the reference-free evaluation of the Paper2Code benchmark and introduce Paper2Code-Extra (P2C-Ex), a refined protocol that incorporates repository-level information and better aligns with the original reference-based metric. We conduct extensive evaluation, validating the effectiveness and robustness of our proposed methods, and observing improvements, including >10% relative performance gain beyond the previous state-of-the-art using open-source backbone models and significantly reduced hallucination in evaluation. Our work is available on GitHub: https://github.com/KOU-199024/HiRAS.

Practitioners have reported a directional pattern in AI-assisted code generation: AI-generated code tends to fail quietly, preserving the appearance of functionality while degrading or concealing guarantees. This paper introduces the Reward-Shaped Failure Hypothesis - the proposal that this pattern may reflect an artifact of optimization through human feedback rather than a random distribution of bugs. We define failure truthfulness as the property that a system's observable outputs accurately represent its internal success or failure state. We then present AIRA (AI-Induced Risk Audit), a deterministic 15-check inspection framework designed to detect failure-untruthful patterns in code. We report results from three studies: (1) an anonymized enterprise environment audit, (2) a balanced 600-file public corpus pilot, and (3) a strict matched-control replication comparing 955 AI-attributed files against 955 human-control files. In the final replication, AI-attributed files show 0.435 high-severity findings per file versus 0.242 in human controls (1.80x). The effect is consistent across JavaScript, Python, and TypeScript, with strongest concentration in exception-handling-related patterns. These findings are consistent with a directional skew toward fail-soft behavior in AI-assisted code. AIRA is designed for governance, compliance, and safety-critical systems where fail-closed behavior is required.

Automated simulator construction requires distributional fidelity, distinguishing it from generic code generation. We identify two failure modes in long-horizon LLM agents: contextual drift and optimization instability arising from conflating structural and parametric errors. We propose SOCIA-EVO, a dual-anchored evolutionary framework. SOCIA-EVO introduces: (1) a static blueprint to enforce empirical constraints; (2) a bi-level optimization to decouple structural refinement from parameter calibration; and (3) a self-curating Strategy Playbook that manages remedial hypotheses via Bayesian-weighted retrieval. By falsifying ineffective strategies through execution feedback, SOCIA-EVO achieves robust convergence, generating simulators that are statistically consistent with observational data. The code and data of SOCIA-EVO are available here: https://github.com/cruiseresearchgroup/SOCIA/tree/evo.

LLMs are widely used for code generation and mathematical reasoning tasks where they are required to generate structured output. They either need to reason about code, generate code for a given specification, or reason using programs of thought. The typical approach to code generation is to prompt the model and generate samples until an appropriate program is obtained. Within this process, sampling $n$ programs from the language model requires $n$ GPU compute-intensive generations which becomes prohibitively expensive for larger values of $n$. In this work, we address this limitation by exposing the LLM's distribution within the generated programs themselves. We propose a novel test-time framework we dub probabilistic programs of thought to obtain more samples from the model with fewer LLM generations. Given a program generated by a model and the associated next-token probabilities, we build a probabilistic program that compactly represents exponentially many deterministic programs. Since performing probabilistic reasoning in this probabilistic program is much cheaper, our approach allows sampling new programs without any additional GPU compute and little CPU overhead. We instantiate our approach on benchmarks for code generation, code understanding and mathematical reasoning and report improvements in performance with fewer generations from the LLM.

In LLM-based code generation, multiple code candidates are often generated in parallel from the same prompt -- for example, in best-of-N sampling or multi-candidate code completion. These requests can share KV caches through a common prefix, yet the extent to which their Mixture-of-Experts (MoE) expert routing overlaps, and how this overlap varies across layers, remains insufficiently understood. We study Qwen3.5-35B-A3B-FP8 (256 routed experts, top-8) by performing tree-search-based branching generation from a shared prefix (851 completed codes, temperature 0.7) and analyzing the results with a compiler-output-based alignment (gcc -S -O0 assembly) that controls for token-identity confounds. Our findings are threefold: (1) At positions where both sequences generated the same token, Jaccard similarity reaches 0.649 (40x random), while even at positions with different tokens it remains 0.175 (11x random). (2) A layer-wise decomposition reveals a crossing pattern: same-token routing similarity exceeds different-token similarity across all layers, but dips in the middle layers (L14-20), while different-token similarity peaks in the middle layers at 14x random. (3) In tree-search code generation, 67% of successfully compiled codes concentrate in the top three assembly-equivalent groups, and 99.6% of within-group differences consist of comments and blank lines. We show that diversity in top-P search, including beam search, poses a significant challenge. These results refine the "context-independent routing" claim of prior work through layer-wise decomposition and suggest opportunities for improving search efficiency in LLM code generation.

Neural Architecture Search (NAS) aims to automatically discover high-performing deep neural network (DNN) architectures. However, conventional algorithm-driven NAS relies on carefully hand-crafted search spaces to ensure executability, which restricts open-ended exploration. Recent coding-based agentic approaches using large language models (LLMs) reduce manual design, but current LLMs struggle to reliably generate complex, valid architectures, and their proposals are often biased toward a narrow set of patterns observed in their training data. To bridge reliable algorithmic search with powerful LLM assistance, we propose LLMasTool, a hierarchical tree-based NAS framework for stable and open-ended model evolution. Our method automatically extracts reusable modules from arbitrary source code and represents full architectures as hierarchical trees, enabling evolution through reliable tree transformations rather than code generation. At each evolution step, coarse-level planning is governed by a diversity-guided algorithm that leverages Bayesian modeling to improve exploration efficiency, while the LLM resolves the remaining degrees of freedom to ensure a meaningful evolutionary trajectory and an executable generated architecture. With this formulation, instead of fully agentic LLM approaches, our method explores diverse directions beyond the inherent biases in the LLM. Our method improves over existing NAS methods by 0.69, 1.83, and 2.68 points on CIFAR-10, CIFAR-100, and ImageNet16-120, demonstrating its effectiveness.

Large Language Models (LLMs) excel at code generation but remain heavily reliant on large-scale annotated solutions and verification-based supervision, which constrains scalability and hinders sustained self-improvement. Recent solver--verifier frameworks exploit program execution as an automatic supervision signal, but their effectiveness degrades as solvers become moderately strong: verifier-generated tests increasingly confirm semantic correctness rather than exposing the remaining failure modes. We propose \textbf{ACE}, a self-evolving code generation framework based on a solver--adversary architecture that prioritizes active failure discovery through execution-centric supervision. A single LLM alternates between generating candidate programs and producing adversarial unit test inputs optimized to induce execution-level failures, such as runtime errors, exceptions, or non-termination. Supervision is derived solely from execution outcomes: robust programs are selected for supervised fine-tuning, while adversarial tests are optimized via Kahneman--Tversky Optimization using execution-derived preferences. Notably, the entire training loop requires no ground-truth code or external reward models. Experiments on CodeContests, MBPP, and LiveCodeBench demonstrate that ACE consistently outperforms strong solver--verifier baselines, achieving 3--7% absolute gains in pass@1, with larger improvements on out-of-distribution benchmarks, while maintaining competitive or improved inference efficiency.

Code search, framed as information retrieval (IR), underpins modern software engineering and increasingly powers retrieval-augmented generation (RAG), improving code discovery, reuse, and the reliability of LLM-based coding. Yet existing code IR models remain largely text-centric and often overlook the visual and structural aspects inherent in programming artifacts such as web interfaces, data visualizations, SVGs, schematic diagrams, and UML. To bridge this gap, we introduce MMCoIR, the first comprehensive benchmark for evaluating multimodal code IR across five visual domains, eight programming languages, eleven libraries, and show the challenge of the task through extensive evaluation. Therefore, we then propose CodeMMR, a unified retrieval model that jointly embeds natural language, code, and images into a shared semantic space through instruction-based multimodal alignment. CodeMMR achieves strong generalization across modalities and languages, outperforming competitive baselines (e.g., UniIR, GME, VLM2Vec) by an average of 10 points on nDCG@10. Moreover, integrating CodeMMR into RAG enhances code generation fidelity and visual grounding on unseen code generation tasks, underscoring the potential of multimodal retrieval as a core enabler for next-generation intelligent programming systems. Datasets are available at HuggingFace.