Two-Stage Reinforcement Learning for Sound and Adversarial Test Generation in Code LLMs

Reinforcement learning (RL) has substantially advanced code generation with large language models (LLMs) through executable feedback. The feedback for coding problems mainly comes from specific test cases, where high-quality test cases are often scarce since they should be both sound and discriminative. We thus turn to study the auto-generation of test cases using the learned model. We find this is naturally an adversarial RL problem: the model is expected to generate effective test cases as counterexamples, depending on the solver's current failure modes. We propose Test Cases Scaling (TCS), a two-stage RL framework for effective test generation. Both stages train a test generator from a rolling policy-aligned buffer: Stage 1 generates tests consistent with the reference solution, and Stage 2 restricts the buffer to current failure modes and learns counterexample tests. Across TACO and LiveCodeBench, TCS improves both pass@1 and inference-time answer selection according to generated tests. We find the learned test generator also enables effective selection among other LLM outputs.

CodePoisonRAG: Knowledge Poisoning Attacks on Retrieval-Augmented Code Generation

Retrieval-Augmented Code Generation (RACG) improves LLM-based software development by retrieving external code artifacts, documentation, and patches, and incorporating them into the generation context. This reliance on external knowledge introduces a critical trust boundary: poisoned artifacts can influence generated code without modifying the underlying LLM. Prior work shows that selecting existing vulnerable examples can increase the general vulnerability rate of RACG outputs, but leaves open whether a black-box attacker can construct a single task-matched artifact that propagates an attacker-selected weakness. We introduce CodePoisonRAG, a targeted upstream knowledge-poisoning framework that transforms benign fixed-code entries into poisoned artifacts. Its attack chain combines CWE-specific Vulnerability Injection, which embeds a selected source-to-sink flow while retaining task alignment, with Semantic Mislabeling, which adds false safety claims without repairing the vulnerable behavior. The attacker has no access to the victim's deployed knowledge base, retriever, re-ranker, generator, prompt, or defense mechanism and injects at most one artifact per anticipated programming task. We construct 85 poisoned artifacts covering ten CWE classes across Java and C, yielding an aggregate corpus-poisoning ratio of 0.7%. Across three generators, all 85 artifacts appear among the Top-3 results for their corresponding queries, and CodePoisonRAG achieves attack success rates between 0.80 and 0.93. Against CodeGuarder, which injects vulnerability-specific security knowledge into the generation context, the attack retains success rates between 0.40 and 0.71. These results show that RACG poisoning extends beyond the incidental propagation of existing vulnerabilities to the targeted construction and propagation of attacker-selected weaknesses.

HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?

As agents move from research prototypes to deployed tools, their capability increasingly depends on model-external execution infrastructure, commonly termed the agent harness. Changing this harness while holding model weights fixed can substantially alter task performance. Current agent evaluations typically report downstream performance under a chosen harness, leaving a model's ability to develop the harness itself comparatively underexplored. We introduce HarnessDev, a benchmark that shifts the unit of evaluation from task outputs to runnable infrastructure. HarnessDev covers two stages. In Creation, the agent starts from a minimal seed and a small number of cases, then builds a complete execution system. In Evolution, it starts from its own created harness and iteratively revises it using downstream execution feedback, with the goal of improving benchmark performance. We then evaluate each constructed harness on capability (task success on held-out benchmarks) and efficiency (execution-token cost). The reported Creation results cover six creator LLMs, four domains, and five downstream benchmarks totaling 2,207 unique downstream instances, with hidden evaluation tasks withheld from development. We find that generated harnesses remain substantially behind mature human-engineered references on code and on search and research, while matching or exceeding the selected references on writing and machine-learning experimentation, with large variation in execution cost. Evolution produces some performance gains, but they are unstable and transfer only partially to held-out tasks. Experiments with a fixed runtime model further show that the gains depend strongly on the model executing the harness, indicating limited transfer across models.

Hints Help But Do They Teach? Evaluating Skills Transfer in Code Generation

When a hint turns a failing generated program into a passing one, does it provide missing information or merely steer the model toward a solution it could already produce? We test these hypotheses on HumanEval+ and MBPP+ using executable evaluation. For Qwen2.5-3B-Instruct, adaptive relevant hints rescue 36 of 79 selected failures; an unrelated hint rescues 19, while eight unhinted samples solve 46 and recover 31 of the 36 relevant-hint rescues. Phi-3.5-mini shows the same pattern: relevant hints rescue 42 of 101 failures, an unrelated hint rescues 17, and unhinted sampling solves 57, including 36 of the 42 relevant-hint rescues. Because the hint conditions use different attempt budgets, these comparisons do not isolate a purely semantic effect. Mechanistic tests on Qwen identify a stable activation direction shared by relevant and unrelated hints. Persistently adding this direction yields 14 rescues and 18 regressions, with no detectable net accuracy gain; learned low-rank interventions have a positive but imprecise estimated effect. Full textual specifications solve 22 of 24 context-defined problems, versus 5-11 for tested virtual-KV prefixes. Post-generation hidden-state probes transfer across benchmarks, with pooled AUROC 0.806 and 0.780, but their top-one selection advantage over token confidence is statistically unresolved. Overall, relevant hints can rescue failures, but most rescued solutions are already reachable through ordinary sampling, and the internal interventions tested here do not establish task-general capability transfer.

Predicting Program Exit Code with LLMs and Programming Language Semantics

Large language models (LLMs) have shown proficiency in various software engineering tasks, such as code generation and translation. However, a key limitation in their performance may be their (lack of) understanding of programming-language semantics. Even when explicit semantics are given, it remains unclear whether LLMs apply those rules or lean on priors learned during pre-training instead. We study if LLMs lean on priors or given semantics with a novel task--Program Executability Prediction (PrEx)--that asks models to predict whether a program is semantically valid or invalid (and, if invalid, which formal rule it violates) given the program's syntax and operational semantics. Because PrEx requires both valid and invalid programs, we build a dataset with systematically generated invalid transformations derived from valid programs. We evaluate open-source coding LLMs under two semantic formalisms and two semantic shifts across Human-Written, LLM-Translated, and Fuzzer-Generated program splits. Our findings show that LLMs lean on pre-training priors rather than systematically applying the given rules, performing especially poorly on modified semantics and degrading further as program complexity increases. PrEx is available at https://github.com/EngineeringSoftware/prex.

WiseSpec: Requirements-Driven Agents for Code Generation

Code generation aims to automatically generate source code from task requirements and has attracted significant attention with the rapid advancement of large language models (LLMs). Despite remarkable progress, LLMs often struggle to generate correct code for complex software engineering tasks because task descriptions are frequently incomplete, ambiguous, or lack critical contextual information. Existing approaches primarily improve the capabilities of coding agents through more sophisticated tools, skills, and workflows, while largely overlooking the quality of the task requirements themselves. To address this limitation, we draw inspiration from software requirements engineering and propose WiseSpec, a novel requirements-driven agent framework for repository-level code generation. WiseSpec automatically constructs structured and information-rich requirements, assesses their quality through execution-based evaluation, and iteratively refines them to better guide code generation. Experimental results show that WiseSpec consistently outperforms all baselines, achieving an average improvement of 13.17% in %Resolved.

LLM Post-Training as Brownfield Maintenance: An Industrial Perspective on Dataware Engineering

Industrial post-training is a brownfield regime. Teams inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixture budgets without regressing the rest. The maintained artifact is increasingly dataware: behavior governed by a curated post-training mixture, updated via bounded mixture patches rather than clean-slate retraining. From an industrial code-generation improvement effort, we offer a maintainer's perspective on why this work is hard in practice, distilling three recurring challenges, zero-sum mixture design, yield as the binding metric, and end-to-end integration under uncertainty, and arguing that progress depends less on one-off recipes than on an engineering discipline for programming dataware. In our case study, interventions that raised the conversion of teacher distillation into usable training data increased accepted supervision by 2.84 times while using the same solution teacher and four solution attempts per candidate problem. In our primary evaluation, the yield-engineered patch improved CodeForces pass@1 by +2.59 points (+3.11 pass@3) and held-out LiveCodeBench v6 pass@1 by +6.11 (+8.05 pass@3), all statistically significant across 16 stochastic evaluations of each benchmark from one fixed checkpoint per condition, with internal AIME and MATH regression suites within tolerance.

CARVE: Verified Expansion for Variable-Length Generation in Diffusion Language Models

Masked diffusion language models predict tokens from a partially observed response canvas, enabling bidirectional conditioning and parallel token refinement. Yet standard masked-diffusion decoders use a rigid inference interface: the number of masked positions allocated to the answer is fixed before generation begins. Choosing this length is difficult. A short canvas can truncate reasoning or code, while a long canvas wastes computation and can perturb denoising. We introduce CARVE (Counterfactual-Aware Reveal with Verified Expansion), a training-free variable-length algorithm for masked diffusion LMs. Starting from a shorter canvas, CARVE can grow the response during decoding by inserting additional [MASK] positions. Rather than keeping every insertion, CARVE tests a candidate expanded canvas and asks a counterfactual question: would the model make similar predictions for the unresolved positions in the original canvas if the extra masked space were present? The inserted masks are kept only when they induce low Jensen-Shannon (JS) divergence on aligned unresolved positions. This makes length growth a verified stability decision rather than a pure confidence heuristic. CARVE applies without retraining to both full-canvas and blockwise diffusion decoders. Across code generation and mathematical reasoning benchmarks, CARVE consistently improves average performance over fixed-length baselines across all evaluated model families. Crucially, CARVE achieves these accuracy gains while reducing inference cost, reaching half the FLOPs of fixed-length decoding in some settings.

Autoregressive Mosaics: Probing 2D Spatial Reasoning in Text-Only Language Models

Large language models (LLMs) trained only on text and code can sometimes generate programs that draw recognizable images. However, it is unclear whether this reflects an internal representation of 2D spatial layout or simply the ability to translate spatial descriptions into code. We introduce Autoregressive Mosaics (AM-Bench), a benchmark that separates these factors: First, a translation task gives a model a fully specified geometry of a picture in words as a prompt and asks for the code that produces it. Second, a layout task requires the model to compose an image from an underspecified prompt. Across eight open-weight text-and-code-only models, all models reliably translate specified geometry into code, but their open-ended layout performance differs substantially, indicating that these differences are not explained by code-generation ability alone. An output-medium ablation further shows that the interface or medium of expression that the model uses matters: replacing procedural code with raw SVG improves layout scores across all models. Finally, probing model activations shows that a coarse layout plan is present before generation, but reflects only the layout implied by the prompt. During generation, models track the evolving geometric state instead of executing an initially fixed plan. Overall, these results show that 2D spatial performance in text-only LLMs depends on both the model and the output medium, and is not explained by code-generation ability alone.

Interpreting and Steering for Safe and Correct Code Generation

Large language models (LLMs) frequently generate source code containing vulnerabilities, yet little work studies the internal mechanisms that distinguish safe from vulnerable generation in them. In this work, we systematically perform a mechanistic interpretation of LLMs, aiming at both understanding how code safety-vs-vulnerability is represented or driven by components in an LM and turning the insights into actionable steering strategies to encourage safer code generation. To this end, we introduce CodeSec-Pairs, a dataset of 9,342 Python safe-and-vulnerable contrastive code pairs, sampled from Llama-3.1-8B-Instruct. Utilizing the dataset, we explore approaches to localize layers and attention heads that relate to code safety, and further experiment with different steering strategies for inference-time vulnerability reduction. In particular, we propose DuoSteer, a double-steering approach that simultaneously applies safety and code-correctness steering to attention heads. In experiments over five vulnerability types, DuoSteer leads to an average of -26.9% vulnerability rate reduction and +7.5% functional correctness improvement, which outperforms not only other steering variants but also prompting and supervised fine-tuning baselines. The advantage also replicates on Qwen-2.5-Coder-7B-Instruct with another 2,500 contrastive pairs sampled from that model.

InteractBench: Benchmarking LLMs on Competitive Programming under Unrevealed Information

Competitive programming is increasingly being used to evaluate the algorithmic reasoning capabilities of large language models (LLMs). However, existing benchmarks primarily focus on full-information tasks where all problem inputs are provided upfront. This overlooks a critical dimension of algorithmic reasoning: the ability of generated programs to operate when key information is not revealed upfront. Interactive problems, a distinctive component of competitive programming, embody this challenge. These problems require programs to engage in multi-round interaction with an interactor (a judge program) under strict protocol constraints and limited query budgets, with new information revealed only in response to queries. To address this gap, we introduce InteractBench, a benchmark comprising 322 high-quality interactive problems curated from Codeforces, AtCoder, IOI, and ICPC. Each problem is packaged with executable local interactors, enabling fully offline evaluation. Unlike existing benchmarks, InteractBench assesses whether model-generated code can acquire information and track state dynamically. Our evaluation reveals a significant interaction gap: even the most advanced reasoning models achieve limited success on interactive problems. Beyond success rates, we propose a fine-grained failure taxonomy to diagnose the root causes of these deficiencies. Although algorithmic logic errors remain dominant, protocol violations and query-budget overruns are frequent. Code is available at https://github.com/kmsgk0/InteractBench.

Evaluating Tiny Recursive Models Across Training for Code Generation

Code generation increasingly relies on large transformer models, whose capability advances with scale. Yet such a scale is costly, creating demand for small models, especially where data is limited. Recursive models address this by reusing a single block to add depth rather than stacking independent layers. Such models are typically evaluated by teacher-forced fit (next-token loss on ground-truth prefixes) or task accuracy, at a single checkpoint, whereas code is produced by free-running generation, where the model extends its own output. Whether a teacher-forced advantage survives free-running generation, and whether it holds across training, remains open. To study both, we compare a ~28M-parameter autoregressive Tiny Recursive Model (TRM-AR) on natural-language-to-Python code generation against parameter-matched and depth-matched controls, tracking fit and generation across 40 epochs and three seeds. The fit ranking between the recursive model and the depth-matched control reverses twice. Selecting each checkpoint by validation loss and examining the trajectory yields a consistent comparison. At equal parameters, TRM-AR fits, generates, and generalizes better than the parameter-matched control while recovering approximately 45% of the validation-loss gap and 57% of the generation-quality gap between the two controls, at roughly 175 times the per-step cost of the parameter-matched control. However, at equal effective depth, the larger transformer fits and generates better at its validation optimum, suggesting TRM-AR's advantage lies in resistance to overfitting, not greater capability. These findings suggest that recursive code generation models should be evaluated jointly on fit and generation across the training trajectory rather than at a single checkpoint.

Super Library Agent: Joint Generation and Maintenance of Multiple Applications Beyond the Single Codebase

Organizations often develop and maintain portfolios of related applications: independently deployable codebases that share substantial domain logic, interface patterns, or operational conventions. As LLM coding agents are increasingly used to generate and maintain such software, a naive application-by-application workflow duplicates shared logic across codebases and allows prolonged agentic maintenance to accumulate verbosity, dead code, and structural erosion. We introduce the Super Library Agent problem, where an agent sequentially generates a portfolio of N related applications while maintaining a shared Super Library of reusable cross-application components. A minimal sequential scaffold can in principle extract shared code and migrate applications to the evolving library, but in practice suffers from low extraction recall and fragile dependency migration. We address these failures with candidate-guided extraction over code chunk summaries, pre-extraction codebase consolidation, and context-aware migration using extraction traces and call-graph information. Across WebGen-Bench and PaperBench, our method preserves application functionality while significantly reducing redundancy and token footprint (verbosity, token length) over zero-shot, and avoiding the structural erosion introduced by naive library construction, with additional reductions in LOC and MDL. Our code is available at https://github.com/sbigstar0310/super-library-agent.

Development of an Autonomous AI Coding Agent using Monte Carlo Tree Search (MCTS) and Gemini LLM Frameworks

The ongoing changes in software engineering requirements have created a substantial need for automated tools which can create secure source code from natural language input. The performance of traditional Large Language Models (LLMs) becomes limited by their "one-shot" capability which results in logical hallucinations together with reduced algorithmic performance during complicated operations. The research presents an autonomous AI Coding Agent which establishes a connection between LLM-generated content and production-ready software through its organized methodology for decision making. Our framework uses the Gemini 2.5 Flash API for essential reasoning capabilities while employing a tailored Monte Carlo Tree Search (MCTS) method to solve code generation challenges as a search operation. The agent uses a "Self-Critic" evaluator system to test different implementation methods which it ranks according to their accuracy and difficulty level before it improves its operational framework through backpropagation. The system operates through a Flask-based web interface which delivers instant feedback together with syntax highlighting features. Our experimental results show that the MCTS-based method achieves a 92% success rate on complex logical prompts while surpassing standard zero-shot generation models.

Towards Fully Automated Medical Imaging Code Generation via Validation-based Context Engineering

Large language models (LLMs) have demonstrated considerable promise in program generation for small-scale and conventional application development; however, they remain limited when applied to complex, domain-specific tasks such as medical image processing. General-purpose models lack explicit domain knowledge and robust validation mechanisms to ensure correctness, often requiring substantial human intervention to produce reliable processing pipelines. To address these limitations, we propose AutoMedImg, a multi-agent framework for fully automated medical image processing code generation. AutoMedImg orchestrates specialised agents across two phases: a Planning Phase that performs dataset analysis and architecture design with semantic and formal verification, and a Coding Phase that generates modules in parallel with static checking, execution testing, and assembly validation. This multi-stage validation mitigates error propagation throughout generation, while comprehensive auto-context engineering combining domain-specific knowledge bases, shared memory, and validation feedback automates context construction without manual prompting. A cross-project adaptive pipeline synthesis mechanism further accumulates validated pipelines and retrieves proven components for new tasks based on project similarity, enhancing generation efficiency through cross-project learning. Extensive evaluation across six diverse and well-established medical imaging datasets with five backbone LLMs demonstrates that AutoMedImg achieves zero human intervention, with Dice scores of up to 0.90 for segmentation tasks and 99% accuracy for classification.

FigMirror: Ground It, Code It, Plot It

Converting scientific figures into executable code has gained increasing attention, yet existing methods primarily focus on reproducing the reference figure itself. A more practical setting is to plot new data while preserving the visual style of a reference figure (e.g., color scheme and typography). Prior approaches mimic the reference through pixel-level optimization and struggle to carry its style to new data. We show that the key to this task lies in the coordinate grounding and coding capabilities present in modern computer-use models. We propose FigMirror, an agentic framework that unlocks these capabilities through Grounded Measurement, which locates visual elements by coordinates and measures their properties through executable code. We further introduce PlotTwin-Bench, an expert-curated benchmark with fine-grained code and image-level style metrics. Experiments show that FigMirror consistently outperforms existing methods on reference-conditioned style transfer. All plots in this paper are generated by FigMirror, except those produced by other methods for comparison. Our code and data are available at: https://github.com/VILA-Lab/FigMirror.

When Teacher Guidance Misleads: Reward-Aligned On-Policy Distillation

On-policy distillation (OPD) has recently emerged as a popular post-training paradigm for large language models (LLMs), providing an efficient way to transfer the knowledge and capabilities of teacher models into student models. However, teacher guidance on student-generated prefixes is not always reliable. Training should optimize the model to generate responses that are more likely to be correct, or equivalently, to get higher outcome rewards. But during OPD, the teacher model may provide guidance that discourages the student from moving toward correct trajectories or moves the student toward incorrect ones, which is misaligned with outcome reward. Such misaligned guidance is unreliable, as it would mislead the optimization process and ultimately degrade model performance. To mitigate misaligned teacher guidance, we propose Reward-Aligned On-Policy Distillation (RA-OPD). The key insight is to keep only trajectories whose induced updates move the student toward correct trajectories or discourage the student from moving toward incorrect ones. Specifically, for each sampled trajectory, RA-OPD checks whether its trajectory-level distillation return is consistent with its outcome reward and then filters out the misaligned trajectories. RA-OPD selects more reliable trajectories to improve student model performance without requiring additional computational cost. We evaluate RA-OPD on math and code benchmarks using models from the Qwen3 family and the DeepSeek-R1 family. Across seven math benchmarks and three code benchmarks, RA-OPD significantly outperforms standard OPD and other tested OPD variants.

Rubric-to-Code Credit Assignment for Reinforcement Learning

Interactive web application generation requires models to produce usable HTML, CSS, and JavaScript applications from natural language requests. Unlike conventional code generation, application quality depends on multiple user-facing functional requirements, each often tied to localized code regions such as event handlers, state updates, DOM fragments, or CSS selectors. Standard GRPO collapses these structured outcomes into a single sequence-level reward and applies the resulting advantage uniformly to all tokens, weakening credit assignment. We propose \textbf{Rubric-to-Code Credit Assignment} (RCCA), a reinforcement learning framework that converts rubric-level functional feedback into localized optimization signals over generated code. RCCA builds training tasks around explicit functional rubrics, uses a hierarchical reward to separate format, source-code, runtime, and functional failures, and aligns evaluator-generated textual attributions with responsible code spans and generated tokens. The resulting model, \textbf{Ling-RCCA-Flash}, scores 41.25 on MiniAppBench, improving Ling-3.0-Flash by 32.20 points and slightly surpassing Claude Opus 4.5. It also reaches 76.19 on ArtifactsBench, improving the SFT model by 4.48 points and establishing a new top score under the official ArtifactsBench leaderboard setting by surpassing the GPT-5 score by 3.64 points, suggesting transferable implementation-level gains.

Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning

Physical understanding and reasoning depend on forming compact and generalizable representations of the world. While modern vision-language models can recognize and explain diverse physical events, they often lack explicit representations of the underlying mechanisms-such as object states, physical parameters, and governing dynamics-needed for reliably reasoning how the world evolves and responds to interventions. In this work, we introduce Code-as-World, a paradigm that represents physical worlds through executable world representations. By expressing physical composition, dynamic evolution, and visual appearance as executable code, Code-as-World provides a compact, quantitatively grounded, and controllable abstraction of the physical world. To construct such representations from multimodal observations, such as natural-language descriptions or real-world videos, we develop an agentic discovery loop inspired by abductive reasoning, where an agent proposes, executes, renders, verifies, and iteratively refines executable world hypotheses. As a concrete application, we use verified executable worlds to provide scalable physical supervision for training vision-language models on quantitative physical reasoning. Experiments show that Code-as-World-VL achieves state-of-the-art performance on QuantiPhy and surpasses leading proprietary models, highlighting the potential of executable world representations as a scalable foundation for physical intelligence.

Zero-Shot Self-Orchestration with Ledger-Based Control for Improved LLM Coding Performance

Multi-agent large language model systems are widely reported to beat single-model baselines, but the evidence is mixed, and comparisons are usually confounded: pipelines change token budgets, tool calls, and prompts simultaneously, so an aggregate gain rarely reveals what actually helped. We investigate the effect of introducing the manager-worker scaffold over a shared filesystem workspace, with no training and no per-benchmark tuning, measured against the same model answering in a single pass. Across nine models -- five open-weight, spanning 9B to ~2.8T parameters, and four frontier closed models -- on the 100 latest hard LiveCodeBench problems, the scaffold's benefit is real but conditional: large and statistically significant for some (Qwen3.8-27B +23.4, GPT-5.6-Luna +10.6 and GPT-5.6-Terra +8.0, each over five paired passes; Kimi-K3 +30.4 and Minimax-M3 +11.0 over five paired passes with reasoning off, both at $p < 10^{-4}$, and +42 and +12 in a single pass at a 128k cap) and null or negative for others (Qwen3.6-35B -1 to -9 with reasoning off). With the manager, Opus-5 achieves the highest score in the study at 91% in one pass. Running a manager roughly triples the token bill, but it buys accuracy more cheaply than moving to a larger model does: GPT-5.6-Terra with a manager nearly matches Fable 5's single-call accuracy (85.0 against 87.4, $p = 0.59$) at a fifth of the price (\$11.71 against \$61.11 per 100-problem pass, $p < 10^{-4}$), and the Qwen-27B arm does it for \$51.75 on weights anyone can self-host. Our transcript analysis finds several mechanisms behind the gains, of which two recur: context management, in which short worker calls and shared notes organize state and reduce truncation, and problem decomposition. Improvements are modest for large models with reasoning enabled, but larger for some models with reasoning disabled and for smaller models with reasoning enabled.

ReproAgent: Contract-Guided Paper-to-Code Reproduction

Paper-to-code reproduction asks scientific AI agents to turn research papers into executable repositories that preserve the paper's method, protocol and artifacts. This is difficult because the specification is split: explicit paper content such as algorithms, metrics and artifacts is often lost across long agent trajectories, while implicit details such as framework defaults and conventions inherited from related work are absent from the paper. We introduce ReproAgent, a four-stage Prepare--Plan--Generate--Repair pipeline built around a persistent implementation contract with two channels: an implementation-requirement channel that turns paper snippets into code obligations, and a reference-evidence channel that retrieves content and structure evidence from related repositories. Both are bound to work packages, projected into file-level contracts, and consumed across generation and repair. On PaperBench Code-Dev, ReproAgent reaches the highest mean score among same-backbone scaffolds under both Claude-Sonnet-4.5 and Gemini-3-Flash. End-to-end channel ablations and per-paper cases support the contribution of both channels. Code and experimental artifacts are publicly available.

SA-Bench: Evaluating Semantic Alignment in LLM-Based Paper Reproduction

LLM agents can generate paper reproduction code, yet often produce scientifically unfaithful implementations. We define this failure mode as semantic drift, where generated code silently diverges from the paper's specifications. We introduce SemanticAlign-Bench(SA-Bench), a diagnostic benchmark covering 30 papers from ICLR, ICML and NeurIPS 2025. For each paper, we decompose its specifications into atomic and verifiable implementation claims, which we call Semantic Alignment Units (SAUs) and evaluate repositories along four diagnostic dimensions spanning numerical, methodological, protocol and ordering drift. In total, we construct 1,491 SAUs across five ML domains and evaluate 12 generator configurations (4 models $\times$ 3 scaffolds). Even the strongest configuration (Claude+PaperCoder) achieves a mean SAU score of only 0.301 out of 1.0, with an overall mean of 0.221 across 360 evaluations. A failure taxonomy reveals that agents attempt most requirements but implement them incorrectly, with implementation mismatch and stubs accounting for the majority of zero-scored claims. Our analysis further indicates that scaffolds optimized for executability provide limited leverage for scientific reproduction; narrowing the gap requires scaffolds that prioritize semantic specification verification. The benchmark, annotations and evaluation pipeline are publicly available.

Robust Code RL via Faulty-Code-Driven Test case Synthesis and Dense Reward Shaping

Reinforcement learning from verifiable rewards (RLVR) has emerged as a pivotal technique for enhancing the code generation capabilities of Large Language Models (LLMs). However, the efficacy of RLVR in coding implementations is fundamentally limited by the comprehensiveness of test cases, because insufficient test coverage in code validation often causes false positives, further leading to reward hacking and policy degradation. To mitigate the reward bias stemming from the suboptimal quality of current automated generation methods, we propose the RobustTests framework, which introduces a faulty-code-driven test case synthesis strategy that leverages "near correct" faulty codes to guide the model in precisely capturing latent logical discrepancies and further integrates validator agents with behavioral feature clustering to facilitate the granular filtering of invalid and redundant test cases. To address false negatives caused by inherent hallucination noise in synthetic test cases, RobustTests also incorporates a stepwise dense reward function based on pass rates, bolstering training robustness through fine-grained feedback. By employing this pipeline, we construct a high-quality dataset that augmented the test cases in CodeContests, encompassing a broader spectrum of faulty code scenarios and significantly enhances diagnostic utility. Experimental results demonstrate that, by leveraging a moderately challenging subset of problems from CodeContests for training, RL fine-tuning of Qwen3-32B via RobustTests achieves an absolute 3% performance gain on the LiveCodeBench benchmark compared to baseline methods, confirming the effectiveness of the RobustTests framework in advancing the code generation proficiency of LLMs.

Knowing When to Ask for Help: Bayesian Self-Escalation in Hierarchical LLM Agents

Current LLM agent systems decide delegation before reasoning begins (a router picks a model) or after a response is complete (a verifier scores it and may retry). We study a third regime: an agent that recognises, during its own reasoning, that it is unlikely to succeed and transfers control to a stronger model. We formulate intra-generation delegation as a Bayesian optimal-stopping problem over a learned competence posterior -- an online estimate of the agent's eventual task success whose sufficient statistics are learned from labelled trajectories, not read off raw entropy. We derive the myopic escalation threshold in closed form, characterise the optimal policy via dynamic programming, and prove that the optimal policy is a time-varying threshold with no shape assumption on the raw signal. We further prove exponential separation of the oracle belief at the Chernoff-information rate of the signal, a regret bound governed by the calibration of the posterior, and a finite-sample guarantee: with n labelled calibration trajectories the deployed plug-in policy's regret decays as 1/sqrt(n). A controlled simulation study confirms each prediction of the theory, including the predicted 1/sqrt(n) rate. We additionally report a real-model validation on a Qwen2.5-Coder 1.5B->7B code cascade (MBPP, 257 tasks), confirming two of three pre-registered predictions: the escalation frontier dominates post-hoc routing at equal cost, and the cumulative competence belief's discrimination rises over generation.

MARS: Multi-Specialist LLM Relay System for Competitive Programming

Large Language Models excel at code generation, yet competitive programming exposes a persistent failure mode: existing multi-agent pipelines distribute work over generic planner, coder, and debugger roles and delegate the choice of algorithmic technique to the backbone alone. We present MARS (Multi-Agent Relay of Specialized LLMs), a prompt-only framework in which each agent is a topic specialist---dynamic programming, graphs, strings, geometry, and so on---grounded by retrieval-augmented generation over an algorithm-theory corpus. Given a problem, retrieval selects a small team of relevant specialists; a starter writes an initial C++17 solution, and each subsequent turn runs the candidate against public examples in a sandbox, lets the active specialist keep, repair, or hand off the draft, and forwards a structured packet to the next specialist. A single infrastructure-fixer pass normalizes boilerplate at the end. On the CodeContests test split with Gemma 4, MARS reaches $0.624 \pm 0.006$ pass rate at $2.3$ recorded pipeline stages per task ($+14.4$ percentage points over direct prompting), closing most of the gap to CodeSIM ($0.731$) at $3.3{\times}$ lower wall-clock cost and substantially smaller variance in per-task token spend. The source code is available on GitHub: https://github.com/fckand/mars.

Names Can Hurt: Spotting Slopsquatting Risks Caused by Package Name Hallucinations in Local Coding LLMs

When a code generating language model fabricates a Python package name, an adversary who has pre-registered that name on PyPI can convert that hallucination into a supply chain compromise. This event has been termed as 'slopsquatting'. We propose a two layer detector to counter this issue. The first layer performs a deterministic PyPI existence check. The second is a Random Forest classifier trained on ten features derived from the package name and its PyPI metadata. An import name reconciler bridges the two, resolving cases such as 'import cv2' versus 'pip install opencv-python' without a security bypass. The detector is embedded in a LangGraph state machine that retries at escalating temperatures and, on repeated failure, routes to a stronger fallback model. Across 300 curated prompts, the pipeline produces hallucination free code on 76% of runs. The primary exhausts its retry budget on 28.7%; intra model retries recover roughly a quarter of those, and cross model fallback recovers a further 16.5% of the remainder. Four findings have been observed. First, half of the flagged hallucinations are packages already registered on PyPI, as low quality lookalikes of well known projects, caught by the classifier rather than the deterministic layer (e.g., pil, faiss, tabula, haystack). Second, hallucination rate scales almost linearly with prompt adversariality, from 0 to 10% on routine coding to 40 to 73% on slopsquat baits. Third, the weaker primary refused 6 of 10 direct baits unaided, suggesting recent instruction tuning provides a baseline defense. Fourth, when primary and fallback share a model family, approximately 84% of primary failures recur on the fallback, motivating cross family pairing. A user study (n = 24) reports mean satisfaction 4.4 out of 5 and 21 of 24 stated adoption intent.

Beyond Executable Models: The Pufibara Agent Harness and the Modelica Agent Workflow Benchmark for Physical System Modeling

AI agents are increasingly used for simulation-driven engineering. Physical system modeling presents different requirements from general-purpose code generation in software engineering, because correctness depends not only on syntax and executability but also on physical consistency and scenario-dependent behavior. We study this challenge in Modelica, an equation-based modeling language in which a model may compile and simulate while still violating its intended physics or engineering requirements. Across successive revisions, an agent may lose track of requirements or rely on simulation evidence produced by an outdated candidate. To address this challenge, we present Pufibara, an agent harness that maintains persistent engineering state across revisions, associates execution and simulation evidence with the candidate that produced it, and makes submission an explicit agent action. To evaluate end-to-end Modelica agent workflows, we also propose a source-grounded method for constructing realistic and independently evaluable tasks. We use this method to build the 232-task Modelica Agent Workflow Benchmark, spanning Model Repair, Model Generation, and Model Tuning. Each submitted candidate is scored by a benchmark-owned evaluator outside the agent loop. We compare Pufibara with Claude Code as complete harnesses under two matched large language model (LLM) backends. With DeepSeek v4 Flash, Pufibara passes 202 tasks, compared with 185 for Claude Code. With Claude Sonnet 5, Pufibara passes 202 tasks, compared with 187 for Claude Code. Under the repository-reported token accounting, Pufibara records 76.4%-82.5% lower logical-token totals. Its sequential runtime is 6.1%-58.4% lower. These findings show that, even under matched LLM backends, complete agent harnesses can differ substantially in both task success and resource use for physical system modeling.

Evaluating Inference-Time Defenses Against Package Hallucination in LLM-Generated Code

LLMs are increasingly used for code generation, yet they frequently hallucinate non-existent software packages, creating exploitable entry points into the software supply chain. We make four contributions to this problem. First, we show that prior evaluation methodologies systematically inflate hallucination rates by misclassifying standard-library modules as hallucinations in some languages. For Python, the overestimation reaches 9.4 percentage points. Second, we evaluate seven inference-time defenses for mitigating package hallucinations, including five guided decoding strategies (Greedy, Contrastive, DoLa, Nudging, and Active Layer-Contrastive Decoding), an iterative self-refinement approach (Self-Refine), and a Retrieval-Augmented Generation (RAG)-based defense.. Across eight models spanning five families and four programming languages (Python, JavaScript, Ruby, Rust), RAG reduces the package hallucination rate (PHR) in 18 of 32 model--language configurations. Third, we introduce Package Utility (PU) to assess whether defenses preserve valid and task-relevant recommendations. Among strategies evaluated, Greedy decoding provides the strongest average mitigation--utility trade-off. Fourth, we stress-test all strategies under adversarial prompts seeded with fabricated package names and find that PHR surges by up to 45 percentage points relative to standard prompts, with Ruby consistently the most vulnerable language (80.9--95.2\%). Under adversarial conditions, RAG and Self-Refine outperform all decoding-only strategies, indicating that robust defense requires either external grounding or iterative self-verification when prompts are actively hostile. Our results recast package hallucination as both a measurement problem and a decoding-time control problem, and they demonstrate that the choice of defense must be matched to the threat model and recommendation utility.

Function-Level Execution Feedback for Code Preference Optimization

Process supervision has improved mathematical reasoning, where intermediate steps are naturally expressed as chains of thought. In code generation, however, process supervision remains underexplored because there is no standard notion of a step. Supervision can target lines, reasoning traces, or program states, making it unclear what to label and optimize. We propose STEP-KTODER, a framework for code preference optimization that defines steps as module-level functions in decomposed multi-function programs and assigns binary correctness labels via automatically generated unit tests. Our method provides a code-specific instantiation of stepwise KTO, combining function-level process supervision with outcome-level feedback on the full program. We evaluate on HumanEval(+), MBPP(+), BigCodeBench, and LiveCodeBench, showing that STEP-KTODER improves over outcome-only KTO and DPO. Further analysis shows that execution-based labels are essential: LLM-as-a-judge annotations systematically over-predict function failures, corrupt positive step labels, and degrade downstream preference optimization. Code is available at: https://github.com/inechnech/STEP-KTODER.

Decision-Support and Modeling with Large Language Models for Geothermal Well Arrays

Geothermal well arrays, which organize multiple geothermal wells into carefully planned geometric configurations, provide opportunities to enhance energy production capacity and increase fault tolerance. The development and adoption of these emerging geothermal technologies could be accelerated through the recent advances in large language models (LLMs) and high-level high-performance languages. A challenge in LLM-based applications is the reliability of the generated outputs, as they can be prone to subjective biases and hallucinations. This study assesses the potential of cutting-edge LLMs - such as ChatGPT, Gemini, Claude, Grok, and domain-specific models like AskGDR - as expert assistants that can synthesize insightful interpretations of complex geothermal data, as well as improve feature capabilities of geothermal models and numerical software. We developed a novel approach, leveraging Google's recently introduced AI assistant, NotebookLM, to accelerate the generation of unpublished quantitative geothermal benchmarks. The rapid generation of these evaluation instruments is essential for assessing the swiftly evolving capabilities of emerging language model technologies. In particular, we use these benchmarks and LLM-based interviews to analyze opportunities and limitations of two promising technologies: geothermal well arrays and closed-loop coaxial wells. Furthermore, we present a case study illustrating how LLMs can facilitate auto-parallelization of geothermal numerical models. Our analysis emphasizes their application in digital twins and underscores the importance of high-level, high-performance code generation. This line of research could play a transformative role in the geothermal sector by enabling the next-generation of decision-support applications, integrating data analysis, informed recommendations, and more dynamic numerical modeling workflows.