arXiv subjects

Kevin Leach

Publications and source records attributed to Kevin Leach.

At least 19 recordsLinked to original sources

arXiv2026

Rebalancing Token Importance in Language Models with TF-IDF Weighted Cross-Entropy Loss

Large language models are typically trained under uniform token weighting, which allows frequent and low-information tokens to dominate learning and can increase the tendency to memorize surface-level text spans. To address this, we present an information-weighted cross-entropy loss that rescales token-level contributions using TF-IDF statistics, emphasizing semantically informative tokens while down-weighting ubiquitous ones. Experiments on five decoder-only LLMs ranging from 1.1B to 13B parameters show consistent reductions in memorized substring length while preserving perplexity and downstream task performance. Under LoRA fine-tuning, TF-IDF reduces average substring memorization length by 14% across all five models. Under full-weight fine-tuning on TinyLLaMA 1.1B, the reduction reaches 58%. Our approach is architecture-agnostic and can be incorporated into existing training pipelines with less than 3% computational overhead, offering a lightweight and principled way to mitigate memorization without disrupting standard training dynamics.

cs.CL↗

arXiv2026

STILL: Recovering Lowered STL Semantics for LLM-assisted C++ Decompilation

LLM-assisted decompilation improves readability and re-executability, but still underperforms on stripped C++ functions that use the Standard Template Library (STL). Compilation, optimization, and symbol stripping remove or obscure source-level semantics such as container types and library-call structure, while traditional decompiler output often fails to recover them. We present STILL, a structured semantic interface that predicts function-level STL container semantics from stripped control-flow graphs and renders them as compact hints for LLM refinement. On StlBench, STILL predicts common container-level STL semantics, with the strongest cross-dataset results for stable string and vector slices. On stripped HumanEval decompilation, these hints enable DeepSeek-chat refinement to reach 28.4% executability, compared with 17.4% for no-hint refinement and 8.9% for raw Ghidra decompilation; hint utility is downstream-backbone-dependent, with decompilation-specialized models requiring lightweight adaptation to benefit from the same interface.

cs.SE↗

arXiv2026

GraphAlignCoder: Aligning Program and Proof Graphs for Code Generation

Code large language models (LLMs) can generate syntactically plausible programs that nevertheless violate hidden semantic constraints. Existing execution-feedback training methods identify whether a completed program fails, but provide limited supervision about how a correct solution should be organized. We introduce GraphAlignCoder, a training framework that transfers explicit correctness structure into code generation. GraphAlignCoder constructs an implementation graph that captures control and dependence among program regions. In parallel, a constrained Lean pipeline produces proof traces, from which we extract a formal proof-flow graph. The model first learns executable code together with graph-derived descriptions of why individual program regions are correct, and then consolidates this knowledge into code generation. GraphAlignCoder consistently outperforms the base model, code-only SFT, and CodeRL across all benchmarks. Compared with CodeRL, it increases the solved count from 38 to 50 on LiveCodeBench v6 and from 16 to 23 on BigCodeBench Hard, corresponding to relative gains of 31.6% and 43.8%, while also improving BigCodeBench Full from 359 to 363 tasks. The ablation study further shows that verification-graph injection produces the initial reasoning gain, while verification to code consolidation is essential for robust cross-benchmark transfer.

cs.SE↗

arXiv2026

Stronger Alignment between Brain Activity and LLM Embeddings during Code Writing compared to Prose Writing

Programming is a critical skill underlying modern software systems, yet the cognitive processes supporting code writing are only beginning to be understood, limiting educational practices and developer tools. At the same time, Large Language Models (LLMs) are increasingly used to assist programming. These models themselves are not well understood and can exhibit undesirable behavior like introducing security vulnerabilities. Given evidence that some cognitive representations may be shared between LLMs and the brain, we seek to improve our understanding on both fronts by relating these two systems to one another. We used Voxelwise Encoding Models (VEMs) to relate LLM embeddings to brain activity measured with functional Magnetic Resonance Imaging (fMRI) during naturalistic writing tasks. Using participants' (n = 23) keystrokes as prompts, we extracted LLM embeddings to predict voxelwise Blood Oxygen Level Dependent (BOLD) signal, quantifying alignment as the correlation between predicted and recorded signal. To assess whether this alignment is specific to programming or generalizes to other generative processes, we compared code writing to prose writing. Alignment was strongest in the right frontal pole, and brain activity was significantly better predicted by LLM embeddings during code writing than prose writing (p < 0.001, FDR-corrected). Within participants, the best-modeled voxel locations for code writing were 66% consistent across LLM layers but varied substantially between participants (39% similarity). Our findings suggest stronger alignment between human and LLM representations during structured code generation, with implications for designing AI systems that predict code generation but support natural language tasks.

cs.HC↗

arXiv2026

SCOPE: Leveraging Subgoal Critiques for Code Generation

Code generation with large language models (LLMs) remains unreliable because generated programs can appear correct while still violating key semantic requirements in the natural language specification. Existing feedback-based methods improve over coder-only generation, but they often rely on unstructured critique or execution signals that do not explicitly identify what the code is semantically missing. We present SCOPE, a prover-initialized subgoal critic for code generation. SCOPE adapts a Lean-oriented prover model to produce three parseable feedback fields for downstream code generation: subgoals, gap analysis, and a robustness checklist. Our approach combines supervised fine-tuning, process-aligned reinforcement learning (RL), and feedback-guided inference, with two complementary rewards during RL: a dense reward for structured critique quality and a sparse reward based on whether the critique improves the coder's execution score. Experiments show that SCOPE improves over the compared feedback baselines. On LiveCodeBench V6, SCOPE achieves 39.4% pass@1, compared with 36.6% for Reflexion and 20.6% for the coder-only baseline. On BigCodeBench (Hard), it reaches 42.6%, surpassing Reflexion at 36.5% and coder-only generation at 34.5%. Further analysis shows that SCOPE's gains are concentrated in tasks with concrete semantic constraints and that its code corrections are more localized than Reflexion's.

cs.SE↗

arXiv2026

Revising RVL-CDIP: Quantifying Errors and Test-Train Overlap

RVL-CDIP is a popular dataset for benchmarking document classifiers. However, the dataset contains ample amounts of label errors as well as non-trivial amounts of test-train overlap, both of which may impact model performance metrics. In this paper, we address these two problems by (1) finding and fixing label errors, and (2) detecting and addressing test-train overlap. We produce several variations of RVL-CDIP with label error and test-train overlap fixes, and benchmark document classification performance on these new RVL-CDIP variations. Our rigorous analysis of RVL-CDIP finds that the corpus contains 12\% label error and approximately 35% test-train duplication. Remediation sees improvements in classification accuracy when errors are removed, but sees decreases in accuracy when duplicates are removed. We additionally evaluate models on RVL-CDIP-N, an out-of-distribution benchmark, finding that training on error-corrected data substantially improves OOD generalization, with supervised models gaining an average of 8.1 percentage points in accuracy and improvements as large as 14 percentage points.

cs.CL↗

arXiv2026

VERITAS: Verifier-Guided Proof Search for Zero-Shot Formal Theorem Proving

LLM-based formal provers often collapse rich verifier signals (syntax errors, type mismatches, partial goal progress) into a binary pass/fail bit. We present VERITAS, a zero-shot framework that routes every verifier signal back into proof search through a two-phase protocol: Best-of-N sampling first, then a critic-guided MCTS pass that ingests Phase 1 failures as explicit negative examples. The protocol preserves every theorem solved by its own Phase 1 sweep, so Phase 2's additional solves are attributable to feedback-driven exploration. VERITAS reaches 40.6% on miniF2F (vs. an independently run Best-of-5 at 36.9%, Portfolio 26.2%) and 7.3% on VERITAS-CombiBench, a 55-theorem combinatorics benchmark we release on which Best-of-5 (1.8%) falls below Portfolio (3.6%), exposing that unguided sampling hurts when correct lemma names must be recovered iteratively from verifier feedback. Artifacts are available on GitHub.

cs.LG↗

arXiv2026

Constraint-Guided Multi-Agent Decompilation for Executable Binary Recovery

Decompilation -- recovering source code from compiled binaries -- is essential for security analysis, malware reverse engineering, and legacy software maintenance. However, existing decompilers produce code that often fails to compile or execute correctly, limiting their practical utility. We present a multi-agent framework that transforms decompiled code into re-executable source through Multi-level Constraint-Guided Decompilation (MCGD). Our approach employs a hierarchical validation pipeline with three constraint levels: (1) syntactic correctness via parsing, (2) compilability via GCC, and (3) behavioral equivalence via LLM-generated test cases. When validation fails, specialized LLM agents iteratively refine the code using structured error feedback. We evaluate our framework on 1,641 real-world binaries from ExeBench across three decompilers (RetDec, Ghidra, and Angr). Our framework achieves 84-97% re-executability, improving baseline decompiler output by 28-89 percentage points. In comparison with state-of-the-art LLM-based decompilation methods using the same GPT-4o backbone, our approach (84.1%) outperforms LLM4Decompile (80.3%), SK2Decompile (73.9%), and SALT4Decompile (61.8%). Our ablation study reveals that execution-based validation is critical: compile-only approaches achieve 0% behavioral correctness despite 91-99% compilation rates. The system converges efficiently, with 90%+ binaries reaching correctness within 2 iterations at an average cost of $0.03-0.05 per binary. Our results demonstrate that constraint-guided agentic refinement can bridge the gap between raw decompiler output and practically useful source code.

cs.SE↗

arXiv2026

SynthFix: Adaptive Neuro-Symbolic Code Vulnerability Repair

Large Language Models (LLMs) can generate plausible code patches, but plausibility is not enough for automated repair: a patch must compile, pass tests, and remove the target vulnerability. We present SynthFix, a neuro-symbolic repair framework that combines supervised repair learning with compiler-informed feedback. During training, a lightweight router selects between Supervised Fine-Tuning (SFT) for common repair patterns and Reward Fine-Tuning (RFT) for examples that benefit from symbolic feedback. The reward combines static structure, lint/compile checks, security scanning, and public execution tests where available; at inference time, the same evidence guides best-of-K candidate selection under a greedy floor. Across five code LLMs (1.3B-7B) on pyrepair, CodeFlaws, and SVEN, SynthFix improves deployable repair metrics over SFT-only and RFT-only baselines, with relative gains up to 54 percent in functional correctness and 14 percent in security clearance. Our code and data are available at https://github.com/CoderDoge1108/SynthFix.

cs.SE↗

arXiv2026

EyeLayer: Integrating Human Attention Patterns into LLM-Based Code Summarization

Code summarization is the task of generating natural language descriptions of source code, which is critical for software comprehension and maintenance. While large language models (LLMs) have achieved remarkable progress on this task, an open question remains: can human expertise in code understanding further guide and enhance these models? We propose EyeLayer, a lightweight attention-augmentation module that incorporates human eye-gaze patterns, as a proxy of human expertise, into LLM-based code summarization. EyeLayer models human attention during code reading via a Multimodal Gaussian Mixture, redistributing token embeddings based on learned parameters (\mu_i, \sigma_i^2) that capture where and how intensively developers focus. This design enables learning generalizable attention priors from eye-tracking data and incorporating them into LLMs seamlessly, without disturbing existing representations. We evaluate EyeLayer across diverse model families (i.e., LLaMA-3.2, Qwen3, and CodeBERT) covering different scales and architectures. EyeLayer consistently outperforms strong fine-tuning baselines across standard metrics, achieving gains of up to 13.17% on BLEU-4. These results demonstrate that human gaze patterns encode complementary attention signals that enhance the semantic focus of LLMs and transfer effectively across diverse models for code summarization.

cs.SE↗

arXiv2026

Do Hackers Dream of Electric Teachers?: A Large-Scale, In-Situ Evaluation of Cybersecurity Student Behaviors and Performance with AI Tutors

To meet the ever-increasing demands of the cybersecurity workforce, AI tutors have been proposed for personalized, scalable education. But, while AI tutors have shown promise in introductory programming courses, no work has evaluated their use in hands-on exploration and exploitation of systems (e.g., ``capture-the-flag'') commonly used to teach cybersecurity. Thus, despite growing interest and need, no work has evaluated how students use AI tutors or whether they benefit from their presence in real, large-scale cybersecurity courses. To answer this, we conducted a semester-long observational study on the use of an embedded AI tutor with 309 students in an upper-division introductory cybersecurity course. By analyzing 142,526 student queries sent to the AI tutor across 396 cybersecurity challenges spanning 9 core cybersecurity topics and an accompanying set of post-semester surveys, we find (1) what queries and conversational strategies students use with AI tutors, (2) how these strategies correlate with challenge completion, and (3) students' perceptions of AI tutors in cybersecurity education. In particular, we identify three broad AI tutor conversational styles among users: Short (bounded, few-turn exchanges), Reactive (repeatedly submitting code and errors), and Proactive (driving problem-solving through targeted inquiry). We also find that the use of these styles significantly predicts challenge completion, and that this effect increases as materials become more advanced. Furthermore, students valued the tutor's availability but reported that it became less useful for harder material. Based on this, we provide suggestions for security educators and developers on practical AI tutor use.

cs.HC↗

arXiv2025

SUGAR: A Sweeter Spot for Generative Unlearning of Many Identities

Recent advances in 3D-aware generative models have enabled high-fidelity image synthesis of human identities. However, this progress raises urgent questions around user consent and the ability to remove specific individuals from a model's output space. We address this by introducing SUGAR, a framework for scalable generative unlearning that enables the removal of many identities (simultaneously or sequentially) without retraining the entire model. Rather than projecting unwanted identities to unrealistic outputs or relying on static template faces, SUGAR learns a personalized surrogate latent for each identity, diverting reconstructions to visually coherent alternatives while preserving the model's quality and diversity. We further introduce a continual utility preservation objective that guards against degradation as more identities are forgotten. SUGAR achieves state-of-the-art performance in removing up to 200 identities, while delivering up to a 700% improvement in retention utility compared to existing baselines. Our code is publicly available at https://github.com/judydnguyen/SUGAR-Generative-Unlearn.

cs.CV↗

arXiv2025

Context-Guided Decompilation: A Step Towards Re-executability

Binary decompilation plays an important role in software security analysis, reverse engineering, and malware understanding when source code is unavailable. However, existing decompilation techniques often fail to produce source code that can be successfully recompiled and re-executed, particularly for optimized binaries. Recent advances in large language models (LLMs) have enabled neural approaches to decompilation, but the generated code is typically only semantically plausible rather than truly executable, limiting their practical reliability. These shortcomings arise from compiler optimizations and the loss of semantic cues in compiled code, which LLMs struggle to recover without contextual guidance. To address this challenge, we propose ICL4Decomp, a hybrid decompilation framework that leverages in-context learning (ICL) to guide LLMs toward generating re-executable source code. We evaluate our method across multiple datasets, optimization levels, and compilers, demonstrating around 40\% improvement in re-executability over state-of-the-art decompilation methods while maintaining robustness.

cs.SE↗

arXiv2025

DPO-F+: Aligning Code Repair Feedback with Developers' Preferences

Large Language Models (LLMs) are increasingly used in software engineering tasks, especially code repair. However, developers often struggle to interpret model outputs, limiting effective human--AI teaming, where humans and AI work toward a shared objective. Prior work mainly optimizes generated code, giving less attention to natural-language feedback that supports comprehension and iterative improvement. We present \textsc{DPO-f+}, a framework that aligns code-repair feedback with the needs of different developer groups, including novices and proficient developers. It (1) defines feedback-alignment metrics across seven fixed dimensions with task-specific descriptions; (2) automatically constructs pairwise preference datasets from code-repair tasks; (3) fine-tunes models using Direct Preference Optimization (DPO) augmented with a reward model; and (4) provides an automated protocol for evaluating feedback quality. Empirically, \textsc{DPO-f+} outperforms both the baseline and standard DPO in feedback accuracy and overall alignment. On novice programming tasks, \textsc{DPO-f+} improves Pass@1 by 5.71 percentage points (pp) over the baseline and 3.30 pp over DPO. On \textsc{SWE-Bench}, it improves issue-resolution rate by 1.67 pp over DPO and 4.67 pp over the baseline. It also improves feedback alignment by both LLM judges and a human study with 200 developers: beginners preferred \textsc{DPO-f+} in 71.5% of comparisons, with overall preference above chance ($p=0.0057$). By better aligning feedback with developer needs, \textsc{DPO-f+} turns LLM assistance from a one-shot output into a collaborative sense-making workflow, enhancing code comprehension and human--AI teaming in software engineering.

cs.SE↗

arXiv2025

EyeMulator: Improving Code Language Models by Mimicking Human Visual Attention

Code Language Models (CodeLLMs) learn token importance from data correlations, whereas human developers attend selectively to semantically salient code. We present EyeMulator, a model-agnostic method that injects human visual-attention priors into CodeLLM fine-tuning without architectural changes. EyeMulator distills eye-tracking data into semantic salience and gaze-transition priors, then uses them to reweight token-level training losses. Across six backbones, two data regimes, and three CodeXGLUE tasks, the reported configurations yield positive matched-metric deltas in all 36 model-task-setting cells. Effects are largest for structure-preserving completion and translation, while summarization shows smaller but positive METEOR deltas. Session-mode and component-ablation analyses further show that reading, writing, semantic, and transition-derived priors provide complementary signal. Human-attention artifacts are available at https://zenodo.org/records/17205682.

cs.SE↗

arXiv2025

CodeGrad: Integrating Multi-Step Verification with Gradient-Based LLM Refinement

While Large Language Models (LLMs) have demonstrated remarkable capabilities in code generation, they often produce solutions that lack guarantees of correctness, robustness, and efficiency. This limitation is particularly acute in domains requiring strict constraints. CodeGrad introduces a principled framework that integrates rigorous verification techniques directly into an iterative LLM-based generation loop. It uniquely treats code as a differentiable variable, converting structured feedback and mathematical constraints into a textual pseudo-gradient. This gradient guides the model to iteratively refine solutions, ensuring they are not only functional but also robust and mathematically justified. We evaluate CodeGrad on the HumanEval, HumanEval+, and LiveCodeBench benchmarks. Our implementation outperforms strong baselines, achieving an absolute improvement of up to 27% on HumanEval and a 41% relative improvement on the challenging LiveCodeBench V6. StructuredGrad generates mathematically justified code that is robust and efficient, paving the way for reliable AI-assisted software development in high-stakes applications.

cs.SE↗

arXiv2025

A Human Study of Cognitive Biases in Web Application Security

Cybersecurity training has become a crucial part of computer science education and industrial onboarding. Capture the Flag (CTF) competitions have emerged as a valuable, gamified approach for developing and refining the skills of cybersecurity and software engineering professionals. However, while CTFs provide a controlled environment for tackling real world challenges, the participants' decision making and problem solving processes remain under explored. Recognizing that psychology may play a role in a cyber attacker's behavior, we investigate how cognitive biases could be used to improve CTF education and security. In this paper, we present an approach to control cognitive biases, specifically Satisfaction of Search and Loss Aversion, to influence and potentially hinder attackers' effectiveness against web application vulnerabilities in a CTF style challenge. We employ a rigorous quantitative and qualitative analysis through a controlled human study of CTF tasks. CTF exercises are widely used in cybersecurity education and research to simulate real world attack scenarios and help participants develop critical skills by solving security challenges in controlled environments. In our study, participants interact with a web application containing deliberately embedded vulnerabilities while being subjected to tasks designed to trigger cognitive biases. Our study reveals that many participants exhibit the Satisfaction of Search bias and that this bias has a significant effect on their success. On average, participants found 25% fewer flags compared to those who did not exhibit this bias. Our findings provide valuable insights into how cognitive biases can be strategically employed to enhance cybersecurity outcomes, education, and measurements through the lens of CTF challenges.

cs.CR↗

arXiv2025

Enhancing Code LLM Training with Programmer Attention

Human attention provides valuable yet underexploited signals for code LLM training, offering a perspective beyond purely machine-driven attention. Despite the complexity and cost of collecting eye-tracking data, there has also been limited progress in systematically using these signals for code LLM training. To address both issues, we propose a cohesive pipeline spanning augmentation and reward-based fine-tuning. Specifically, we introduce (1) an eye-tracking path augmentation method to expand programmer attention datasets, (2) a pattern abstraction step that refines raw fixations into learnable attention motifs, and (3) a reward-guided strategy for integrating these insights directly into a CodeT5 supervised fine-tuning process. Our experiments yield +7.16 in CodeBLEU on the CodeXGlue benchmark for code summarization, underscoring how uniting human and machine attention can boost code intelligence. We hope this work encourages broader exploration of human-centric methods in next-generation AI4SE.

cs.SE↗