Lukas Twist

Lukas Twist, Helen Yannakoudakis, Jie M. Zhang

Explicit reasoning models are trained to produce intermediate reasoning traces before final answers, but downstream fine-tuning is often performed on ordinary instruction-response data that contains no such traces. We show that this mismatch can induce reasoning-trace collapse: a fine-tuned model continues to produce plausible final answers while losing the structurally valid explicit reasoning traces that made it a reasoning model in the first place. We introduce a structural evaluation framework that separates answer correctness from reasoning-trace validity, measuring valid, empty, missing, and truncated reasoning alongside reasoning-conditioned task performance. Using this framework, we study four open-weight reasoning models and find that standard supervised fine-tuning can rapidly suppress valid reasoning traces, and that answer-only metrics can substantially obscure this failure: in several settings, performance conditional on valid reasoning remains high while the rate of valid reasoning falls sharply. We further show that simple loss-masking strategies can substantially mitigate collapse without requiring teacher-generated reasoning traces. These results suggest that evaluations of fine-tuned reasoning models should report structural reasoning reliability metrics in addition to final-answer performance, especially when adaptation data does not contain explicit reasoning traces.

Lukas Twist, Shu Yang, Hanqi Yan et al.

Large Language Models (LLMs) increasingly exhibit strong reasoning abilities, often attributed to their capacity to generate chain-of-thought-style intermediate reasoning. Recent work suggests that exposure to code can further enhance these skills, but existing studies largely treat code as a generic training signal, leaving open the question of which properties of code actually contribute to improved reasoning. To address this gap, we study the structural complexity of code, which captures control flow and compositional structure that may shape how models internalise multi-step reasoning during fine-tuning. We examine two complementary settings: solution-driven complexity, where complexity varies across multiple solutions to the same problem, and problem-driven complexity, where complexity reflects variation in the underlying tasks. Using cyclomatic complexity and logical lines of code to construct controlled fine-tuning datasets, we evaluate a range of open-weight LLMs on diverse reasoning benchmarks. Our findings show that although code can improve reasoning, structural properties strongly determine its usefulness. In 83% of experiments, restricting fine-tuning data to a specific structural complexity range outperforms training on structurally diverse code, pointing to a data-centric path for improving reasoning beyond scaling.

Jingzhi Gong, Ruizhen Gu, Zhiwei Fei et al.

Agent skills provide modular, task-specific guidance for LLM- based coding agents, but manually tuning skill bundles to balance success rate, cost, and runtime is expensive and fragile. We present SkillMOO, a multi-objective optimization framework that automatically evolves skill bundles using LLM-proposed edits and NSGA-II survivor selection: a solver agent evaluates candidate skill bundles on coding tasks and an optimizer agent proposes bundle edits based on failure analysis. On three SkillsBench software engineering tasks, SkillMOO improves pass rate by up to 131% while reducing cost up to 32% relative to the best baseline per task at low optimization overhead. Pattern analysis reveals pruning and substitution as primary drivers of improvement, suggesting effective bundles favor minimal, focused content over accumulated instructions.

Lukas Twist, Jie M. Zhang, Mark Harman et al.

Large Language Models (LLMs) are increasingly used to generate code, influencing users' choices of libraries and programming languages in critical real-world projects. However, little is known about their systematic biases or preferences toward certain libraries and programming languages, which can significantly impact software development practices. To fill this gap, we perform the first empirical study of LLMs' preferences for libraries and programming languages when generating code, covering eight diverse LLMs. Our results reveal that LLMs exhibit a strong tendency to overuse widely adopted libraries such as NumPy; in up to 48% of cases, this usage is unnecessary and deviates from the ground-truth solutions. LLMs also exhibit a significant preference toward Python as their default language. For high-performance project initialisation tasks where Python is not the optimal language, it remains the dominant choice in 58% of cases, and Rust is not used a single time. These results indicate that LLMs may prioritise familiarity and popularity over suitability and task-specific optimality. This will introduce security vulnerabilities and technical debt, and limit exposure to newly developed, better-suited tools and languages. Understanding and addressing these biases is essential for the responsible integration of LLMs into software development workflows.

Lukas Twist, Jie M. Zhang, Mark Harman et al.

Large language models (LLMs) are increasingly used to generate code, yet they continue to hallucinate, often inventing non-existent libraries. Such library hallucinations are not just benign errors: they can mislead developers, break builds, and expose systems to supply chain threats such as slopsquatting. Despite increasing awareness of these risks, little is known about how real-world prompt variations affect hallucination rates. Therefore, we present the first systematic study of how user-level prompt variations impact library hallucinations in LLM-generated code. We evaluate six diverse LLMs across two hallucination types: library name hallucinations (invalid imports) and library member hallucinations (invalid calls from valid libraries). We investigate how realistic user language extracted from developer forums and how user errors of varying degrees (one- or multi-character misspellings and completely fake names/members) affect LLM hallucination rates. Our findings reveal systemic vulnerabilities: one-character misspellings in library names trigger hallucinations in up to 26% of tasks, fake library names are accepted in up to 99% of tasks, and time-related prompts lead to hallucinations in up to 84% of tasks. Prompt engineering shows promise for mitigating hallucinations, but remains inconsistent and LLM-dependent. Our results underscore the fragility of LLMs to natural prompt variation and highlight the urgent need for safeguards against library-related hallucinations and their potential exploitation.