arXiv subjects

Anji Li

Publications and source records attributed to Anji Li.

7 recordsLinked to original sources

arXiv2026

Unseen-Codebases-Domain Data Synthesis and Training Based on Code Graphs

In the context of newly release software frameworks, large language models (LLMs) often exhibit poor performance and a high rate of hallucination, as they are not exposed to such environments during training. Although inference-time augmentation techniques such as retrieval-augmented generation (RAG) can partially mitigate hallucinations, knowledge injection through prompting alone is insufficient to enable models to fully understand the intrinsic relationships among different components of a codebase, or to reason about the correct compositions and apply. Although explicit knowledge injection can be achieved through post-training, compared with public code domains, unseen codebases typically provide only source code and lack large volumes of high-quality, usage-oriented code that can be directly leveraged as training data. Consequently, existing data synthesis approaches are insufficient to adequately capture unseen codebases usage scenarios when restricted to source code alone. To address these challenges, we propose UCD-Training, a two-stage training framework for reasoning-aware data synthesis grounded in a code graph constructed from unseen codebases. UCD-Training first parses the source code to build a code graph, then conducts dependency-preserving continued pretraining (CPT) using file-level dependency data, followed by graph-grounded supervised fine-tuning (SFT) on three types of synthesized data augmented with explicit reasoning traces: (1) single-hop relation reasoning data, (2) compositional API reasoning data, and (3) codebase utilization data. We further introduce a new benchmark, UnseenCodeBench, for code generation on unseen codebases and conduct comprehensive experiments across multiple codebases.

cs.SE↗

arXiv2025

Knowledge Matters: Injecting Project and Testing Knowledge into LLM-based Unit Test Generation

Automated unit test generation using large language models (LLMs) holds great promise but often struggles with generating tests that are both correct and maintainable in real-world projects. This paper presents KTester, a novel framework that integrates project-specific knowledge and testing domain knowledge to enhance LLM-based test generation. Our approach first extracts project structure and usage knowledge through static analysis, which provides rich context for the model. It then employs a testing-domain-knowledge-guided separation of test case design and test method generation, combined with a multi-perspective prompting strategy that guides the LLM to consider diverse testing heuristics. The generated tests follow structured templates, improving clarity and maintainability. We evaluate KTester on multiple open-source projects, comparing it against state-of-the-art LLM-based baselines using automatic correctness and coverage metrics, as well as a human study assessing readability and maintainability. Results demonstrate that KTester significantly outperforms existing methods across six key metrics, improving execution pass rate by 5.69% and line coverage by 8.83% over the strongest baseline, while requiring less time and generating fewer test cases. Human evaluators also rate the tests produced by KTester significantly higher in terms of correctness, readability, and maintainability, confirming the practical advantages of our knowledge-driven framework.

cs.SE↗

arXiv2025

EvolMathEval: Towards Evolvable Benchmarks for Mathematical Reasoning via Evolutionary Testing

The rapid advancement of Large Language Models (LLMs) poses a significant challenge to existing mathematical reasoning benchmarks. However, these benchmarks tend to become easier over time as LLMs can learn from the published benchmarks. This limitation hinder the precise evaluation of the true capabilities of SOTA models. To address this challenge, this paper introduces EvolMathEval, an automated mathematical benchmark generation and evolution framework based on evolutionary testing. Experimental results demonstrate that EvolMathEval can not only generate a large volume of high-difficulty problems through continuous self-iteration, but it can also significantly enhance the complexity of public datasets like GSM8K through evolution, reducing model accuracy by an average of 48\%. Deeper investigation reveals that when solving these evolved problems, LLMs tend to bypass complex multi-step logical reasoning by relying on simplistic and fuzzy conditions, consequently leading to incorrect solutions. We define this phenomenon as the ``Pseudo Aha Moment", which we find accounts for 77\% to 100\% of errors on targeted problems. Code and resources are available at: https://anonymous.4open.science/r/EvolMathEval

cs.AI↗

arXiv2025

FeedbackEval: A Benchmark for Evaluating Large Language Models in Feedback-Driven Code Repair Tasks

Code repair is a fundamental task in software development, facilitating efficient bug resolution and software maintenance. Although large language models (LLMs) have demonstrated considerable potential in automated code repair, their ability to comprehend and leverage diverse types of feedback, which is crucial for iterative self-correction in authentic debugging scenarios, remains insufficiently understood. To bridge this gap, we introduce FeedbackEval, a systematic benchmark constructed from three heterogeneous sources (HumanEval, CoderEval, and SWE-Bench-verified), to evaluate LLMs' feedback comprehension and code repair performance. We conduct a comprehensive empirical study on five state-of-the-art LLMs, including GPT-4o, Claude-3.5, Deepseek-R1, GLM-4, and Qwen2.5, to evaluate their behavior under both single-iteration and iterative code repair settings. Our results show that mixed feedback yields the highest repair success (63.6%), with LLM-Expert and test feedback providing strong targeted gains (62.9% and 57.9%, respectively), while minimal (53.1%) and compiler feedback (49.2%) offer moderate benefits and LLM-Skilled proves least effective (48.8%). Iterative feedback further enhances repair performance, though the marginal benefit diminishes after two or three iterations. Moreover, prompt structure is shown to be critical: structured reasoning (RR, CoT) and dynamic example selection deliver notable improvements, whereas removing semantic cues such as docstrings or role-play causes severe degradation. This work introduces a robust benchmark and delivers practical insights to advance the understanding and development of feedback-driven code repair using LLMs.

cs.SE↗

arXiv2025

A Preliminary Study on the Robustness of Code Generation by Large Language Models

Robustness is a critical factor for reliable code generation by large language models, yet most evaluations focus on correctness and overlook key issues such as missing input validation and inadequate error handling. In this work, we present the first empirical study of LLM-generated code robustness using the CoderEval benchmark. Evaluating four state-of-the-art code LLMs, we find that 35.2% of their outputs are less robust than human-written code, with over 90% of deficiencies caused by missing conditional checks-70% of which occur in the first line. Interestingly, in 63% of cases where a conditional statement is needed but absent, the "if" token still ranks among the top three predictions, suggesting implicit recognition of control flow. To address these issues, we propose RobGen, a model-agnostic framework that improves robustness without retraining. RobGen combines a line-level intervention checker, which decides whether to adjust logits for each generated line, with token-level conditional logit adjustments to promote essential control structures. Experiments show that RobGen reduces the proportion of less robust code by 10%, achieves the highest average Pass@1 (43.57), and adds minimal overhead (+33.4%). As a lightweight and adaptable solution, RobGen effectively enhances the reliability of LLM-generated code across diverse tasks.

cs.SE↗

arXiv2024

Are Decoder-Only Large Language Models the Silver Bullet for Code Search?

Code search is essential for code reuse, allowing developers to efficiently locate relevant code snippets. The advent of powerful decoder-only Large Language Models (LLMs) has revolutionized many code intelligence tasks. However, their effectiveness for the retrieval-based task of code search, particularly compared to established encoder-based models, remains underexplored. This paper addresses this gap by presenting a large-scale systematic evaluation of eleven decoder-only LLMs, analyzing their performance across zero-shot and fine-tuned settings. Our results show that fine-tuned decoder-only models, particularly CodeGemma, significantly outperform encoder-only models like UniXcoder, achieving a 40.4% higher Mean Average Precision (MAP) on the CoSQA$^+$ benchmark. Our analysis further reveals two crucial nuances for practitioners: first, the relationship between model size and performance is non-monotonic, with mid-sized models often outperforming larger variants; second, the composition of the training data is critical, as a multilingual dataset enhances generalization while a small amount of data from a specific language can act as noise and interfere with model effectiveness. These findings offer a comprehensive guide to selecting and optimizing modern LLMs for code search.

cs.SE↗

arXiv2024

iJTyper: An Iterative Type Inference Framework for Java by Integrating Constraint- and Statistically-based Methods

Inferring the types of API elements in incomplete code snippets (e.g., those on Q&A forums) is a prepositive step required to work with the code snippets. Existing type inference methods can be mainly categorized as constraint-based or statistically-based. The former imposes higher requirements on code syntax and often suffers from low recall due to the syntactic limitation of code snippets. The latter relies on the statistical regularities learned from a training corpus and does not take full advantage of the type constraints in code snippets, which may lead to low precision. In this paper, we propose an iterative type inference framework for Java, called iJTyper, by integrating the strengths of both constraint- and statistically-based methods. For a code snippet, iJTyper first applies a constraint-based method and augments the code context with the inferred types of API elements. iJTyper then applies a statistically-based method to the augmented code snippet. The predicted candidate types of API elements are further used to improve the constraint-based method by reducing its pre-built knowledge base. iJTyper iteratively executes both methods and performs code context augmentation and knowledge base reduction until a termination condition is satisfied. Finally, the final inference results are obtained by combining the results of both methods. We evaluated iJTyper on two open-source datasets. Results show that 1) iJTyper achieves high average precision/recall of 97.31% and 92.52% on both datasets; 2) iJTyper significantly improves the recall of two state-of-the-art baselines, SnR and MLMTyper, by at least 7.31% and 27.44%, respectively; and 3) iJTyper improves the average precision/recall of the popular language model, ChatGPT, by 3.25% and 0.51% on both datasets.

cs.SE↗