Zibin Zheng

CodeCritic: towards effective code quality evaluation with large language models

SWE-Prime: Fewer Trajectories, Better Performance

To improve large language models'ability to resolve real-world software issues, prior work has focused on constructing large-scale agent trajectory datasets and performing supervised fine-tuning (SFT) on successful trajectories. However, task success does not guarantee high-quality supervision: successful trajectories...

Efficient SWE Agent Benchmarking via Trajectory-Aware Evaluation

PTA-IRT is proposed, a Privileged Trajectory-Aware Item Response Theory framework that fuses process and outcome signals and consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks.

Adaptive Critical Token-Aware Retrieval for Repository-Level Code Generation

ACToR identifies critical tokens during generation and triggers targeted retrieval on demand to provide repository context at these decisive positions, and designs a position-aware weighting method for dense retrievers to prioritize context that is more informative for generation.

From Static to Dynamic: Benchmarking Real-World Code Review with MCR-Bench

MCR-Bench is introduced, the first defect state-aware benchmark designed for realistic multi-round code review, and in-depth error analysis dissects the distinct drivers of false positives and false negatives, revealing critical weaknesses such as cross-round temporal misalignment and inadequate long-range memory.

RepoReasoner: Evaluating Repository-Level Code Reasoning Ability of Long-Context Language Models

RepoReasoner is introduced, a benchmark for evaluating repository-level code reasoning that assesses two complementary abilities: Output Prediction, which measures fine-grained, stateful execution reasoning across files, and Call Chain Prediction, which evaluates high-level architectural dependency understanding under...