Zibin Zheng

SALA achieves superior reasoning alignment by learning flexible, task-specific operations, outperforming traditional methods that rely on rigid matching criteria.

CoMerge not only preserves task-specific capabilities but also achieves near-optimal performance with just 1,445 scalar coefficients, challenging the need for full-parameter fine-tuning.

Critical tokens can make or break code generation, and ACToR ensures they receive the precise context needed to succeed.

PTA-IRT reveals that incorporating detailed problem-solving trajectories can dramatically enhance the evaluation accuracy of software engineering agents.

Training on just 10% of carefully curated trajectories can lead to performance improvements of over 24% in software issue resolution tasks.

Mainstream LLMs miss 40% of defects in multi-round code reviews, revealing their limitations in real-world software development contexts.

AutoSQL outperforms traditional methods by over 20% in SQL template extraction, revealing hidden efficiencies in ORM codebases.

Even the top-performing LLM struggles with cross-file reasoning, achieving only 69.1% accuracy on a new benchmark designed to reflect real-world software development challenges.

Log-based anomaly detection models are missing 90% of the picture, but AnomalyGen uses LLMs and static analysis to hallucinate realistic training data and close the gap.

LLMs can now add features to entire codebases with 36% higher success by generating multiple design options and validating their architectural impact.