Xiaodong Gu

ACQUIRE turns knowledge gaps into actionable insights, boosting software repair accuracy by over 4% while keeping costs low.

Dockerless achieves a 14.3 AUC point improvement in program verification without the overhead of Docker environments, revolutionizing efficiency in training coding agents.

UNIATTACK can breach multi-layered defenses with unprecedented efficiency, achieving up to 248% higher attack success rates than existing methods.

FastContext cuts coding agent token usage by 60% while boosting resolution rates by 5.5% by decoupling code exploration from task-solving.

Integrating visual graphs with text interfaces allows LLM agents to reduce token consumption by 26% while enhancing their issue-resolution accuracy.

Code generation models can achieve a significant boost in reliability by leveraging a tailored uncertainty estimation framework that outperforms traditional natural language methods.

Coding agents can achieve superior repository exploration, outperforming classical methods by effectively leveraging line-level context for bug diagnosis and code retrieval.

LLM agents struggle to consistently reflect human-like psychology, even when provided with extensive personality profiles and autobiographical memories, suggesting current models lack a deeper understanding of human behavior.