Master's student · Central South University
Code agents. Richer environments.
Better learning.
I study how to make code agents more capable through better tasks, richer environments, and post-training.
I'm a master's student at Central South University and an applied algorithm intern at ByteDance. I received my bachelor's degree from Hunan Normal University in 2025.
Agents, code, and learning from interaction.
9 research works shown.
A six-stage multi-agent pipeline turns sparse CVE metadata into executable tasks with 95% correctness. Fine-tuning Qwen3-32B on sampled trajectories improves LiveCVEBench from 5.3% → 35.8%, surpassing Claude 4.5 Sonnet, and Terminal-Bench from 12.5% → 28.75%.
WrongSelect models test outcomes as a binary matrix and uses matrix rank to remove redundant error patterns. The resulting compact evaluation sets reduce computation and avoid inflated test-case quality scores.
World Rehearsal lets a single policy act, predict environment responses, and learn from its own rehearsals without an external environment. EnvACE improves performance on BFCL-v4, τ²-Bench, and VitaBench.
A benchmark for debugging errors that arise from interactions between multiple libraries, extending code debugging beyond isolated functions and static bugs.
Measuring semantic and security anomalies in HTTP/3-to-HTTP/1.1 translation through automated, agent-assisted testing.
A large-scale benchmark and multi-stage framework for removing facial shadows while preserving fine details and visual fidelity.
Open-vocabulary cytopathology detection with a structured prompt bank, separating visual feature extraction from class-aware prompt matching.
Additional work under review
Observation-preserving counterfactual environments train agents to inspect before acting, with inspect-then-act SFT trajectories and RL/GRPO data. On Terminal-Bench 2.1, SFT improves from 22.2% → 30.8% and Direct RL (GRPO) from 22.2% → 39.4%.
Benchmarking the self-correction capabilities of multimodal models.
— Present
China Commerce & Advertising · Applied Algorithms
Built lightweight routing models for advertising review workflows using Qwen-0.8B's last hidden states. Cleaned distribution-shifted data and expanded training labels aligned with Pass@k.
Recall 52.91% → 77.02% at 90% precision; deployed across 10 tags with measurable impact.
—
Applied Algorithms
—
Data Intelligence Applications · Research
— Present
Master's student
—
Bachelor's degree · Project 211 university
Let's connect