I'm a master’s student at the Institute of Automation, Chinese Academy of Sciences, advised by Prof. Shizhu He. I received my B.E. from UESTC.
My research focuses on language models and AI agents, with interests in capability evaluation, mechanistic interpretability, and reinforcement learning.
As frontier agents become more capable, I believe evaluation is struggling to keep pace. I develop tasks and evaluations to understand what agents can do, where they fall short, and how they can become better AI researchers and scientists, including through RSI and interpretability research.
I am equally interested in agents as user collaborators. I have worked on large-scale optimization of foundation models to better support users. Even as agents become more autonomous, I believe learning from user feedback and adapting to human involvement remain essential.
The framework below brings these directions together: understanding and improving models, evaluating their capabilities, and learning through interaction.
Two papers accepted to EMNLP 2026 (1 Main, 1 Findings).
One paper accepted to ACL 2026 (Main). DEPT ↗
Co-authorInvited talk at NICE on internal policies of LLMs and reinforcement learning. Video ↗
One paper accepted to NeurIPS 2025 Efficient Reasoning Workshop. Zero-Step Thinking ↗
First authorOne paper accepted to ACL 2025 (Main). Neural Incompatibility ↗
First authorOne paper accepted to ICMR 2025. RobustPT ↗
Co-first authorOne paper accepted to ICMR 2024. G-SAP ↗
Co-first author* equal contribution
SAEScientist-Bench evaluates whether AI agents can recover semantically specified SAE features using restricted activation probes and validates them through held-out activation and causal steering.
PostTrainBench⁰ evaluates whether LLM agents can improve frozen models through gradient-free, score-only search under fixed compute and time budgets.
SWE-Touch injects validated, task-conflicting user edits into live coding-agent trajectories to evaluate state awareness and adaptation in shared workspaces.
DSRL warms up models with negative-sample reinforcement on unconditional P(y), then reincarnates the policy into conditional P(y|x) reinforcement learning.
BuPO first optimizes a lower-layer internal policy to rebuild foundational reasoning, then switches to full-model reinforcement learning for refinement.
Mode selection is a harder early-exit problem because it must choose a reasoning mode before observing any question-specific reasoning trajectory.
Cross-scale parametric knowledge transfer requires alignment, yet behavioral and structural differences create neural incompatibility that destabilizes both PostPKT and PrePKT.
DyPRAG trains a lightweight translator to convert retrieved documents into dynamic LoRA parameters, enhancing LLM knowledge efficiently at test time.
RobustPT dynamically disentangles modality-specific prompts before softly aligning them, reducing channel bias and improving robustness when vision or language is missing.
G-SAP builds refined evidence graphs, derives structure-aware prompts for frozen language models, and performs heterogeneous graph-text message passing for commonsense reasoning.
NICE · Invited talk Video ↗
EMNLP 2026 · COLM 2026
NeurIPS 2026 Workshops: Meta-Agents · TTCL · Verify-Agents · Pre-to-Post
ICMR 2025 · NeurIPS 2025 Efficient Reasoning Workshop