Beijing Institute of Technology · Meituan LongCat foundation-model team
I am a Ph.D. student in Computer Science at Beijing Institute of Technology, advised by Prof. Dandan Song. I also work on LLM post-training with Meituan's LongCat foundation-model team.
My research asks how a model's failures can become its next training signal. I build systems that locate the first consequential error, verify outcomes with executable feedback, and assign learning credit to the decisions that caused the failure. This connects my work on production post-training loops, verifiable reinforcement learning, structured reasoning, and self-improving agents.
Deployed models produce a stream of supervision that is both honest and underused: their own failure traces. I built a Bad Case Auto-Remediation loop that discovers recurring failures, identifies the first error, synthesizes graders and repair data, launches training, and runs targeted and general regressions. A related project, Zero Auto Harness, holds the base model fixed and instead lets its tools and middleware improve from execution feedback.
Once correctness can be checked by a program, the hard question becomes which decision deserves the credit. In ToolAgent-RL, I decomposed multi-turn tool-use trajectories into 6,498 state-action decisions and built executable rewards for protocol validity, tool routing, and argument accuracy. My post-training work uses token-level advantage masks to focus sequence-level feedback on attributable errors and preserve correct spans during optimization.
Rather than asking a model to emit an answer directly, I study intermediate structures that an executor can inspect. This line includes CompKBQA for component-wise reasoning and retrieval, GRV-KBQA for decoupled generation, grounding, and validation, and SELF-KBQA for structure-aware subgraph retrieval and token-budgeted evidence condensation.
I also use internal representations to test whether a behavioral result reflects the mechanism we think it does. My work studies hallucination signals in deep representations and uses probing to connect layer-wise features with downstream predictions.
– Present
Research Engineer · LLM post-training
Designed and shipped a closed loop from production failure discovery and first-error localization to repair-agent data generation, automatic graders, token-level advantage masks, training, and regression evaluation. A single failure type can be repaired from as few as 8 automatically mined examples.
Pure-JSON compliance improved from 77.53% to 99.78%; explicit and implicit Markdown-fence cases improved from 34.29% to 77.96% and 55.56% to 88.89%. The target failure share fell from 39.5% to 2.0%; masking added roughly 7 percentage points on the targeted metric while the nine-task general macro average held steady (61.19% to 61.33%).
Helped build a two-agent work/evolve loop that improves tool implementations and middleware from execution traces while keeping the base model fixed. Added cross-task generalization constraints and held-out re-evaluation to reduce benchmark memorization. With GPT-5.4 fixed, five evolution rounds raised SWE-bench Verified 500-task Pass@1 from 51.8% to 61.0%.
–
LLM Research Engineer · Quantitative research
Built a multi-agent workflow spanning strategy generation, sandboxed code execution, and automatic backtesting. Backtest outcomes served as verifiable feedback for selecting and iterating proposed alpha factors—my first production setting where execution could replace manual labels.
Led research-report summarization and market-sentiment factor modeling, addressing long-context truncation, class imbalance, entity-name leakage, and factor collinearity before production deployment. I also built RAG systems with semantic chunking, hybrid dense/sparse retrieval, Milvus, LLM reranking, and evidence attribution.
Ph.D. student in Computer Science · Advisor: Prof. Dandan Song
– Expected
BIT Graduate Special Scholarship · Outstanding Student
B.Eng. in Computer Science
–
ACM-ICPC Asia Regional Silver Medal · GPLT Silver Medal · National Endeavor Scholarship
13 papers, including 4 first- or co-first-authored papers. See also Google Scholar.
ACL 2026 · First author
SELF-KBQA retrieves structurally aligned subgraphs, condenses them under a token budget, and generates executable logical forms.
EMNLP 2025 · First author
A component-wise framework that learns skeleton, entity, relation, and logical-form subtasks and retrieves knowledge dynamically with R³.
EMNLP 2025 · Co-first author
Separates structure generation, semantic grounding, and validation; structure-aware filtering removes 90% of invalid candidates.
EMNLP 2024 · First author
Combines an unsupervised two-stage ranker with multi-hop graph beam search, improving GrailQA zero-shot performance by 8.1 EM.
WWW 2026
ACL 2026
ACL 2026
AAAI 2026
EMNLP 2025
IJCAI 2025
EMNLP 2024
WWW 2024
IEEE Transactions on Knowledge and Data Engineering