💼 Experience

July 2025 – Present · Researcher, Beidou Program

I contribute to the development and releases of the LongCat model family, focusing on mid-training data curation and model evaluation.

Mid-training data curation. Turn user interaction logs into quality-controlled agent trajectories through preprocessing, session aggregation, filtering, and de-identification. Analyze data retention across sources and batches, and use small-model training experiments to inform data admission.
AI at Work evaluation. Build and iterate on tasks, environments, and rubric-based LLM grading for productivity scenarios. Analyze evaluation failures and adapt the evaluation pipeline for comparisons with external models. [FORTE]
Reasoning evaluation. Build challenging benchmarks for mathematical and general reasoning, combining reliable automated grading with analyses of token efficiency, test-time scaling, and cross-lingual robustness. [AMO-Bench] [General365]

August 2019 – June 2025

Long-context understanding. Developed information-intensive training and FILM-7B to address information loss in long contexts. The training data was adopted in the Microsoft Phi-3 model family. [FILM]
Learning from mistakes. Explored mistake-correction training to improve mathematical reasoning and data efficiency. [LEMA]
In-context learning and code generation. Developed skill-based example selection to improve semantic parsing and code generation, with related techniques adopted in Microsoft DBCopilot.
Compositional generalization. Systematically evaluated language models’ ability to recombine learned concepts, and developed lightweight models with near-human compositional generalization performance.