Developing a proposal-stage framework, Reasoning-Aware Latent Diffusion (RALD), that trains masked latent token slots as reasoning states and adaptively allocates latent compute based on uncertainty. Targeting reasoning benchmarks such as Sudoku, Zebra, Countdown, GSM8K, and MATH500, with possible extensions to medical VQA and multimodal medical reasoning.

正在推进proposal-stage框架 Reasoning-Aware Latent Diffusion (RALD)：将masked latent token训练为推理状态，并依据不确定性自适应分配latent compute。计划评估Sudoku、Zebra、Countdown、GSM8K、MATH500，并探索医学VQA与多模态医学推理扩展。

Reliable and verifiable AI generation research. Key outputs: MARIE (ACM MM 2026, co-first author, under review; rebuttal completed), a first-author ACL ARR / EMNLP 2026 submission on uncertainty quantification for LLM-based recommendation, and GOING (TIST, first author, under review). Focus on evidence-committed generation, source-aware reliability evaluation, and training-free RAG robustness.

可靠且可验证的AI生成研究。核心成果：MARIE（ACM MM 2026，共同第一作者，审稿中；rebuttal已结束）、一篇第一作者ACL ARR / EMNLP 2026不确定性量化投稿，以及GOING（TIST，第一作者，审稿中）。研究重点包括证据承诺生成、来源感知可靠性评估与免训练RAG鲁棒性。

Built high-throughput real-time data pipelines for HFT with Python & SQL. Managed a 10M RMB simulation system validating ML-based trend-tracking strategies under volatile market conditions.

用Python & SQL构建高吞吐实时数据管道用于高频交易；管理千万规模仿真系统，验证ML趋势跟踪策略。

Automated preprocessing pipelines for 500+ companies with Pandas; scraped and structured financial logs from Wind/Flush databases.

用Pandas为500余家公司自动化数据预处理；从Wind/Flush抓取并结构化金融日志。

Weekly office hours mentoring undergraduates in ML, statistics, and Python/R.

每周答疑，辅导本科生机器学习、统计学与Python/R。