Full Professor, Tsinghua University
1 paper at NeurIPS 2025
We train a token-level neural router to let SLM following LLM reasoning paths by replacing only divergent tokens