Study Examines Failure Profiling and Trajectory Selection in Reasoning Distillation
A new arXiv paper proposes failure profiling and reachable trajectory selection to improve distilling mathematical reasoning from larger teacher models into smaller language models.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-09 05:00 · arXiv cs.CL
Failure Profiling and Reachable Trajectory Selection for Reasoning Distillation