Get Started
Home
Topics
Search
Library
Research questionHow can reasoning-model fine-tuning improve pass@1 without reducing pass@k coverage?At decision points with multiple valid reasoning paths, fine-tuning may concentrate generation on fewer paths. The causes of this coverage loss and ways to preserve diverse valid solutions remain unclear.
AI
Evaluation & Benchmarks
LLM Pretraining & Post-training
Machine Learning
Reasoning
Latest papersRecent research connected to this question, newest first.Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the RoadThe evidence concerns SFT-based post-training for reasoning models, using controlled cases involving indecipherable nodes in graph branching and different reasoning modes. It examines decision-point data, targeted data synthesis, and diversity-aware decoding, with pass@1 and pass@k as key outcomes.research paper · Sep 1, 2026
Related questions
How should vision-language model post-training balance perception and reasoning under SFT versus reinforcement learning?How should on-policy distillation select examples for data-efficient reasoning post-training?How can large language models allocate reasoning computation to preserve accuracy under limited training and inference budgets?How do pause tokens affect reasoning adaptation while preserving previously learned capabilities during LLM fine-tuning?