Large language models learn better from key reasoning steps than full reasoning paths
Revisiting Complete Reasoning Traces for Post-Training
Computation and Language
Summary
Reasoning steps used to train large language models are often very long and complicated, including many unnecessary detours. The authors found that these models don’t gain much from learning the entire long reasoning process. Instead, they can learn just as well, or even better, when trained on shorter reasoning parts or only the important start and end steps. This means models might fill in missing reasoning details on their own using what they know internally. The study also shows changes in how models reason when trained only on the key steps, encouraging new ways to improve post-training methods.
What this means in practice
- •For machine learning engineers: Reduce training time by focusing on key reasoning steps instead of full reasoning paths when fine-tuning large language models.
- •For ai product developers: Improve model behavior in tasks requiring reasoning by adjusting training to emphasize start and end steps rather than entire reasoning traces.
Authors
Jaehui Hwang, Sangdoo Yun, Byeongho Heo, Dongyoon Han
Abstract
Large language models (LLMs) are often post-trained on pre-collected reasoning trajectories to improve their reasoning capability. Such trajectories tend to be long due to complex, interwoven paths, which often include detours on the path toward the answer. However, it has been underexplored whether LLMs indeed benefit from learning complete trajectories in post-training, such as supervised fine-tuning (SFT). Starting from our pilot study, we find that full trajectories provide only limited benefit, while partial trajectories are effective even under heavy truncation. We analyze redundancy in reasoning trajectories through attention-based analyses and controlled token-removal studies, both of which show that intermediate tokens contribute minimally to final reasoning quality. This suggests that avoiding redundant information may allow LLMs to internally infer coherent alternatives by inferring missing steps from their internal knowledge, given known trajectory endpoints. Furthermore, we show that training LLMs using endpoints leads to consistent changes in reasoning behavior, and that it also benefits post-training methods based on reinforcement learning or on-policy distillation, highlighting the need to revisit complete reasoning traces. Code is available at https://github.com/naver-ai/revisiting-trace.