Large language models learn better from key reasoning steps than full detailed traces
Revisiting Complete Reasoning Traces for Post-Training
Computation and Language
Summary
Large language models are usually trained using long detailed reasoning paths that include many steps to reach an answer. The authors found that models do not gain much from learning the entire long path, and can perform just as well when given only parts of the reasoning or just the start and end points. Their analysis shows many intermediate steps add little unique information, so models can fill in gaps using their own knowledge. Training models on these shorter trajectories changes how they reason and helps improve related training methods.
large language modelspost-trainingreasoning trajectoriessupervised fine-tuningattention analysistoken removalreinforcement learningon-policy distillationredundancy in reasoningmodel behavior
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.