Researchers have revisited a common assumption in large language model (LLM) post-training: that models need complete, pre-collected reasoning trajectories to improve their reasoning abilities. According to a new paper titled Revisiting Complete Reasoning Traces for Post-Training, full trajectories provide only limited benefit, while partial trajectories remain effective even under heavy truncation.
The study began with a pilot investigation and then analyzed redundancy in reasoning traces through attention-based analyses and controlled token-removal experiments. Both approaches indicated that intermediate tokens contribute minimally to final reasoning quality, suggesting that models may infer missing steps from internal knowledge when trajectory endpoints are known.
Based on these findings, the authors propose endpoint-based supervised fine-tuning (E-SFT), which uses only the beginning and ending parts of a reasoning trace. They report that E-SFT retains or even improves SFT performance, and that endpoint-based training leads to consistent changes in reasoning behavior.
The paper further shows that endpoint-based training benefits post-training methods beyond SFT, including reinforcement learning approaches such as GRPO and on-policy distillation (OPD). The authors conclude that these results highlight the need to revisit the role of complete reasoning traces in post-training, and they have released code at https://github.com/naver-ai/revisiting-trace.