Complete reasoning traces add little value in LLM post-training, study finds

Large language models are commonly post-trained, for example through supervised fine-tuning, on collections of reasoning trajectories gathered in advance, with the goal of improving how well the model reasons. Those trajectories tend to run long, because they follow complex, interwoven paths to an answer that often include detours along the way. Whether models actually gain from learning the entire trajectory, detours included, rather than a shorter version, has been largely unexamined. Starting from a pilot study, the authors report that complete trajectories provide only limited benefit, while partial trajectories remain effective even under heavy truncation.

To explain that redundancy, the authors ran attention-based analyses alongside controlled experiments in which tokens were deliberately removed from trajectories. Both approaches point to the same conclusion: tokens in the middle of a reasoning trajectory contribute only minimally to the quality of the final reasoning. Their reading of this is that once the redundant material is stripped out, a model can fill in coherent intermediate steps on its own, drawing on knowledge it already has internally, as long as it is given the trajectory's endpoints.

The paper also reports that training language models directly on those endpoints, rather than on the full path to them, produces consistent changes in the models' reasoning behavior. The effect is not confined to supervised fine-tuning: the authors find it also benefits post-training methods based on reinforcement learning or on-policy distillation. Taken together, they argue this is reason to revisit how complete reasoning traces are used in post-training. Code for the study has been released on GitHub under the organization name naver-ai, in a repository called revisiting-trace; the paper's own text does not name an author or institution.

Key facts

A pilot study reports that complete reasoning trajectories provide only limited benefit in LLM post-training, while partial trajectories remain effective even under heavy truncation.
Attention-based analyses and controlled token-removal experiments both show that tokens in the middle of a reasoning trajectory contribute only minimally to the quality of the final reasoning.
Training models directly on a trajectory's endpoints, instead of the full path, produces consistent changes in reasoning behavior.
The endpoint-training effect also benefits post-training methods based on reinforcement learning or on-policy distillation, not just supervised fine-tuning.
The study's code is released on GitHub under the organization naver-ai, in the repository revisiting-trace; no author or institution is named in the paper's own text.

Why it matters

Post-training on collected reasoning trajectories, for instance through supervised fine-tuning, is a standard way to improve how well large language models reason, and those trajectories are typically long because they follow complex, interwoven paths that include detours before reaching an answer. This study challenges an assumption behind that practice: that a model needs the complete trajectory, detours included, to learn from it. The authors report that full trajectories add only limited benefit, that partial ones remain effective even under heavy truncation, and that training directly on a trajectory's endpoints changes a model's reasoning behavior in a consistent way. If that holds up, it reframes what a reasoning-training example actually needs to contain, and where the effort of collecting one should go.

Who it affects

The direct audience is teams building post-training pipelines for reasoning-focused language models: anyone assembling or curating reasoning-trajectory datasets for supervised fine-tuning, and, per the paper, anyone using post-training based on reinforcement learning or on-policy distillation, since the authors report the endpoint-training effect benefits those methods too. It is also relevant to researchers studying which parts of a chain of reasoning actually carry information, since the paper's token-removal and attention analyses speak directly to that question. It has no direct bearing on people simply using AI products day to day; the subject is a training-pipeline choice, not a feature they would see.

How to use it

The paper comes with released code, published on GitHub under the organization naver-ai in a repository called revisiting-trace, for readers who want to inspect or reproduce the trajectory analysis and the endpoint-based training approach it describes.

How solid is it

What is available here is the paper's own account of a pilot study. The authors back the central claim with two independent analyses, an attention-based analysis and a controlled token-removal experiment, that point to the same conclusion, and they report the endpoint-training effect extending beyond supervised fine-tuning to post-training based on reinforcement learning or on-policy distillation, which suggests the pattern is not confined to one training setup. Working against that: the text does not name a specific model, model size, benchmark or dataset used in the experiments, and it gives no percentage, score or other figure for how limited the benefit from full trajectories is, or for how much of a trajectory has to be removed to count as heavy truncation. Without those specifics, the size of the effect and how far it generalizes cannot be judged independently from what is available here.

Risks and caveats

The paper does not define a threshold for heavy truncation or say how few tokens an endpoint actually retains, so the practical recipe is not fully specified. It also does not name the specific reinforcement-learning algorithm or on-policy-distillation method used, only the general categories, and it makes no claim about compute or training-time savings from using partial trajectories or endpoints: the reported benefit is about reasoning quality, not cost or speed. No author names or institutional affiliations appear in the paper's own text; the linked GitHub code is hosted under the organization naver-ai, which hints at, but does not confirm, an affiliation.

“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.”

— the paper