Reasoning-trace endpoints can match full-trace training, researchers say

Language models can learn from just the beginning and end of a reasoning trace without losing fine-tuning performance—and sometimes improve on it—according to the authors of an arXiv paper submitted on September 7, 2026. Their method, endpoint-based supervised fine-tuning (E-SFT), leaves out the middle of each training example.

How much can safely be removed, and under what conditions? The available abstract and author summary give no model names, benchmarks, scores or truncation ratios. The full paper was not reviewed for this report, so the evidence here supports an account of the researchers’ claim, not a recommendation to adopt the method.

In Revisiting Complete Reasoning Traces for Post-Training, Jaehui Hwang and colleagues report using attention-based analyses and controlled token-removal studies to examine which parts of a trace matter. They say the intermediate tokens contribute little to final reasoning quality. Their proposed explanation is that a model given the endpoints may fill in missing steps from its existing knowledge; that remains an interpretation, not an established mechanism.

The authors also report benefits for reinforcement learning and on-policy distillation. Those claims extend the idea beyond supervised fine-tuning, but the summaries do not quantify the improvements or describe how the resulting reasoning behavior changed.