Researchers train LLMs to check each step of their own reasoning

A new training method teaches language models to pause mid-problem and check each link in their chain of reasoning before committing to a final answer. Posted to arXiv on 12 Aug 2026 by researchers at NTU Singapore, the National University of Singapore, VinUniversity, and A*STAR's Institute for Infocomm Research, the paper introduces a two-stage framework that first strengthens step-level reasoning, then trains the model to verify and revise its own intermediate steps.

The method, called SFS-DPO (Self-Fix Step-DPO), separates the two skills most reasoning models conflate: producing a working chain of steps, and catching a wrong one before it snowballs into a wrong answer. A variant, SFS-DPO-R, adds teacher-supplied explanations to guide error verification.

The authors report that both variants outperform prior step-level training baselines on in-domain and out-of-domain evaluations across multiple language models, with gains in how often the model catches an error and how often its revision actually fixes it. Their premise: a confident final answer that papers over a flawed intermediate step is the failure mode that breaks multi-step reasoning, so step-level verification is what closes it. The full paper is on arXiv.

The honest scope: the results are author-reported on benchmarks, not independently replicated, and self-correction in language models remains an open problem. Step-level verification that sticks on a held-out math or code set is a measured advance on a narrow target, not a general fix for reasoning failures or a deployment-ready capability.