Language models may not need a teacher to get better at reasoning — just a mirror.
Researchers have published a technique called Self-evolving Post-Training, or SePT, that lets a model improve its reasoning by repeatedly sampling its own outputs and training on them. There are no external reward signals, no human-labeled data, and no separate critic model. The loop works like this: sample questions, have the model generate answers at a set temperature, train on those answers, then repeat — each round using the freshest version of the model to generate the next batch. Tested across six math reasoning benchmarks, SePT outperformed a strong baseline of the same untuned model evaluated at its optimal decoding temperature.
The interesting part is what SePT does not require. Most recent reasoning improvements — reinforcement learning from human feedback, process reward models, verifier-guided search — depend on something outside the model to judge correctness. SePT sidesteps that entirely, which cuts both the cost and the pipeline complexity. If this holds up outside math benchmarks, it could make reasoning fine-tuning substantially more accessible for teams without the budget for large-scale annotation.
The caveats are real: math is an unusually clean domain where self-generated answers can be verified by structure alone, and the researchers themselves note the importance of temperature dynamics and online data refresh in making the loop work. Whether SePT transfers to messier, open-ended reasoning tasks is the question no one has answered yet — and that answer will determine whether this is a general technique or a tidy math trick.