Can small models teach themselves to reason?

​As frontier LLMs improve at programming and mathematical reasoning, often via reinforcement learning, or large reasoning datasets, we ask: can smaller models (<10B) improve using only their own outputs?

In this work (THINK, PRUNE, TRAIN, IMPROVE: SCALING REASONING WITHOUT SCALING MODELS released by Stanford University researchers), the team investigates the conditions that enable this model self-improvement.

They introduce Think, Prune, Train, a scalable framework that iteratively fine-tunes small models on their own reasoning traces, using ground-truth pruning to ensure data quality. This approach leads to improved performance on math and coding benchmarks.

Join the BuzzRobot community to stay in touch