Small language models stumble badly at physics — and a new training framework aims to stop the bleeding before one wrong step ruins everything downstream.
Researchers published a method that detects the first reasoning error in a multi-step physics problem, generates targeted structured feedback, and then trains the model to revise its own solution using policy gradient optimization with KL regularization. Crucially, the model never sees the correct answer as a training target — it learns to fix its work without being handed the solution. The external verifier that catches errors is used only during training, not at inference time, so no preference data or human annotation is required.
The results are notable for small models, which are often written off as incapable of reliable scientific reasoning. Across five physics benchmarks, the framework improves accuracy by 17-20% over standard chain-of-thought prompting and 10-16% over the next-best baseline. Calculation errors dropped from 56.9% to 23.5%; miscomprehension errors fell from 22.3% to 12.0%.
The harder problem — conceptual errors — proved more stubborn. Those fell from 89.7% to 68.7%, which is progress, but still means small models misunderstand the underlying physics nearly seven times out of ten in the worst cases. That gap matters for anyone tempted to deploy lightweight models in scientific or educational contexts where getting the concept right is the entire point.