A new paper on arXiv investigates a simple but potentially powerful signal in large language model reasoning: the density of information in each step of a chain-of-thought. The authors ask whether variation in this density predicts whether the model's final answer is correct, and whether it can be used as a target for steering reasoning at inference time.
The study prompts frozen pretrained models to solve problems and examines the chain-of-thought steps. By measuring information density across those steps, the researchers aim to link that measure to answer correctness. The abstract does not yet report specific results, but the framing suggests a possible method for improving reliability without retraining.
If the hypothesis holds, information density could become a controllable inference-time target. That would mean adjusting the model's reasoning process on the fly, rather than changing its weights. The use of frozen models hints that the approach could be applied to existing systems without additional training.
The abstract is brief and leaves many details—such as the exact measure of information density and the tasks used—unstated. Still, the idea offers a fresh angle on interpretability and control for chain-of-thought reasoning.