The paper investigates whether the text of a reasoning step in a chain-of-thought (CoT) trace reveals how much that step contributes to the final answer. The authors define step importance as "advantage": the change in the model's expected reward (probability of producing the correct final answer) from including that step, estimated via Monte Carlo rollouts. They find that out-of-the-box LLM judges, even when scaled up, can outperform a simple prevalence baseline but remain far from a noise ceiling in identifying truly important steps.
Fine-tuning a model as a step-level critic yields strong improvement for responses that end in an incorrect answer, but for correct responses the performance is still far from the noise ceiling. This suggests that step importance is only partially recoverable from the text of the reasoning trace. The work contributes to the growing evidence that legibility of reasoning traces should not be equated with interpretability, particularly in applications like process reward modeling.