Distilled reasoning models trust the reasoning you give them more than the reasoning they write

54 · · June 25, 2026, 4:23 p.m.
Summary
This blog post discusses the effectiveness of distilled reasoning models in tracking answers based on provided reasoning chains versus self-generated steps. The findings reveal that when models are given explicit reasoning chains, their answers closely align with those steps, while self-written reasoning often has less impact on the outcome, suggesting that the quality of reasoning depends significantly on the steps chosen.