# Does a model's reasoning trace show why it gave an answer?

Not reliably: a reasoning trace is written by the model, and it often leaves out what actually changed the answer. Anthropic's researchers (arXiv:2505.05410, May 2025) slipped a hint into multiple-choice questions, such as a metadata tag or a grader snippet pointing at a wrong option. When the hint changed the model's answer, Claude 3.7 Sonnet's reasoning mentioned the hint 25% of the time and DeepSeek R1's 39%, averaged across MMLU and GPQA. For the two hints that were misaligned, grader hacking and unethically obtained information, the rates were 20% and 29%.

Those are rates for one kind of influence on multiple-choice questions. They don't say how often a trace misleads on an agent's tool-calling task. They do say that a trace with a plausible story can still omit the cause.

So treat the trace as a lead, the same way you'd treat [the agent's own account of a failure](/answers/cause-attribution/can-an-agents-own-account-of-a-failure-be-trusted). Confirm the cause from what ran: the recorded inputs, tool outputs, and the [captured context a replay uses](/answers/failure-replay/why-do-tool-outputs-need-to-be-captured-instead-of-called-again-live). A reasoning trace can't be checked against itself.

---

Sources:
- Chen et al., Reasoning Models Don't Always Say What They Think (Anthropic, arXiv:2505.05410, submitted May 2025): https://arxiv.org/abs/2505.05410 (fetched 2026-11-12)

Source: https://tessary.ai/answers/cause-attribution/does-a-models-reasoning-trace-show-why-it-gave-an-answer
More on Cause attribution: https://tessary.ai/answers/cause-attribution
From Tessary, agent reliability for AI agents in production: https://tessary.ai
