Reasoning Trace

A reasoning trace is the explicit chain of logical steps that an AI model displays as it works through a problem. Rather than jumping directly to an answer, the model shows intermediate thoughts, calculations, and deductions that lead to its conclusion.

How reasoning traces work

Modern "reasoning models" like OpenAI's o1 and o3, Anthropic's Claude with extended thinking, and DeepSeek-R1 are specifically trained to produce detailed reasoning traces. When given a complex problem, they generate a visible thought process — breaking down the problem, considering approaches, working through steps, checking their work, and arriving at a conclusion.

This is distinct from chain-of-thought prompting, where the user asks for step-by-step reasoning. Reasoning models produce traces by default as part of their architecture and training.

What reasoning traces look like

A reasoning trace might include problem decomposition ("I need to find X, which requires knowing Y and Z"), strategy selection ("I will approach this by first calculating..."), intermediate calculations, self-correction ("Wait, that does not account for..."), verification ("Let me check this result by..."), and a final answer.

Benefits of reasoning traces

• Transparency: Users can see how the model arrived at its answer, making it easier to trust or question the result.
• Error detection: When a reasoning step is wrong, it is visible and can be identified. With hidden reasoning, you only see the final wrong answer.
• Debugging: If the output is incorrect, the trace shows where the reasoning went wrong.
• Learning: Users can learn from the model's problem-solving approach.

Limitations

Reasoning traces may not accurately reflect the model's actual internal computation — they are generated text that might rationalise rather than truly explain. Traces significantly increase token usage, making responses more expensive. And longer traces do not always mean better answers.

When reasoning traces matter most

Reasoning traces are most valuable for complex problems — math, logic, code debugging, scientific analysis, and strategic planning. For simple tasks like writing an email or answering a factual question, the overhead of a detailed trace is unnecessary.