Reasoning Traces

Reasoning traces are the visible, step-by-step thinking processes that AI models produce when working through complex problems. Rather than simply providing a final answer, the model shows its work — each logical step, intermediate conclusion, and decision point — allowing users to follow and verify the reasoning.

Why reasoning traces matter

A model that says "The answer is 42" provides no basis for trust. A model that says "First, I calculated X. Then I considered Y. Based on these factors, the answer is 42" gives you the ability to verify each step. This transparency is critical for high-stakes decisions.

Types of reasoning traces

• Chain-of-thought traces: The model works through a problem sequentially, each step building on the previous one. "To find the total cost, I first multiply the quantity by the unit price, then add shipping..."
• Deliberative traces: The model explicitly considers multiple options and explains why it chose one over others. "Option A has lower cost but higher risk. Option B has higher cost but is more reliable. Given the constraints, Option B is preferred because..."
• Tool-use traces: The model explains which tools it is calling and why. "I need to check the current exchange rate, so I am calling the currency API..."
• Extended thinking traces: Models like Claude and OpenAI's o-series produce extended reasoning traces that can be hundreds of tokens long, working through complex problems methodically.

Reasoning traces for verification

The primary value of reasoning traces is verifiability:

• If the final answer is wrong, you can identify exactly where the reasoning went astray.
• Intermediate steps can be checked independently against known facts.
• Logical errors and unsupported assumptions become visible.
• Domain experts can evaluate the reasoning without needing to reproduce the analysis from scratch.

Reasoning traces for debugging

When an AI application produces unexpected results, reasoning traces are invaluable for diagnosis:

• Was the model working with incorrect information?
• Did it misinterpret the question?
• Was the reasoning logically valid but based on a wrong premise?
• Did it skip an important consideration?

Limitations

Reasoning traces should not be taken at face value:

• Faithfulness: A model's stated reasoning may not accurately reflect its actual internal processing. The trace is generated text, and like all generated text, it can be misleading.
• Post-hoc rationalisation: Models sometimes generate plausible-sounding reasoning for conclusions they reached through different (possibly flawed) internal processes.
• Length does not equal quality: Longer reasoning traces are not necessarily better. Some models pad traces with redundant steps.

Best practices

• Require reasoning traces for any decision with significant consequences.
• Verify critical steps independently rather than trusting the trace implicitly.
• Use traces for debugging and quality improvement, not just as window dressing.
• Train your team to read and evaluate reasoning traces effectively.