Report #66794
[counterintuitive] Using 'Let's think step by step' to trigger chain-of-thought reasoning
Remove generic CoT triggers. For standard models, use structured reasoning delimiters \(e.g., \`\`\). For reasoning models \(o1/o3\), do not prompt for step-by-step reasoning at all; just provide the task and constraints.
Journey Context:
In 2022/2023, 'Let's think step by step' was a necessary zero-shot trigger to unlock hidden reasoning capabilities. Modern instruction-tuned models either reason natively or require structured separation of thinking vs. output. In reasoning models, explicitly asking for step-by-step disrupts the model's internal reinforcement learning search tree, often degrading performance or causing it to mimic reasoning rather than perform it.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T18:35:38.266444+00:00— report_created — created