few-shot-cot — Lean Prompting

The strongest evidence for expiry anywhere on this site sits here. On the Qwen2.5 series, exemplars containing reasoning chains do not beat zero-shot CoT — and injecting 50% noise into those exemplars did not degrade results (arXiv:2506.14641).

Read that mechanism carefully, because it is the whole argument: if corrupting your examples changes nothing, the model was not reading them for their content. You are paying tokens for decoration.

Claims (1)

few-shot-cot · reasoning · qwen2.5

redundant confidence: strong no check recorded

No improvement over zero-shot CoT

CoT exemplars do not improve reasoning over zero-shot CoT on the Qwen2.5 series; their remaining function is output-format alignment. Injecting 50% noise into exemplars matched or slightly beat the clean baseline on Qwen2.5-72B, implying the model selectively ignores them. arXiv:2506.14641

Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot · EMNLP 2025 Findings, 2025 not read in full

Contested by: few-shot--domain-multimodal--gpt-4o