The original paper by Wei and colleagues used worked reasoning examples in the prompt and found improvements on arithmetic, commonsense, and symbolic reasoning benchmarks. Results depended on the model and task; the technique is not a universal improvement for every request.
Few-shot prompting supplies demonstrations. A zero-shot reasoning cue asks for steps without those examples. Both differ from reasoning models trained to spend additional computation before answering.
Example: a model can calculate each step of a pricing problem correctly while starting from the wrong discount policy. Check the evidence and result, rather than treating a detailed explanation as an audit trail.
Turpin and colleagues found that generated chains can rationalize answers influenced by biasing cues without mentioning those cues. That is a reason to distinguish useful intermediate work from guaranteed transparency. Dan Klein’s episode 54 discusses the gap between plausible outputs and reliable deployment. For the broader progression from prompting to inference-time reasoning, use the show’s reasoning overview.
Sources
- Wei et al.: Chain-of-Thought Prompting — Original benchmark evidence for prompting with intermediate reasoning demonstrations.
- Turpin et al.: Unfaithful Explanations — Shows that generated explanations can omit influences on a model's answer.
Go deeper
- What is the difference between prompt, context, and memory engineering? AI, decoded · Prompt vs. Context vs. Memory Engineering
- When should you use a reasoning model instead of a standard LLM? AI, decoded · Reasoning Model vs. Standard LLM: When to Use Which
- Kojima et al.: Large Language Models are Zero-Shot Reasoners paper
Compare the paper's no-example reasoning cue with worked demonstrations, then inspect its separate answer-extraction prompt. Results concern the models and benchmark tasks studied.
