Chain of Thought
Chain-of-thought prompting asks the model to emit intermediate reasoning steps before the answer, which reliably lifts arithmetic, symbolic, and multi-hop tasks.
Definition
Wei et al. (2022) show that few-shot exemplars with written steps, or the phrase “let’s think step by step,” raise accuracy on math and logic for large enough models. The tokens are not a proof — they are extra compute and a different output distribution.
Why it matters
CoT is the cheapest “test-time compute” knob. It also creates a new failure mode: fluent wrong scratchpads that persuade humans more than a short wrong answer would.
Case
GSM8K jumps when a 100B+ model is allowed to scratchpad. A 7B model often writes confident nonsense steps. Size is part of the effect.
How to read it
Score the final answer on a locked set. Do not confuse a pretty trace with correctness. If you train on traces, you are doing process supervision — different paper, different claims.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsWei et al. · 2022 · NeurIPSCoT prompting.