Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Reference: Wei, Wang, Schuurmans, Bosma, Ichter, Xia, Chi, Le, Zhou (2022). NeurIPS 2022. Source file: 2201.11903.pdf. URL

Summary

Shows that prompting a sufficiently large language model with a few exemplars that include intermediate natural-language reasoning steps — a chain of thought — dramatically improves performance on arithmetic, commonsense, and symbolic reasoning benchmarks. The technique requires no finetuning, no verifiers, and no task-specific training data beyond a handful of (input, chain-of-thought, output) triples.

Prompting PaLM-540B with eight chain-of-thought exemplars achieves then-state-of-the-art accuracy on GSM8K math word problems, surpassing even a finetuned GPT-3 with a verifier. Gains emerge only at scale (~100B parameters), making CoT one of the canonical examples of emergent capability. CoT underpins most modern agent reasoning loops (ReAct, ToT, Reflexion, LLM-Planner) referenced throughout AI Agents Under Threat and LLM Agents.

Key Ideas

• A prompt of input/rationale/output triples unlocks multi-step reasoning without any parameter updates.
• CoT is emergent: negligible benefit below ~100B parameters, striking gains above.
• Works across arithmetic, commonsense, and symbolic domains — the method is task-general.
• Rationale generation enables interpretability windows and lays the groundwork for reasoning-trace auditing in agent security.
• The externalised reasoning trace is itself an attack surface — later exploited by indirect Prompt Injection and adversarial rationale manipulation.

Connections

• LLM Agents
• AI Agents Under Threat
• Tool Use
• Prompt Injection
• Hallucination

Conceptual Contribution

• Claim: Providing a small number of exemplars with explicit intermediate reasoning steps in the prompt elicits multi-step reasoning in sufficiently large language models, yielding state-of-the-art performance without finetuning.
• Mechanism: Few-shot prompt template ⟨input, chain-of-thought rationale, output⟩; the model learns to generate its own rationales before the final answer at inference time.
• Concepts introduced/used: chain-of-thought prompting, emergent reasoning, prompt-only adaptation — the substrate for ReAct-style agent loops that structure LLM Agents and the reasoning traces attacked in AI Agents Under Threat.
• Stance: empirical
• Relates to: The reasoning-trace paradigm CoT popularises is the target of Prompt Injection, goal-hijacking jailbreaks, and the brain-layer threats catalogued in AI Agents Under Threat.

Tags

#llm #prompting #reasoning #foundational #emergent-capability