Chain of Thought
Chain-of-thought prompting elicits or represents intermediate reasoning traces before a final answer, giving an LLM workflow additional structure but not by itself a guarantee of correctness, faithfulness, or efficient execution.
核心思想
A prompt or agent workflow asks a model to decompose a task into intermediate steps, which can help planning, tool use, or verification. The resulting trace is model output: it may be useful for control and debugging, but it can be incomplete, post-hoc, or inconsistent with the final decision unless an external evaluator checks it.
为什么重要
Agent and automated-research systems use intermediate traces to coordinate planning and artifact production. The systems question is how to turn a trace into an auditable action: constrain it with tools, tests, code execution, or verifiers, and report its token/latency/cost impact.
关键观察 / 隐含假设
- 观察：reasoning traces require external evaluation to become reliable system actions. AI-Scientist-arXiv24 and Auto-Research-arXiv25 use agent workflows with evaluator boundaries.
- 观察：long reasoning can interact with memory and serving cost. SkipKV-MLSys26 examines a systems path affected by such workloads.
- 假设：a visible trace faithfully explains model behavior. Kosmos-AI-Scientist-arXiv25 and RD-Agent-Quant-arXiv25 illustrate why tool/verifier evidence is stronger than narrative trace alone.
设计空间与取舍
- Free-form vs structured traces：structure improves control but can constrain exploration.
- Trace length vs cost/latency：more steps consume context and serving resources.
- Self-critique vs independent verification：model-generated criticism is weaker than executable or external checks.
引用本概念的论文
- AI-Scientist-arXiv24 — agent workflow with evaluation stages.
- Auto-Research-arXiv25 — automated-research pipeline.
- SkipKV-MLSys26 — serving/memory implications of reasoning workloads.
- RD-Agent-Quant-arXiv25 — tool-oriented agent workflow.
- Kosmos-AI-Scientist-arXiv25 — long-horizon agent/research context.