CoT deep reading

Series · 1 posts

← Back to Paper Reading

  • CoT: Make the Model Write the Reasoning, but Do Not Treat It as an Agent That Moves

    Intermediate Agent runtime, safety, and evaluation
    CoT deep reading: Part 1 , Note: Aug 27, 2026 , Paper: 2022 , NLP

    A source-grounded reading of Wei et al., NeurIPS 2022: few-shot exemplars with intermediate steps elicit multi-step reasoning in large frozen models. PaLM 540B on GSM8K moves from 17.9 to 56.9; this is still a prompt, not tools, an environment, or memory paging.

    Understand it in 90 seconds
    Problem
    Standard few-shot prompting gives only $\langle$question, answer$\rangle$ pairs. Multi-step arithmetic, commonsense, and symbolic tasks stay weak, and scaling the model alone often leaves those curves flat.
    Core insight
    Change the exemplar to $\langle$question, intermediate reasoning, answer$\rangle$. The decision point moves from “answer immediately” to “write the work, then answer.” Weights stay frozen. This is still a prompt, not an agent.
    Strongest evidence
    PaLM 540B on GSM8K moves from 17.9 to 56.9, above Cobbe et al.’s finetuned GPT-3 + verifier at 55 (Table 1, Figure 2). Figure 4 / Table 2 show the gain appearing around 100B parameters.
    Main boundary
    No environment, no tools, no memory paging. Small models often get worse. A chain can be unfaithful, or luckily reach the right number. Self-consistency (Wang et al., 2022a) is a later paper; the main tables here use greedy decoding.
    Read the full deep dive

For speaking invitations, internal engineering sessions, or architecture exchange, see the topics and public work I can bring into the conversation.

Speaking & contact