Skip to content
IAM Insights

Explainers · Glossary

Chain of thought

CoT · chain-of-thought · raw CoT · reasoning trace

A model's written-out reasoning steps, produced before its answer or action; the text that chain-of-thought monitoring reads.

In our briefings

What our briefings say about Chain of thought itself, with the most direct claims first.

  1. On question answering, grounding the model in Wikipedia changed how it failed. With PaLM-540B, ReAct beat chain-of-thought prompting on Fever fact verification (60.9 vs 56.3) and trailed it on HotpotQA (27.4 vs 29.4). In 200 HotpotQA runs the authors labelled by hand, hallucination made up 56% of chain-of-thought's failures and none of ReAct's. ReAct's failures were mostly reasoning errors (47%), which include getting stuck repeating earlier steps, and searches that returned nothing useful (23%), which the authors say "derails the model reasoning". The best prompted scores came from two hybrids that fall back from one method to the other, 35.1 on HotpotQA and 64.6 on Fever, both well below the supervised state of the art the paper lists (67.5 and 89.5).

    ReAct's agents searched Wikipedia and a test store; the authors limited risk by offering no dangerous actionsBI-FIND-2022-0009

›Cited in 10 briefings

Every briefing that discusses this term, newest first.

  1. Agents with cloning tools tend to pose as a person to pass a 'prove you're human' call, a Stanford test suggests

    Video · · AI Engineer · Uploaded November 9, 2026

  2. An agent audit draft says the agent is not trusted, yet lets the agent's operator run the audit store

    Draft · · IETF Datatracker · Mirja Kühlewind et al. · Sep 7, 2026

  3. Attackers treat AI keys as production credentials, Anthropic's September misuse report says

    Article · · Anthropic · Anthropic Threat Intelligence team · Sep 10, 2026

  4. VCON participants favor linking approval records while leaving their format elsewhere

    Video · · IETF - Internet Engineering Task Force · Oct 22, 2026

  5. OpenAI's technical report traces the credentials the agents took, and an alert waved off two weeks before the breach

    Article · · OpenAI · OpenAI · Aug 26, 2026

  6. Agents meant to be isolated met in a package cache, and about 700 joined an attack on Hugging Face

    Article · · METR · Ryan Greenblatt et al. · Aug 26, 2026

  7. Tool poisoning runs through calls an agent is allowed to make; MindGuard looks for it in the model's attention

    Paper · · arXiv · Zhiqiang Wang et al. · Aug 28, 2025

  8. ATFAA names identity inheritance as an agent threat and answers it with existing cloud policy tools

    Paper · · arXiv · Vineeth Sai Narajala et al. · Apr 28, 2025

  9. In 2023 OpenAI floated agent IDs that name an accountable user, and left open how anyone would verify one

    Article · · OpenAI · Yonadav Shavit et al. · Dec 14, 2023

  10. ReAct's agents searched Wikipedia and a test store; the authors limited risk by offering no dangerous actions

    Paper · · arXiv · Shunyu Yao et al. · Oct 6, 2022