ReAct interleaves reasoning with actions, so the model thinks about what to do, calls a tool, reads the observation, and thinks again. Nearly every coding and browser agent runs some variant of this pattern, which makes it the baseline for reasoning about how a poisoned observation can redirect an agent.
Read in this order
- ArticleReAct: Synergizing Reasoning and Acting in Language Models (Google Research blog)Start here: a short, readable overview of the idea with worked examples.
- PaperReAct: Synergizing Reasoning and Acting in Language ModelsThe original paper, for the exact method and benchmark results.