ReAct Thought Editing

Let humans correct a few reasoning traces mid-episode so later actions realign — cheaper than retyping action sequences or retraining the policy.

Because ReAct thoughts are readable language in the same transcript as acts, humans can inspect and edit them on the fly: delete a hallucinated belief, insert a subgoal hint, or change reasoning style. Later actions often realign after only a couple of thought edits — far less effort than typing tens of corrective actions, and unlike Act-only or classical RL where changing a few past actions rarely reshapes the policy.

This is stronger than dialogue-only goal updates: edits can rewrite internal beliefs the flexible thought space supports. Expose it as an agent UX steering affordance (editable thoughts + regenerate forward) and pair with human approval gates when acts have real side effects. Coupling LLMs to web or physical action spaces still risks harmful lookups or irreversible acts — keep blast-radius limits in the harness even when thoughts look aligned.

---

Sources

• ReAct (Yao et al.), pp. 1–15