Google's Thinking-in-Bets logs agent checks before user pushback to cut wrongful revisions

Google researchers describe Thinking-in-Bets, a method in which an LLM agent privately records its main claims, supporting checks and confidence before it sees a user's reply.

When challenged, the agent re-checks only the disputed claims. According to a Rohan Paul post summarizing the paper, Gemini agents abandoned correct requirements after a bare "that's wrong" complaint 12.2% of the time without the method, versus 0.02% with it. In report reviews, the pass rate was 89% when the checklist was written before the complaint and 65% when written after, according to the same post.