● ResearchNov 13, 2026 · 12:001 source

Cleaning an Agent's History Hurts When the World Still Holds the Error

From arxiv.org

AI;DR

Tool-using language-model agents accumulate histories of their own mistakes. On AgentCE-Bench, adding a truthful note beside a preserved error raised recovery from 51% to 66% when the world state was held fixed.

Read originalSign in to suggest
Read original