When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents
Runxuan Tang and colleagues (Nanyang Technological University, Georgia Tech, University of Washington) propose Falsification-Gated Self-Evolution (FGSE), which treats every lesson an agent wants to save to memory as a hypothesis to be tested before it is committed.
Ask this paper
Failure mode. Lessons written straight from a successful episode often encode shortcuts or overly broad conditions, so they help on the source task and then regress archived behavior on similar but different tasks.
Structured hypothesis. Each candidate lesson states a precondition, behavior change, expected effect and verifier. It is applied only in a temporary state and tested on target-transfer, falsification and archived-regression probes.
Commit, refine or reject. Measured gain and risk decide whether the lesson enters persistent memory, is revised, or is dropped.
Results. FGSE reaches 39.8 WebArena success, 78.8 AppWorld completion and 0.724/0.487 Pass^1 on tau-Retail/tau-Airline, while cutting falsification failures to 5.0%, regression damage to 1.3% and harmful committed updates to 3.6%.
Abstract
LLM agents increasingly operate in long-horizon, verifier-rich environments, but current self-improvement mechanisms often write episode-derived lessons directly into memory, making local success an unreliable signal for reusable and safe knowledge. This paper addresses the core question of when an experience-derived behavioral update should be admitted into persistent agent state. We propose Falsification-Gated Self-Evolution (FGSE), a verifier-grounded framework that represents each candidate lesson as a structured hypothesis with an explicit precondition, behavior change, expected effect, and verifier. FGSE applies the hypothesis only in a temporary state, tests it on target-transfer, falsification, and archived-regression probes, and then commits, refines, or rejects it according to measured gain and risk. Across web, app, tool-use, and long-term memory benchmarks, FGSE achieves strong task performance, including 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass$^1$ on $τ$-Retail/$τ$-Airline, while reducing falsification failures to 5.0%, regression damage to 1.3%, and harmful committed updates to 3.6%. These results suggest that testing memory updates as hypotheses supports more reliable long-term adaptation.