When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents
FGSE tests an agent’s proposed lessons before saving them to persistent memory, reporting harmful committed updates at 3.6%.
Each lesson specifies a precondition, behavior change, expected effect and verifier. The framework trials it temporarily against transfer, falsification and archived regression probes, then commits, refines or rejects it based on measured gains and risks. Across web, app, tool-use and long-term memory benchmarks, the authors report a 39.8% WebArena success rate and 78.8 average completion on AppWorld. They also report falsification failures of 5.0% and regression damage of 1.3%.
ArXiv · AI/CL/LG's note
Each lesson specifies a precondition, behavior change, expected effect and verifier. The framework trials it temporarily against transfer, falsification and archived regression probes, then commits, refines or rejects it based on measured gains and risks. Across web, app, tool-use and long-term memory benchmarks, the authors report a 39.8% WebArena success rate and 78.8 average completion on AppWorld. They also report falsification failures of 5.0% and regression damage of 1.3%.
ArXiv · AI/CL/LG's note
score 4