A small RL-trained model at each node of a recovery graph diagnoses drift in autonomous LLM agents on AppWorld

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv:2608.14109 (Aug 14, 2026, El Hamraoui, Jose, Bureau, Plana) proposes a plug-and-play drift recovery module: instead of retraining the expensive main agent model, a single small language model is RL-trained to specialize at each node of a recovery graph covering drift classification, operation detection, risk evaluation and recovery decision. Rewards combine rule-based structural checks on XML output schema with LLM-as-judge semantic quality. Evaluated on the public AppWorld benchmark, it keeps the small model on-schema while making correct recovery decisions — a cheap sidecar pattern for production agents that wander.