Learning Stateful Predictive Knowledge From Experience

Short summary

Stateful Knowledge Learning (SKL) shifts LLM agents from trajectory-level reflection to maintaining explicit, declarative predictive assessments anchored to state. Two algorithms—SKL-SD via self-distillation and SKL-RL via reinforcement learning—train agents to extract state-grounded predictive knowledge. Experiments on WebShop, ScienceWorld, and ChessPuzzles show SKL significantly outperforms reflection-based training paradigms.

• •SKL replaces trajectory-level reflection with state-anchored predictive knowledge
• •Two algorithms: SKL-SD (self-distillation) and SKL-RL (reinforcement learning)
• •Outperforms reflection-based paradigms on WebShop, ScienceWorld, and ChessPuzzles

Generated with AI, which can make mistakes.