Can agents learn from failure without updating their weights?
Explores whether language models can improve through trial and error by storing reflections in episodic memory rather than fine-tuning. This matters because it suggests a fundamentally different path to agent adaptation.
Continuously consolidated agent memory follows an inverted-U utility curve LLM agents depend on raw experience but disregard condensed experience External critics improve revision while internal self-assessment degrades it Models degrade sharply when prior errors contaminate their context history Self-revision amplifies confidence in wrong answers, but diverse debate prevents it
Why do Generation-Then-Comprehension and AI Delegation produce opposite learning outcomes? Which AI interaction patterns preserve learning while which ones degrade skill formation? Can language agents be represented as optimizable computational graphs? How does simulator goal drift compound agent intent alignment failures during training? What training signals would models need to learn reciprocal common-ground construction? Can environmental scaffolding replace internal memory scaling in agent design? Do dynamic environments enable different kinds of agent-environment coevolution? Can explicit numerical signals override learned linguistic defaults in fine-tuned models? Why does visual similarity retrieval fail for embodied agents? Why does storing past judgments in memory make current evaluations worse? How do agents ground their judgments in evidence instead of pattern matching? Can continuum memory systems prevent catastrophic forgetting in neural networks?See all 122 inquiring lines on this note →