LLM Daydreaming

Source: https://gwern.net/llm-daydreaming Author: Gwern Date: 2025-07-15

Summary

Gwern’s speculative piece on what it would mean for LLMs to “daydream” — process information between conversations, consolidate memories, and improve without explicit training. Analogous to how human sleep consolidates memories.

Key Claims

• Current LLMs are stateless between queries — all “thinking” happens during inference. No background processing.
• Daydreaming proposal: during idle compute time, LLMs run inference on their own outputs, consolidating and restructuring knowledge.
• Connection to memory consolidation: human sleep moves short-term memories to long-term storage. LLM daydreaming would do something analogous to the context window.
• Technical challenges: daydreaming requires the model to update its own weights — currently this only happens during training, not inference.
• The value: models could improve from experience between sessions without explicit human feedback.

Concepts

• Agent Memory — daydreaming as memory consolidation mechanism
• Continual Learning — daydreaming as a continual learning approach
• Autonomous Research — self-directed learning during idle time