about
Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories (arxiv.org)
5 points by olalonde 17 days ago | hide | past | pdf | 1 comment on HN

In plain words: A model gets a sleep cycle: it folds short-term memories into permanent weights by having a smaller version teach a bigger one, then dreams up its own practice problems to rehearse. On long-horizon and continual tasks, this beat the same models without sleep.

Abstract

The past few decades have witnessed significant advances in the design of machine learning algorithms, from early studies on task-specific shallow models to more general deep Large Language Models (LLMs). Despite showing promising results in tasks that require instant prediction or in-context learning, existing models lack the ability to continually learn and effectively transfer their temporal in-context knowledge to their long-term parameters. Inspired by human learning process, we introduce a ''Sleep'' paradigm that allows the models to continually learn, distill their short-term fragile memories into stable long-term knowledge with replay, and recursively improve themselves with ''Dreaming'' process. In more detail, sleep consists of two stages: (1) Memory Consolidation: an upward distillation process, called Knowledge Seeding, where the memories of a smaller-self are distilled into a larger network to provide more capacity while preserving the knowledge. As a proof of concept, we present a new Generalized Distillation process for {Knowledge Seeding} (i.e., the combination of on-policy distillation with Reinforcement Learning (RL)-based imitation learning); (2) Dreaming: a self-improvement phase, where the model uses RL to generate a curriculum of synthetic data to rehearse new knowledge and refine existing capabilities without human supervision. Our experiments on long-horizon, continual learning, knowledge incorporation, and few-shot generalization tasks support the importance of the sleep stage.

Ali Behrouz, Farnoosh Hashemi, Adel Javanmard, Vahab Mirrokni
arXiv:2606.03979 · cs.LG, cs.AI · submitted Jun 2, 2026 · updated Jul 10, 2026
abstract · pdf · html · A version of this work has been publicly available from September 2025 on OpenReview

add comment on HN
Also discussed: Jun 2026 (1 point, 0 comments)

Worth knowing this is the Titans / Nested Learning line from Google Research (Behrouz, Mirrokni), so "sleep" here is a concrete two-stage recipe: distil the in-context memories into a larger copy of the network ("knowledge seeding"), then let the model generate its own curriculum with RL ("dreaming"). The abstract names the task families but not a single number, so the question for the paper is how much of the continual-learning gain survives against the plain baseline of continued training on the same replayed data. The brain analogy is the weakest part; the replay-then-distil recipe may be useful regardless.