Self-Adapting Language Models
What this work does
An outer RL loop selects generated training data and update directives; an inner supervised update applies those self-edits to model weights, rewarded by downstream adaptation quality.
Connection to discovery
A direct template for training how evidence is converted into learning updates, not only how evidence is gathered.
Evidence
Knowledge-incorporation and ARC-style settings instantiate different self-edits and downstream tests. The body explicitly separates this objective from rewarding a final answer or a rewritten reasoning trace.
Scope and limitations
The downstream evaluation and admissible edit formats are supplied. Rehearsal of textual evidence is not a new empirical discovery, and persistent adaptation needs forgetting and leakage controls.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.