Context Language Models
The paper proposes models that edit their own context as a working file, cutting compute while improving long-task results.
The authors say Context Language Models let the model decide what to keep, rewrite, or discard in context instead of leaving that control to an external harness. In zero-shot tests, they report higher accuracy with fewer FLOPs on BrowseComp-Plus, EdgeBench, and a 24-hour multi-repository agent-swarm task. They also show context-management behavior can be improved through natural-language skill optimization and online reinforcement learning. A serving method called Suffix Cache Reuse cuts server-side compute further at matched performance. HF Daily Papers' note
The authors say Context Language Models let the model decide what to keep, rewrite, or discard in context instead of leaving that control to an external harness. In zero-shot tests, they report higher accuracy with fewer FLOPs on BrowseComp-Plus, EdgeBench, and a 24-hour multi-repository agent-swarm task. They also show context-management behavior can be improved through natural-language skill optimization and online reinforcement learning. A serving method called Suffix Cache Reuse cuts server-side compute further at matched performance. HF Daily Papers' note
score 6