A paper submitted to arXiv on 29 September proposes something simpler than most context-management research: stop treating a model’s context as an append-only transcript and start treating it as a file the model is allowed to edit.
The authors — a group including Rulin Shao, Mike Lewis, Luke Zettlemoyer and Pang Wei Koh — call the result Context Language Models, or CLMs. The context is mirrored into a storage space the model can write to. The model can append newly generated tokens, or it can use Bash to edit the file freely, and every modification is synchronised immediately back into the live context.
The result
On BrowseComp-Plus, a deep research benchmark, a CLM reaches 11.4% higher accuracy than the strongest baseline while using 21.5% fewer prefix-reuse FLOPs. Better and cheaper at the same time is the part that makes the paper interesting, because context management is usually a trade: summarise to save tokens and lose detail, or keep everything and pay for it.
Trained on deep research tasks, Qwen3.5-9B improves from 28.8% to 42.5% on the same benchmark under the CLM setup. That beats a Codex-style summary harness trained with the same recipe by 0.4 points while using 38.8% fewer FLOPs — a narrow win on accuracy and a wide one on cost.
How it is trained and steered
The authors add a success-gated efficiency advantage to stepwise GRPO, which rewards efficient trajectories among the ones that succeeded rather than rewarding brevity on its own. They also show the behaviour can be steered in plain language: instructions evolved through a standard skill-optimisation loop improved held-out accuracy by up to 35.9 points on a context-management task while reducing compute.
A serving optimisation called Suffix Cache Reuse addresses the obvious objection, which is that editing the middle of a context destroys the prefix cache that makes long-context inference affordable.
Why it matters beyond the benchmark
The design extends naturally to multi-agent systems, where several agents’ contexts coexist as separate files. That is close to how the current generation of coding agents already works in practice, using scratch files and notes, except that here the model’s own working memory is the file rather than a workaround beside it.
What to watch
Whether a frontier lab ships the mechanism rather than publishing it. The paper’s code is released under a non-commercial licence, which limits who can build on it directly.
