Research2026-09-29
Researchers introduced Context Language Models that treat context as a file. Models can make unrestricted updates to this file. This approach outperforms existing context management strategies on multiple tasks. It enables both in-context and parametric learning of context strategies. CLMs can be steered with natural-language instructions. An online reinforcement learning method improves Qwen3.5-9B performance on BrowseComp-Plus. Suffix Cache Reuse for CLM serving reduces server-side compute by 35%. The paper was submitted to arXiv on 29 September 2026.
What changed
Context management was handled by external harness control rather than model behavior.
What it unlocks
Models can now natively manage and update their own context as files.
- 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus
- 5% higher scores with 59% fewer FLOPs on EdgeBench
- 65% greater improvement on multi-repository agent-swarm task
- 35.9-point held-out accuracy gain on context-management task
Sources