Researchers Introduce Context Language Models for Native Context Management

Researchers have introduced Context Language Models (CLMs), a new class of language models that natively manage their own context rather than relying on external harnesses. By treating the context window as a file that the model can update without restriction, CLMs learn to prioritize critical information and naturally extend to multi-agent systems where multiple contexts coexist.

Performance and Efficiency

Building CLMs zero-shot with existing models demonstrates significant improvements over current state-of-the-art context management strategies across various tasks:

• BrowseComp-Plus: Achieved 11.4% higher accuracy with 21.5% fewer FLOPs.
• EdgeBench (12-hour): Delivered 5% higher scores with 59% fewer FLOPs.
• Multi-repository agent-swarm (24-hour): Showed 65% greater improvement with the same compute resources.

Learning and Optimization

Shifting context management from external control to intrinsic model behavior enables both in-context and parametric learning of management strategies. The paper details two key optimization methods:

• Skill-Optimization Loop: CLMs can be steered with natural-language instructions, improving held-out accuracy by up to 35.9 points on context-management tasks while reducing compute.
• Online Reinforcement Learning: Applying this method to Qwen3.5-9B improved performance on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs.

Serving Infrastructure

To support efficient deployment, the authors co-designed Suffix Cache Reuse for CLM serving. This technique reduces server-side compute by 35% relative to standard SGLang at matched performance levels.