University of Washington and Meta Superintelligence Labs · Self-Managed Context

Context Language Models

Language models that manage their own context by treating it as a file they can rewrite, an idea that extends to multi-agent systems where each agent's context is a file. Built zero-shot from existing models, CLMs report 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus and 5% higher scores with 59% fewer FLOPs over a 12-hour EdgeBench run. Online reinforcement learning lifts Qwen3.5-9B by 47.6% on BrowseComp-Plus while using 12% fewer FLOPs, and a co-designed Suffix Cache Reuse cuts serving compute by 35% against SGLang, all self-reported. Posted 29 September; code, a Harbor harness and training modules are at facebookresearch/context-language-models, still active this week.