Meta, MIT and UW unveil Context Language Models that manage their own context

Researchers from Meta, MIT and the University of Washington introduced Context Language Models (CLMs), which let a model edit its own context as a file instead of relying on external summarization and retrieval. Zero-shot CLMs gained 11.4% accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, while RL lifted Qwen3.5-9B from 28.8% to 42.5% with 12% fewer FLOPs.

• Zero-shot CLM: +11.4% accuracy and 21.5% fewer FLOPs on BrowseComp-Plus
• In-context learning improved ContextBench by up to 35.9 points at lower compute
• RL raised Qwen3.5-9B on BrowseComp-Plus from 28.8% to 42.5%
• Risks include losing key data and new prompt-injection channels