Meta paper: agents editing their own context gain 11.4% accuracy with 21.5% fewer FLOPs
omarsar0 · x · 2026-10-02
A new paper from Meta and colleagues proposes Context Language Models (CLM): the model's live interaction context is exposed as a read-write file that the model edits via Bash, deciding natively what to keep, rewrite, or remove.
- Unlike RLMs, which place large inputs in external variables processed recursively, CLMs let the model directly edit its own live context, including information accumulated during execution.
- Applied zero-shot, CLMs achieve 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus versus existing context-management strategies.
The core theme: giving agents write access to their own context may improve both quality and efficiency.
More from coding & agent
- Blogger builds Elon Musk interview archive site in 1 hour with Tencent Hy4 flash, open-sources two Skills — vista8 · 2026-11-16
- VibePrevisBench: 242-task benchmark tests coding agents on 4D Blender scene generation — Kling-Keye · 2026-11-16
- raylib author threatens to block users over AI-generated GitHub issues — yacineMTB · 2026-11-16
- Dev open-sources local Apple Health MCP server that feeds workouts to AI agents and pushes plans back to Apple Watch — PrinzPwnage · 2026-11-16
- Coding agents crack a 1994 Jeep's ECU via a $3 Arduino and start controlling it — LeviTurk · 2026-11-16
- 'My favorite PRs': raylib author hails a change that adds features by deleting 100+ lines — yacineMTB · 2026-11-16