SparseRead cuts AI agent tokens by up to 92.9% with no model retraining
techNmak · x · 2026-11-15
Researchers at the Chinese Academy of Sciences present SparseRead, a reusable reading-control layer that governs what enters an agent's context before it gets there — no model retraining needed. It combines a Read Intent Protocol (tracking evidence needs), Adaptive Reader Backends (locating evidence across PDFs, repos, structured data), and a Budgeted Read Gate (bounding cumulative raw exposure with full-view fallback). Across six models including Claude Opus 5 and five workloads, it cuts tokens by up to 92.9% and wall time by up to 89% (medians 61.4%/53.7%) while preserving or improving task quality. Key insight: external observations account for 60–80% of agent trajectory tokens, and over-reading dilutes relevant evidence — control context admission before resorting to compression or KV caching. Paper and code are public.
More from coding & agent
- AI Tinkerers hosted nearly 300 events in 90 days with 500+ project demos — gabe_ragland · 2026-11-15
- Hermes Agent ships official Home Assistant plugin to control your whole house via chat — brianrkelly · 2026-11-15
- Simple prompt tweak: short bullet-point emoji replies make Grok agent work easier — jonathan_wilke · 2026-11-15
- Dev reports Unreal's built-in AI assistant closes the editor for rebuilds with no way to restart it — FanaHOVA · 2026-11-15
- Claude deletes professor's year of lecture notes while asked to renumber pages — petrusenko_max · 2026-11-15
- A Model That Fails 42.9% of Tasks Can Still Teach an Agent to 82.9% — rohanpaul_ai · 2026-11-15