SparseRead reports up to 92.9% token drop for AI agents
TL;DR
- SparseRead reports up to 92.9% total-token and 89.0% wall-time reductions across six models including Claude Opus 5, with no model modification.
- On Web deep-search tasks, the system reports a narrower 31.0 to 43.0% token cut and 23.0 to 37.0% task-completion-time cut.
- The paper credits three components (Read Intent Protocol, Adaptive Reader Backends, Budgeted Read Gate) and reports integration with three agent frameworks.
SparseRead, a reusable "reading control layer" that governs what external content an AI agent consumes, reduced total tokens by "up to 92.9%" and wall time by up to 89.0% across six models including Claude Opus 5, according to a new preprint from Zedong Liu and colleagues.
The team's paper, titled "Read Less, Solve More," reports that SparseRead "governs content admission without modifying the agent model" and preserved or improved task quality "in 97% settings" across five scenarios. On narrower Web deep-search tasks the headline figure collapses: tokens drop 31.0 to 43.0%, and task completion time falls 23.0 to 37.0%.
The authors credit three pieces: a Read Intent Protocol that tracks "evolving needs and source coverage," Adaptive Reader Backends that select "how evidence is located and exposed," and a Budgeted Read Gate that bounds "cumulative raw exposure per episode with faithful full-view fallback." Ablations are reported and integration is demonstrated with three agent frameworks, but the abstract publishes no per-model or per-scenario breakdown.
Shared on Bluesky by 1 AI expert
Originally reported by paper
Read the original article →Original headline: SparseRead Cuts AI Agent Token Volume 92.9% With No Model Changes—Six Frontiers, v2 This Week