# SparseRead

> Token-efficient reading for AI agents.

SparseRead is a training-free reading layer that admits only the evidence needed for the next decision, instead of dumping whole files, PDFs, or repositories into context.

- Website: https://zedong-liu.github.io/SparseReading/
- Code: https://github.com/Zedong-Liu/SparseReading
- Paper: https://arxiv.org/abs/2608.22237
- PDF: https://arxiv.org/pdf/2608.22237

## What it is

A thin adapter around one token-efficient reading runtime, not a new agent. A Read Gate decides whether to read sparse; Reader Backends admit a bounded evidence slice. It is not post-hoc summarization, KV-cache compression, or an always-on replacement for native tools.

## Supported agents

Claude Code, OpenCode, OpenClaw, and NanoBot.

## Results (paper, vs Naive full reading)

- Token volume down by up to 92.9%.
- Wall time down by up to 89.0%.
- Quality held or improved on 26 of 30 NanoBot cells.
- Keep “up to” on the maxima. Token volume is not automatically dollar cost.

## 中文

SparseRead 是面向 AI Agent 的省 token 阅读层：先拦截整文件 / PDF / 仓库的全文灌入，再按下一步决策需要的证据读取。论文 arXiv:2608.22237。