Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents
Zedong Liu, Jiaan Wu, Xinyang Ma, Le Xu, Kai Wang, Yuanchao Hu, Dingwen Tao, Guangming Tan
University of Chinese Academy of Sciences · Institute of Computing Technology, Chinese Academy of Sciences · Songshan Lake Materials Laboratory
Published November 7, 2026 · on the feed November 12, 2026 · ▲ 19 · arXiv 2608.22237
The abstract
Web-facing AI agents repeatedly access webpages, documents, repositories, and tool responses, yet reading interfaces often expose entire objects when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute relevant evidence. We present SparseRead, a reusable reading control layer that governs content admission without modifying the agent model. Its Read Intent Protocol tracks evolving needs and source coverage; Adaptive Reader Backends select how evidence is located and exposed; and the Budgeted Read Gate bounds cumulative raw exposure per episode with faithful full-view fallback. A unified Reader interface supports interchangeable evidence acquisition across heterogeneous resources. Across six models, including Claude Opus 5, and five scenarios, SparseRead reduces total tokens by up to 92.9% and wall time by up to 89.0%, preserving or improving task quality in 97% settings. On Web deep-search tasks, it reduces tokens by 31.0--43.0% and task completion time by 23.0--37.0%. Ablations validate the three designs, while integration with three agent frameworks demonstrates portability.
The Brief
1:56A short podcast summary of the paper. Written and voiced by AI.
Transcript
Read Less, Solve More. Token-Efficient Sparse Reading of External Content for AI Agents. By Zedong Liu and seven others, from the University of Chinese Academy of Sciences, the Institute of Computing Technology of the Chinese Academy of Sciences, and one other institution. Published on the seventh of November.
An agent working on the web keeps opening things: webpages, documents, repositories, responses from tools. The reading interfaces it uses often hand over the entire object, even when only a few scattered pieces of evidence are needed. That over-reading costs tokens and latency, and it can dilute the evidence that actually matters by burying it in everything else.
SparseRead is a reusable reading control layer. It governs what content is admitted into the agent's view, and it does this without modifying the agent model at all. The agent stays the same. What changes is how much of the outside world is let in.
Three designs carry it. A Read Intent Protocol keeps track of the agent's evolving needs and of how much of each source has been covered. Adaptive Reader Backends decide how evidence is located and how it is exposed. A Budgeted Read Gate puts a bound on the cumulative raw exposure in an episode, with a faithful full-view fallback when the whole object is needed. Under all three sits a unified Reader interface, so evidence can be gathered the same way across very different kinds of resource.
Tests covered six models, Claude Opus 5 among them, and five scenarios. SparseRead cuts total tokens by up to ninety-two point nine percent and wall time by up to eighty-nine percent. Task quality is preserved or improved in ninety-seven percent of settings.
On Web deep-search tasks, tokens fall by thirty-one to forty-three percent, and task completion time by twenty-three to thirty-seven percent. Ablations support each of the three designs, and integration with three agent frameworks shows the layer can be carried from one to another.
Instead of a smarter reader, this is a stricter doorkeeper. By deciding what the agent gets to see, and how much, the work can be done with a fraction of the reading, and the quality of the answers mostly holds or improves.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Agents read far too much. Every extra page costs.
Your web agent keeps opening webpages, documents, repositories and tool outputs. Each time, the interface hands it the whole object.
The road ahead
- 1 · The problem ~7 min Web agents get handed whole pages, files and tool outputs when they need only a few scattered pieces, and that waste piles up.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min SparseRead is a layer between the agent and its sources that controls how much content gets in, without changing the agent model.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each read goes through tracking the need, picking a reader, checking the exposure budget, and falling back to a full view when needed.
- 6 · The experiment ~6 min You’ll guess: While SparseRead cut tokens and wall time, in what share of settings did task quality hold up or improve?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.