Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents
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.
Community
Hi, we’re the authors of SparseRead! We built it for a problem you might recognize: you ask an agent to explain one error, and it reads the entire log.
SparseRead lets the agent preview a file, read the relevant sections with source locations, and ask for more when it needs to. If the task requires the whole file, the agent can still read it directly.
In our evaluation across 125 tasks and six models, the largest reductions compared with full reading were 92.9% in token use and 89.0% in wall time. Task scores stayed the same or improved in 26 of 30 model–workload combinations.
Code and setup: https: https://github.com/Liuzedooong/SparseReading
Web of project: https://sparseread.pages.dev/
If you try it on your own files or repos, we’d like to hear where reading less helps and where it gets in the way.
Interesting work👍🏻 Does it work with Claude Code?
Yep, no problem! The SparseRead plugin works across major harnesses like CC, Codex, and OpenCode.
Over-reading is something every agent builder has felt, and a gate that bounds raw exposure with a full-view fallback is a safe design. Cutting tokens by up to 92.9 percent is large. In which cases did the fallback to the full view trigger most often?
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- From Delivery to Stateful Exploration: Rethinking the Index for Agentic Search (2026)
- Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses (2026)
- The Missing Complement: State-Conditioned Minimal Sufficient Evidence for Coding Agents (2026)
- CORE: Coarse-to-Fine Search with Context Management for Deep Research (2026)
- Programmatic Search Agents: Extending Agentic Search Beyond Query Reformulation (2026)
- TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents (2026)
- SkillBrowser: Selective Skill Exposure for Scalable Skill-Using Agents (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Hi authors, thanks for the work. A question on Table 1: the NanoBot rows for DeepSeek-Flash and DeepSeek-Pro do not match any of their five scenarios in Figure 6. Which workload does Table 1 use? Details in this GitHub issue: https://github.com/Liuzedooong/SparseReading/issues/1
The plugin fails to load in OpenCode. Investigation with Qwen showed that the plugin is probably written for v1 API, whereas the current OpenCode is v2.2.6, Idk why would you these days even use v1?
Get this paper in your agent:
hf papers read 2608.22237 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper