Paper review · November 13, 2026
Read less, solve more: token-efficient sparse reading for AI agents (SparseRead)
Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents
Adding a reading layer that makes agents read only the evidence they need from web pages, documents and tool responses cut tokens by up to 92.9% while maintaining quality.
Problem
Web agents often receive an entire page, file or tool result even when the next decision needs only a few paragraphs. The authors call this 'over-reading'. Over-reading increases tokens and latency, and costs accumulate because content read once is fed in again on every later model call. Relevant evidence can also get buried, leading to wrong tool calls or failures to terminate. In the authors' measurements, full-reading agents used up to 17.3 times more tokens with no quality gain.
Method
Without changing the model, a read-control layer sits between the agent and external sources. (1) Read Intent Protocol: each time the agent accesses a source, it states what it needs as a Read Intent and receives an EvidencePack containing source-attributed evidence and a record of 'which items have evidence and which are still missing'. Judging sufficiency and deciding when to stop are left to the agent. (2) Adaptive Reader Backends: among Readers for each source type (text, PDF, tables, JSON, etc.), it picks the lowest-cost configuration among those that pass a calibrated confidence threshold. (3) Budgeted Read Gate: each episode has a budget for raw-text exposure, and it switches to a full view when partial reading is inefficient.
Results
- Across six models, including Claude Opus 5, and five scenarios, it cut total tokens by up to 92.9% and run time by up to 89.0%, and maintained or improved task quality in 97% of settings.
- On web deep-search tasks, it cut tokens by 31.0–43.0% and task completion time by 23.0–37.0%.
- After fixing how Read Intents are recorded, total tokens dropped 2.1–6.3x on long-book QA, multi-file auditing, PDF report QA and web search.
- Conversely, forcing sparse reading on full-statistics tasks and rule-based coding tasks increased tokens by 309% and 210%, respectively. The authors presented this result as evidence that a mechanism to fall back to a full view is needed.
- The authors reported confirming the effect of the three design components through ablations and showing portability by attaching the layer to three agent frameworks.
Limitations
On tasks that require seeing the whole document, sparse reading actually increases cost, so performance depends on the Read Gate's switching decisions. Calibrating Reader confidence requires development request data with evidence locations marked. Discarding earlier partial-read results when switching to a full view may prevent prefix cache reuse. The input text was cut off partway, so detailed per-scenario numbers and comparisons with baselines (such as RAG) could not be checked.
Why it matters
Tokens spent reading external sources make up a large share of agent operating costs. This paper provides evidence that cost and latency can be cut substantially through the reading interface alone, without changing the model. It is a reason to reconsider 'return everything' defaults in agent frameworks and tool design.
Key terms
- Over-reading
- The practice of putting an entire page or file into the model's context when the next decision needs only part of the evidence. It raises token costs and buries important evidence.
- EvidencePack
- The result bundle returned by SparseRead. It contains source-attributed evidence, the range processed, the amount exposed, and whether evidence was found for each required item.
- Full-view fallback
- A mechanism that falls back to showing the full source text when partial reading is inefficient or fails to pass the confidence threshold.
Original
Authors: Zedong Liu, Jiaan Wu, Xinyang Ma, Le Xu, Kai Wang, Yuanchao Hu, Dingwen Tao, Guangming Tan · Posted 2026-08-23 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.