Zedong Liu, Jiaan Wu, Xinyang Ma, Le Xu, Kai Wang, Yuanchao Hu, Dingwen Tao, Guangming Tan
7 min
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.
Alex: So it distinguishes the overall research question from the particular thing the agent needs to check now. What comes back from a successful read?
Sam: An EvidencePack: evidence linked to its source, together with records of what was processed and exposed. It also records which requested pieces have support and which remain unresolved. Crucially, that coverage record doesn't declare the evidence sufficient. The agent still decides whether to answer, refine, or verify.
Alex: That distinction seems essential. Finding something for every requested item doesn't necessarily mean you've found enough to answer reliably. How does the system handle different kinds of sources?
Sam: Replaceable Readers handle documents, structured data, or collections of files. A document Reader might retrieve passages. A structured-data Reader might select columns or perform an exact calculation. They return evidence through the same interface, without requiring a change to the agent model.
Alex: A calculation could require the entire source, though. Would that defeat the point of reading sparsely?
Sam: Here's the key distinction: processing a source isn't the same as showing it to the model. The paper's table example scans every row to calculate a sum. The model receives only the result, the row count, and its source. Complete processing can still mean very little model exposure.
Alex: But repeated little reads could eventually cost more than opening the whole thing. How do they prevent selective access becoming an expensive detour?
Sam: That's the Budgeted Read Gate, a controller that limits cumulative exposure. It sets the budget using the size of a faithful complete view. Before returning another batch, it checks whether that batch fits. If not, it withholds the batch and supplies the complete view instead.
Alex: Does that mean the system guarantees lower cost than ordinary reading?
Sam: No. The guarantee is at most twice that complete view's raw exposure within one reading episode, excluding bounded previews and protocol metadata. Sparse evidence can use up the budget, and fallback can add a complete view. This doesn't bound billed tokens, execution time, answer quality, or the whole session.
Alex: So the guarantee limits one particular downside. The practical savings still have to come from finishing before that budget is exhausted. What did the experiments show?
Sam: The main evaluation crossed six models with five task scenarios. Those covered long documents, multi-file diagnosis, structured analysis, web search, and scientific script generation. The comparison was the unmodified agent framework, which could already search and make partial reads using its native tools.
Alex: That's a stronger comparison than forcing the baseline to read everything. How often did saving text actually preserve the task result?
Sam: Task scores were preserved or improved in twenty-nine of thirty model-and-scenario settings. The exception was one model on scientific script generation. Median completion time fell by about fifty-four percent relative to the unmodified agent.
Alex: Those are broad results, but the task mix matters. Did this work beyond large, static documents, where selective reading seems especially well suited?
Sam: On web deep search, every tested model improved its score while using fewer tokens and finishing faster. The authors also integrated the same protocol and Readers into different agent frameworks through thin adapters. Costs fell throughout that portability experiment, although some task scores decreased.
Alex: And against retrieval or compression specifically? Otherwise we wouldn't know whether the new control layer earns its complexity.
Sam: In that comparison, SparseRead preserved or improved scores throughout, but wasn't always cheapest or fastest. Fixed, one-shot retrieval was faster in every web setting tested, while losing score in many settings overall. Iterative retrieval and the compression approaches also had quality or cost regressions. That supports the combined design, not a claim that retrieval itself is inferior.
Alex: How far can they attribute those gains to the individual parts, rather than the whole package?
Sam: Removing intent information worsened repeated reading and evidence diagnostics. Text-retrieval tests also showed that the evidence-location method matters. But the authors explicitly say that comparison doesn't isolate the benefit of their adaptive selector over fixed configurations. And source links establish provenance, not relevance or sufficiency.
Alex: So this is encouraging engineering evidence, not a promise that sparse reading is always safer. Who should read the full document, and where should they start?
Sam: Researchers building document-heavy agents should start with the Read Intent Protocol and Budgeted Read Gate sections. Then read the retrieval and context-reduction comparison, followed by the component ablations. Look closely at what gets tracked, when fallback happens, and which guarantees concern exposure rather than actual cost.
Alex: And for someone who uses these systems but won't implement a reading layer?
Sam: Remember this: a system can process the whole source without making the model read the whole source. The useful question isn't simply how much context a model can hold. It's what evidence its next decision actually needs.
Alex: That's a useful question to keep asking.