SparseRead cuts agent reading tokens by up to 92.9% and keeps task quality

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

cs.AI

2026-08-23

SparseRead, a model-agnostic read gate, cuts tokens by up to 92.9% and wall time by up to 89.0% across six models and five tasks, holding or raising quality in 97% of settings.

What problem this solves

Web agents receive webpages, PDFs, repositories, and tool payloads as whole objects. The next decision often needs a few passages. External observations already account for 60%-80% of tokens in a trajectory, and text admitted early is replayed on later calls.

Annotated evidence is a thin slice. On LooGLE the median supporting evidence is 3.17% of source tokens per question, and 2.89% across five questions from one document. Answerable Qasper questions with evidence annotations sit near 5.8%. Those figures measure sparsity, not the reading required to locate a passage. A Reader can scan a whole table, sum a column, and expose only the result, the row count, and the source.

On paired high-sparsity runs, an uncontrolled agent spends up to 17.3 times as many tokens without a higher score. Extra text dilutes evidence: wrong tool calls, confused sources, and failures to stop. Long context and KV-cache compression cheapen text that is already inside the window. ACON and observation masking shrink or hide observations after admission. Retrieval can fetch passages. Missing is one admission contract for the current need, explicit coverage, a cap on raw exposure, and a faithful full-view fallback.

Method

SparseRead is a reading layer between the agent and external artifacts. Weights stay frozen, and there is no extra compression model. The control plane handles intent, routing, and admission. The evidence plane is a set of swappable Readers. The contract on both sides is Read Intent in, EvidencePack out.

Each access submits a structured intent for the decision at hand. It outranks a session intent drawn from the opening user request; the session intent only fills blanks. A vague intent receives a bounded preview. Five actions stay fixed: scout, focus, collect, refine, and verify. Adding a Reader does not add actions. The pack carries source anchors, scope processed, amount exposed, and per-slot coverage. Coverage marks which needs have support and which are still open. The agent decides sufficiency. The protocol never declares the task done.

Accesses to one artifact version under one objective form an episode, with one evidence record and one exposure budget. A SHA-256 fingerprint identifies the version. A change invalidates stored evidence for that artifact.

Readers cover text and PDF, tables and JSON, multi-file collections, and script libraries. Locating evidence and choosing an exposure form are separate. Text can use lexical match, BM25 windows, or hybrid retrieval. A table Reader can project columns, or compute over every row and return only the result and its provenance. Configurations must clear a reliability threshold; among those that do, the layer picks the lowest estimated cost. Reliability is calibrated on annotated development requests by whether the evidence was actually obtained, not by mean recall. If none qualify, the full view is returned and the threshold stays put.

The gate sets L to the cheapest faithful full view and freezes L for the episode. Cumulative sparse exposure E is measured in characters and may not exceed L. Small objects, requests that need a complete model-visible view, and missing Readers take the native full read. A batch that does not fit the remaining budget is withheld, and the full view is delivered intact. Evidence is not cut in a way that drops its anchors. Unweighted exposure for one episode is therefore at most 2L.

After coverage is recorded, total tokens on long-book QA, multi-file audit, PDF reports, and web search fall by 2.1x to 6.3x. High-sparsity QA and audit drop 80%-88% of tokens. Forced sparse reading raises tokens by 309% on global statistics and by 210% on rule coding, so the gate switches before the next batch exceeds L. NanoBot v0.2.0 registers the tools directly. OpenCode v1.17.14 and OpenClaw v2026.6.11 attach through thin adapters. SKILL.md only advises; the core enforces the budget.

Results

NanoBot runs six models on five scenarios, 30 settings: Qwen3.6-Plus, DeepSeek-V4-Flash, DeepSeek-V4-Pro, GLM-5.1, Kimi-K2.5, and Claude Opus 5. The baseline is the unmodified Naive agent. Costs are end to end and include retrieval.

SliceTokensWall timeScore
All 30down 15.7%-92.9%, median 61.4%down 23.0%-89.0%, median 53.7%held or up in 29/30
Long contextdown 76.5%-92.9%down 67.1%-89.0%held or up
Multi-file diagnosisdown 26.8%-74.5%within the overall rangeheld or up on all six
Structured analysisdown 15.7%-64.1%within the overall rangeheld or up on all six
Web deep searchdown 31.0%-43.0%down 23.0%-37.0%up 5-14 points on all six
Science scriptsdown 82.7%-89.1%down 28.3%-63.0%held or up on five

The only loss on this grid is GLM-5.1 on science scripts. The abstract's 97% is these 29 of 30. Opus 5 saves 41.0%-89.0% tokens and 35.0%-78.1% wall time across the five scenarios.

Fifteen settings compare SparseRead with static RAG, iterative RAG under the same iteration and timeout limits, ACON, and observation masking that keeps the latest ten observations. SparseRead holds or raises the score in every cell, cuts tokens 27.0%-92.9%, and cuts wall time 23.0%-88.9%. It uses the fewest tokens in six cells and the least time in nine. On multi-file diagnosis it is lowest on both costs for all five models, 17.8-34.5 points above Naive. Static RAG is faster on all five web settings, and scores below Naive in nine of the fifteen cells. Iterative RAG loses score in eight and is slower than Naive in eight. ACON loses score in five and raises both costs in six. Masking loses score in five.

On NanoBot, OpenCode, and OpenClaw together, all 18 model pairings cut tokens and time, and 15 hold or raise the score. Median token cuts are 65.0%, 65.2%, and 30.6%; median time cuts are 59.3%, 63.9%, and 28.8%. Three drops run from 0.013 to 0.031: DeepSeek-V4-Pro on NanoBot and on OpenCode, and Kimi-K2.5 on OpenCode. The NanoBot row for Kimi-K2.5 is a 92.9% token cut, a score change of +0.598, and an 85.4% time cut.

Combining session and access intent cuts repeated reads from 4.8 to 1.5 per task, untraceable evidence from 28% to 8%, checks that close a gap from 37% to 89%, and answers that fail for missing evidence from 25% to 5%. On 741 questions from 47 LooGLE documents, a hit requires 80% of the annotated evidence. Lexical match hits 39.5% at 780 characters, BM25 60.1% at 1,497, and hybrid 65.3% at 1,490. Wider BM25 windows use 2,106 characters and hit 56.7%. Checking the budget before delivery keeps exposure ratios between 1.62 and 1.93, with a full-view upgrade in 5 of 15 runs. Checking after delivery reaches 2.10-2.33 times the full view.

Why it matters

When external observations dominate trajectory tokens, the useful cut is at admission, before another expansion of the context window. Web deep search saves only about a third of tokens, and every model still gains 5-14 points. The extra material was diluting the evidence.

The same runtime attaches to three frameworks with no retraining. OpenClaw's median token cut is about half that of NanoBot and OpenCode, so the contract moves and the savings still track how eagerly the host dumps whole objects. Forcing the sparse path on low-sparsity work wastes tokens. The gate's job is to stop at L and return the full view.

Limitations

The 2L bound caps unweighted raw exposure inside one episode. It does not cap billed tokens, wall time, or replay on later calls. Prefix caching changes the replay bill on its own. Spent tokens are not returned when fallback drops earlier sparse observations, and the bound does not sum across episodes.

The 30-cell grid leans high-sparsity. Cases where forced sparse reading increases tokens stay in the motivation experiments, outside that grid. Structured analysis can save as little as 15.7%. A 92.9% cut is the top of the range, and 97% means 29 of 30 on NanoBot. Across three frameworks the score holds in 15 of 18. GLM-5.1 loses on science scripts; a missed library lookup and an early fallback are not reported separately.

The text-locator comparison does not isolate the configuration selector. A new artifact type without annotated calibration goes straight to the full view. Static RAG is faster on all five web settings. The teaser names the smaller model Qwen3.5-35B; section 6 uses Qwen3.6-Plus.

Terms

Source

Related papers

All paper explainers