Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents

AI summary

Overview

Research area: AI agent infrastructure — specifically, how LLM agents decide what to read from external artifacts (webpages, documents, repositories, tool responses) before each decision.

Technical level: Intermediate. The paper is readable with a general understanding of LLM agents and retrieval, but it introduces a formal protocol, a cost model, and a bounded-exposure proposition.

Scope (one sentence): The paper presents SparseRead, a model-agnostic reading control layer that makes agents read only the evidence their current reasoning step needs, and evaluates it across six models, five scenarios, and three agent frameworks.

What This Paper Is About

Agents typically receive an entire webpage, file, or tool response when their next decision requires only a few passages, and that excess content is replayed across many later model calls. The paper names this mismatch "over-reading" and shows it raises token cost and latency while sometimes diluting relevant evidence and hurting task quality. The goal is a reusable control layer, working with unchanged agent models, that decides which content is admitted to context, how it is located, and when it is cheaper to fall back to a full view.

Key Contributions

1. Read Intent Protocol. A cooperative, stateful protocol that converts broad artifact access into a bounded reading process. Each access carries a structured Read Intent; each read returns an EvidencePack with source-anchored evidence, the scope examined, the amount exposed, and per-slot coverage. The agent — not the protocol — decides sufficiency and stopping. The protocol also tracks artifact versions via SHA-256 fingerprint, invalidating stale evidence when an artifact changes.
2. Adaptive Reader Backends. A unified Reader interface that separates how evidence is located or processed from what is exposed to the model. A Reader configuration is a registered Reader plus a location/processing method plus an exposure form; the layer selects the lowest-cost configuration that clears a calibrated reliability threshold, falling back to a full view when none does.
3. Budgeted Read Gate. An inference-time controller that decides when sparse reading is worth attempting. It bounds cumulative raw exposure per episode by the length of a faithful full view and retains an explicit full-view fallback, with Proposition 4.1 guaranteeing total unweighted artifact exposure of at most 2L per episode.
4. Evaluation. Comparison against strong native agents, static RAG, agentic RAG, ACON, and observation masking across five scenarios and six models, plus ablation of each component and a portability study across three agent frameworks.

Main Findings

• Evidence is measurably sparse. For long-document QA, annotated supporting evidence covers a median of 3.17% of source tokens per question and 2.89% across five questions from one document in LooGLE; the share is about 5.8% for analyzed answerable Qasper questions with evidence annotations.
• Over-reading compounds. Content exposed at step i may persist across R_i later calls, giving a replay-weighted exposure of the sum of |e_i|·R_i. In paired high-sparsity runs, full-read agents spend up to 17.3× more tokens than SparseRead without improving task score.
• Broad efficiency gains across 30 settings. Across six models and five scenarios, SparseRead reduces tokens by 15.7–92.9% and wall time by 23.0–89.0%, with median reductions of 61.4% and 53.7%; score is preserved or improved in 29 settings, with a regression for GLM-5.1 on AI for Science.
• Scenario-specific results. Long-context tasks save 76.5–92.9% of tokens and 67.1–89.0% of time; multi-file diagnosis and structured analysis preserve or improve score for every model with token reductions of 26.8–74.5% and 15.7–64.1%; AI for Science saves 82.7–89.1% of tokens and 28.3–63.0% of time with score preserved or improved for five models.
• Web deep search improves quality while cutting cost. All six models gain 5–14 percentage points while saving 31.0–43.0% of tokens and 23.0–37.0% of task completion time.
• Strong models still benefit. Claude Opus 5 saves 41.0–89.0% of tokens and 35.0–78.1% of time across all five scenarios.
• Best single-setting result. On NanoBot with Kimi-K2.5, SparseRead cuts tokens by 92.9% with a score change of +0.598 and an 85.4% time saving.
• Baseline comparison. In 15 settings across three scenarios and five models, SparseRead preserves or improves score in every setting, saving 27.0–92.9% of tokens and 23.0–88.9% of time; it achieves both the fewest tokens and least time for all five models on multi-file diagnosis, with score gains of 17.8–34.5 percentage points. RAG Static is faster in all five Web settings but scores below Naive in nine of 15 settings; RAG Agent loses score in eight settings and exceeds Naive's time in eight; ACON loses score in five settings and increases both costs in six; observation masking loses score in five.
• Portability holds. Across 18 settings spanning NanoBot v0.2.0, OpenCode v1.17.14, and OpenClaw v2026.6.11, token use and wall time are lower in all 18; score is preserved or improved in 15, with three decreases ranging from 0.013 to 0.031. Median token reductions are 65.0%, 65.2%, and 30.6%; time savings are 59.3%, 63.9%, and 28.8%.
• Intent tracking matters. Combining session and access intents reduces repeated reads from 4.8 to 1.5 per task (68.8%) and untraceable evidence from 28% to 8%; verification checks that resolve an evidence gap rise from 37% to 89%, and answers failing from missing evidence fall from 25% to 5%.
• Location method quality varies. On 47 LooGLE documents and 741 questions (hit = at least 80% annotated evidence coverage), lexical matching achieves 39.5% at 780 characters, BM25 60.1% at 1,497, and Hybrid 65.3% at 1,490; expanding BM25 windows raises exposure to 2,106 characters while lowering hit rate to 56.7%.
• Gate timing matters. Checking budget before delivery keeps all six exposure ratios within 1.62–1.93 of the full view, whereas checking after delivery lets three upgrades reach 2.10–2.33 times the full view.
• Sparse reading has a regime boundary. Forced sparse reading increases tokens by 309% on global statistics and 210% on a rule-coding task, while it reduces tokens by 80–88% on high-sparsity QA and audit tasks.

Methodology in Plain English

The researchers first measured the problem: they annotated how much of a source actually supports an answer, and measured how much content native agents expose, including content carried forward into later calls. They then built a control layer that sits between the agent and the external resource, leaving the model untouched.

The layer works in three moves. First, the agent states what it needs for its next decision in a structured form, resolved against the broader session objective; vague requests get a bounded object preview instead. Second, the layer picks a Reader configuration for that artifact type and intent — for example hybrid retrieval for long text, column queries for tables, key-path lookup for JSON or logs, or an exact computation over a whole table that returns only the result, row count, and provenance. Third, a gate compares the size of the next evidence batch against a per-episode budget set to the length of a faithful full view, and withholds the batch in favor of the full view if it cannot fit intact, retiring earlier sparse observations so they are not replayed.

Evaluation used five scenarios (long-context reading, multi-file audit and diagnosis, structured analysis, Web deep search, AI for Science) built from LooGLE, PinchBench, Workspace-Bench-Lite, QwenClawBench, SpreadsheetBench Verified, FRAMES, BrowseComp-Plus, and LAMMPS script-generation tasks adapted from MDGYM. Models were Qwen3.6-Plus and DeepSeek-V4-Flash as medium models, DeepSeek-V4-Pro, GLM-5.1, and Kimi-K2.5 as leading open models, and Claude Opus 5 as a closed model. Costs were measured end to end, including retrieval and reranking for RAG baselines and auxiliary model calls for ACON.

Why This Matters

The paper reframes context reduction as an access-control problem rather than a post-hoc compression problem: it decides what enters the context, not how to shrink what already entered. It also shows that more context is not a free substitute for better reading, since full-read agents consumed up to 17.3× more tokens without quality gains — a finding relevant to both agent architecture research and serving economics.

Real-world applications:

• Web research and deep-search assistants that iterate over many pages, where the paper reports 31.0–43.0% token savings and 5–14 percentage point quality gains.
• Repository and multi-file code auditing, where evidence must be reconciled across many workspace files and token reductions reach 74.5%.
• Spreadsheet and structured-data analysis, where a Reader can scan an entire table yet expose only a result, row count, and source.
• Scientific workflow assistance, such as LAMMPS script generation grounded in scientific script libraries, where 82.7–89.1% of tokens are saved.

Industry relevance: the design requires no model retraining and no separate generative compression model, and it was integrated into three existing frameworks (NanoBot, OpenCode, OpenClaw) through thin adapters. Because the exposure budget and intent validation are enforced in the core runtime rather than by model compliance, the guarantees do not depend on a particular model's behavior — which makes the approach practical for serving systems that must bound cost per episode.

Future Directions

• Extend the guarantee beyond raw exposure. Proposition 4.1 bounds unweighted artifact exposure per episode at 2L, but the paper states it is not a bound on replay-weighted context, billed tokens, task latency, or performance against an oracle, and it applies per episode rather than across a session. Tightening these is a natural next step.
• Improve the cost/reliability calibration. The Reader ablation compares fixed location methods and explicitly does not isolate the incremental benefit of the configuration selector; better calibration of reliability and follow-up-read cost estimates remains open.
• Reduce residual regressions. A score regression for GLM-5.1 on AI for Science and three score decreases of 0.013–0.031 in the cross-framework study point to cases where intent handling or fallback policy could be sharpened.
• Broaden the Reader ecosystem. The interface is designed so retrieval and compression methods (the paper names SWE-Pruner and Squeez as possible compression backends) and new artifact types can register without changing the agent-facing action set of scout, focus, collect, refine, and verify.

Target Audience

Agent-infrastructure and applied-LLM engineers building tool-using agents that read external content; researchers working on context management, retrieval, and observation pruning; and platform teams who need predictable per-task token and latency budgets. Readers without a background in agent loops will find the ablation and cost numbers accessible, while the protocol and gate sections reward readers interested in system design.

Authors’ 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…