Source-linked AI summary
Read Less, Solve More: Token-Efficient Sparse Reading for AI Agents
Zedong Liu, Jiaan Wu, Xinyang Ma, Le Xu, Kai Wang, Yuanchao Hu, Dingwen Tao, Guangming Tan
TL;DR
Long-horizon agents often expose entire external artifacts even when only sparse evidence is needed, creating token, latency, and evidence-dilution costs. SparseRead controls artifact admission before broad content reaches context through a stateful protocol, extensible Reader Backends, and a regime-aware Read Gate. Across six models and five scenarios, it reduces token use by up to 92.9% and job time by up to 89.0% while preserving or improving task quality.
Problem
Long-horizon agents repeatedly process external artifacts, but current interfaces expose broad content even when decisions require only bounded evidence slices.
Method
SparseRead combines pre-context artifact admission control with a stateful reading protocol, extensible Reader Backends, and an inference-time Read Gate.
Results
Across six models and five scenarios, SparseRead reduces token use by up to 92.9% and job time by up to 89.0% while preserving or improving task quality.
Takeaways & Limitations
SparseRead shifts agent efficiency from post-hoc context compression to pre-reading control and transfers across heterogeneous artifacts and agent frameworks.
Takeaways & Limitations
Sparse reading may add overhead without reducing token usage for tasks requiring whole-table computation, exact scanning, or direct small-file inspection.
Abstract
from arXiv · showhide
Long-horizon agents increasingly rely on repeated access to external artifacts, yet current reading interfaces often expose entire objects even when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute task-relevant evidence, while existing context-reduction methods mainly intervene after broad content has already entered the trajectory. We present SparseRead, a training-free, model-transparent reading layer that controls content admission before unnecessary evidence reaches the model context. SparseRead combines a regime-aware Read Gate, extensible Reader Backends, and a stateful protocol for bounded, source-anchored evidence acquisition with explicit refinement, verification, stopping, and fallback. Across six frontier models, including Claude Opus 5, and five workload scenarios, SparseRead reduces token volume by up to 92.9% and wall time by up to 89.0%, while preserving or improving task quality. Its consistent gains across three agent frameworks further demonstrate broad portability.
1 Introduction
Long-horizon agents repeatedly process external artifacts, but much exposed reading is irrelevant to the current decision. SparseRead addresses this over-reading by controlling artifact admission before broad content enters model context.
- The cost of over-reading: 60–80% of agent-trajectory tokens come from external observations, while only 6–15% of exposed reading contributes to the current decision in representative high-sparsity tasks.Full-read agents can consume up to 17.3× more tokens without achieving higher task quality.
- The cost of over-reading: Over-reading exposes artifacts at object granularity even when a decision requires only a bounded evidence slice.Excess content can dilute relevant evidence, inducing erroneous tool use, source confusion, and failures to stop.
- The missing capability: Existing context-reduction methods act after broad content enters the trajectory and therefore do not control the read itself.Long-context methods, cache compression, model-based compression, and retrieval address context cost or selection but not pre-read artifact admission.
- The missing capability: Sparse Reading limits exposure to what the agent’s current reasoning requires and expands the read only when additional evidence is needed.The abstraction follows human reading: inspect structure, focus on relevant passages, revisit unresolved details, and stop when the information need is satisfied.
- SparseRead: SparseRead combines a stateful Sparse Reading protocol, extensible Reader Backends, and an online Read Gate that selects the operating mode during inference.It integrates with existing agent frameworks and LLMs without parameter changes, additional training, or an auxiliary compression model.
- Evaluation: Across six models and five scenarios, SparseRead reduces token use by up to 92.9% and job time by up to 89.0% while preserving or improving task quality.It achieves the lowest token and time costs in every comparison with SOTA baselines and transfers across three agent frameworks.
2 Quantifying the Cost of Over-Reading
The paper measures over-reading as a mismatch between broad artifact exposure and the smaller amount of evidence needed for a decision. Selective reading reduces exposure while preserving or improving quality, showing that broad context is not a free substitute for reading control.
- More context does not improve task quality: Over-reading is a measurable failure mode, not only a cost problem, because exposing more content does not reliably improve task quality.The paper first quantifies useful admitted reading, then tests whether more exposed content improves quality.
- Useful evidence is sparse: Only 6–34% of the reading budget is useful in high-sparsity settings such as long-document QA, PDF QA, and audit bundles.These figures make full reading highly wasteful in the cited scenarios.
- More context does not improve task quality: Post-read compression cannot recover the cost of fully admitting a broad read or prevent trajectory drift caused by over-reading.Sparse Reading intercepts before ingestion by extracting task-relevant evidence and avoiding full-read overhead.
- More context does not improve task quality: Up to 17.3× more tokens are spent by full-read agents than SparseRead agents without improved task scores in paired high-sparsity runs.Across model and task groups, selective reading preserves or improves score while reducing exposed context.
3 Why Compression Alone Is Not Enough
Compression alone reduces admitted content but does not provide the provenance, completeness signals, or regime control needed for reliable sparse reading. SparseRead addresses these gaps with explicit protocol state and selective intervention.
- Compression Alone: Compression can reduce content while still failing to provide provenance and rationale sufficient for trustworthy reading decisions.The paper identifies missing provenance and compression rationale as barriers to terminating broad reading.
- Compression Distrust: 1.7–6.2× more requests and 2.1–12.5× more tokens occur without a protocol that exposes sparse-result coverage.The reported failure is uncertainty about whether evidence is complete or safe to use, rather than evidence size alone.
- Compression Distrust: Compression distrust can trigger repeated rereading, broad expansion, and another underspecified result.The loop arises because short answers lack source anchors, unresolved requirements, and a bounded next action.
- Design Implication: Sparse Reading requires explicit protocol state to establish trust and prevent blind rereading loops.SparseRead maintains compact protocol state rather than relying only on a compressor.
- Regime Boundary: Sparse reading suits localized extraction, whereas whole-table computation, exact scanning, and small-file inspection favor native access.The latter tasks may incur overhead without reducing token usage.
- Regime Boundary: 80–88% token reduction is reported for long-document QA, while forced sparse reading increases tokens by 309.2% on T67.The boundary motivates selective rather than universal enforcement; the passage also reports a 210.5% increase on T59.
4 SparseRead Design
SparseRead separates reading control from evidence acquisition and treats reading as stateful, sequential evidence collection. A Read Gate selects intervention strength, while Readers execute anchored actions under explicit closure and fallback rules.
- Architecture: SparseRead separates a control plane that regulates context admission from an evidence-compression plane implemented by Reader Backends.This separation lets new Readers register capabilities without changing the agent-facing macro-action set or admission policy.
- Architecture: The Unified Reader Interface fixes stable request, action, and evidence semantics between policy and object- or domain-specific access.Backend changes therefore need not propagate into the agent interface.
- Agent-Facing Protocol: An Object Preview bounds the initial view, Evidence State accumulates results, and the control plane validates actions, enforces budgets, and updates state.The agent remains the explicit policy actor for sequential evidence acquisition.
- Agent-Facing Protocol: The agent writes a Read Intent that guides evidence search, while Evidence State preserves a verifiable record and current search boundary.The intent can be narrowed in later rounds without introducing an auxiliary model.
- Reader Routing: The macro-actions scout, focus, collect, refine, and verify separate semantic reading decisions from mechanical backend selection.The Router maps the selected action to a compatible Reader and returns anchored observations.
- Stopping and Fallback: Closure occurs when evidence supports the pending decision and residual needs are resolved, accepted as non-critical, or explicitly taken as risk.Stalling, unusable results, or budget exhaustion trigger native fallback.
- Reader Backends: Evaluated Readers invoke no additional learned model during backend execution, keeping read-time overhead low and predictable.Model-based pruning can be wrapped as Readers, but that extensibility path is not evaluated in this work.
- Read Gate: The deterministic Read Gate selects force, native, or advisory using hard constraints and a weighted margin over exposure, protocol cost, and omission risk.Boundary cases go to advisory; force enters the sparse protocol, native bypasses it, and advisory preserves both paths.
5 Implementation and Agent Integration
SparseRead integrates as a shared runtime across agent harnesses, with each framework supplying only a thin adapter for routing broad object access into the core protocol.
- Runtime Integration: The SparseRead core owns the preview/read protocol, typed Readers, gate policy, artifact state, and bridge server.This shared runtime centralizes the reading controls used across integrations.
- Framework Adapters: NanoBot registers SparseRead tools directly, while OpenCode and OpenClaw use thin adapters exposing common preview and read calls.Adapters translate local read/open events into the same core calls and retain compatibility support.
6 Evaluation
Across models, scenarios, baselines, and frameworks, SparseRead reduces reading cost while generally preserving or improving task quality. Its gains extend to strong models, scientific workloads, context-reduction comparisons, and multiple agent frameworks.
- Effectiveness Across Models and Scenarios: SparseRead reduces token volume and wall time in all 30 model–scenario cells, with maximum reductions of 92.9% and 89.0%, respectively.Task score is preserved or improved in 26 cells.
- Effectiveness Across Models and Scenarios: Across six models in AI for Science workloads, SparseRead reduces token use by 82.7–89.1% and wall time by 28.3–63.0%.These results support transfer across scientific workloads as well as models.
- Comparison with Context-Reduction Baselines: SparseRead reduces tokens by 63.4–86.3% and wall time by 44.7–72.3% against ACON and CT observation masking across six settings.It achieves the lowest token volume and wall time in every setting, while its score never falls below Naive; the wall-time reductions correspond to 1.8–3.6× speedups.
- Portability Across Agent Frameworks: SparseRead reduces token volume and wall time in all 15 framework–model cells, preserving or improving score in 12.Median token reductions are 69.0%, 71.8%, and 28.7% across NanoBot, OpenCode, and OpenClaw; median wall-time savings are 64.4%, 64.9%, and 28.2%.
- Read-Gate Behavior: The Read Gate reduces token cost in 38 matched pairs and preserves or improves score in 39, with a 79.6% median token reduction in clear-win cases.Among 24 gate/pass cases, 20 reduce token cost and 23 preserve or improve score; exceptions concentrate in low-sparsity tasks.
7 Related Work
Prior approaches reduce context or tool-observation costs after information has been exposed, whereas SparseRead controls evidence admission before reading broad content. This targets both consumed cost and trajectory drift from over-reading.
- Post-Read Context Reduction: Context compression and KV-cache methods intervene after content has been read, so they cannot recover costs or trajectory drift induced by over-reading.These approaches include assembled-context compression and inference-time KV-state optimization.
- Pre-Reading Control: SparseRead selects evidence before admission without rewriting the existing prefix, moving efficiency control to the reading stage.This distinguishes its intervention point from post-hoc context reduction.
8 Conclusion
SparseRead makes reading an explicit, controllable part of agent execution through a stateful protocol, extensible Readers, and a dynamic Read Gate. Its gains transfer across artifacts, frameworks, and increasing model capability, while motivating reading control as a broader agent-design principle.
- Conclusion: SparseRead moves agent efficiency from post-hoc context compression to pre-reading control over external objects.The framework combines a stateful protocol, extensible Readers, and a dynamic Read Gate.
- Conclusion: The abstraction transfers across heterogeneous artifacts and agent frameworks while remaining model-independent and prefix-cache-friendly.The conclusion presents this transfer as a design property of the framework.
- Conclusion: Persistent gains as model capability increases suggest that over-reading is a structural inefficiency in agent interaction with external information.The paper frames reading as an explicit execution-path capability rather than an unrestricted tool side effect.