Knows: Agent-Native Structured Research Representations

Abstract: Research artifacts are distributed primarily as reader-oriented documents like PDFs. This creates a bottleneck for increasingly agent-assisted and agent-native research workflows, in which LLM agents need to infer fine-grained, task-relevant information from lengthy full documents, a process that is expensive, repetitive, and unstable at scale. We introduce Knows, a lightweight companion specification that binds structured claims, evidence, provenance, and verifiable relations to existing research artifacts in a form LLM agents can consume directly. Knows addresses the gap with a thin YAML sidecar (KnowsRecord) that coexists with the original PDF, requiring no changes to the publication itself, and validated by a deterministic schema linter. We evaluate Knows on 140 comprehension questions across 20 papers spanning 14 academic disciplines, comparing PDF-only, sidecar-only, and hybrid conditions across six LLM agents of varying capacity. Weak models (0.8B--2B parameters) improve from 19--25\% to 47--67\% accuracy (+29 to +42 percentage points) when reading sidecar instead of PDF, while consuming 29--86\% fewer input tokens; an LLM-as-judge re-scoring confirms that weak-model sidecar accuracy (75--77\%) approaches stronger-model PDF accuracy (78--83\%). Beyond this controlled evaluation, a community sidecar hub at https://knows.academy/ has already indexed over ten thousand publications and continues to grow daily, providing independent evidence that the format is adoption-ready at scale.

• The paper introduces a YAML-based sidecar specification, KnowsRecord, to augment PDFs with structured claims, evidence, and provenance.
• It employs deterministic linting and cross-model evaluations to achieve up to 93% token savings while significantly boosting accuracy across LLM agents.
• Enhanced traceability and interoperability with existing scholarly standards enable efficient agent consumption and bridge to agent-native publishing.

Knows: A Sidecar Specification for Agent-Native Structured Research Representations

Motivation and Problem Setting

The exponential increase of agent-assisted and agent-native scientific workflows accentuates a long-standing bottleneck: the distribution of scholarly knowledge as reader-centric documents (PDFs, HTML) optimized for humans rather than LLM-based consumers. Extracting fine-grained, task-relevant information (methods, results, evidence) from narrative prose requires computationally intensive, prompt- and context-sensitive processing that is both unstable and highly inefficient. In agent-centric pipelines, every LLM agent redundantly re-parses unstructured prose, resulting in duplicated computation and inconsistent information recovery.

To address this inefficiency and the absence of a machine-consumable intermediate, "Knows: Agent-Native Structured Research Representations" (2604.17309) introduces a lightweight, YAML-based sidecar specification—KnowsRecord—to supplement existing research artifacts with structured, schema-validated claims, evidence, provenance, and relations. This approach targets immediate gains for current research agents and serves as a bridge toward an ecosystem in which structure is primary and prose becomes a byproduct.

Figure 1: Overview of Knows, where a YAML sidecar augments the PDF with structured claims and evidence, enabling token savings and agent-native consumption.

KnowsRecord Specification and Ecosystem

KnowsRecord v0.9 defines a JSON-schema-conformant YAML document with five primary, schema-enforced collections: artifacts, statements, evidence, relations, and actions, augmented by provenance, versioning, and freshness metadata. Its design covers 30 root-level fields and 23 entity types, validated by deterministic lint. Importantly, KnowsRecords coexist with the original human-readable PDF, requiring zero change to publication workflows or downstream integrations.

Entities resolve inter-/intra-record references (via record_id#local_id) supporting cross-record traceability (e.g., peer reviews referencing claims in the evaluated work). Each structured claim or evidence annotation is explicitly annotated by type, modality, confidence, and extraction fidelity. The schema is forward-extensible via x_extensions, domain profiles (e.g., paper@1, review@1), and semantic versioning. Critically, Knows does not replace but interoperates with scholarly infrastructure standards: DataCite, PROV, CiTO, CodeMeta, RO-Crate, BibTeX, etc.

Figure 2: KnowsRecord architecture and workflow, showing primary entities, sidecar generation/validation, and cross-record linking for reviews.

Retrieval Protocol and Agent Consumption

The normative agent consumption protocol is formalized: upon receipt of a query, an agent attempts to validate and process the KnowsRecord sidecar, retrieving relevant statements/evidence and their source anchors. If coverage/confidence is insufficient, or structural validation fails, the protocol falls back to anchored PDF retrieval. Confidence thresholds (parameterized by $\tau$) and deterministic schema linting enable robust, reproducible behaviors, supporting both sidecar-only and hybrid fallback modes.

Figure 3: Diagram of the agent retrieval protocol, detailing decision points, structural fallback, and the hybrid query path.

This pipeline yields (i) efficient retrieval for task-relevant queries, (ii) audit-ready citation traces, and (iii) cross-record relational reasoning in contexts such as peer review and review-as-sidecar workflows.

Experimental Results

Comprehensive cross-model, multi-discipline evaluation (140 questions, 20 classic papers, 14 disciplines, 6 LLM agents/tiers) demonstrates:

• Weak models (0.8--2B params) gain 29–42 percentage points in accuracy reading sidecars versus PDFs, e.g., 19%→47%/67% (Qwen-0.8B/2B), with 29–86% token savings.
• Sidecar-augmented weak models (75–77% LLM-judged accuracy) approach strong-model PDF performance (78–83%).
• Fallback mode yields highest accuracy for 5/6 models, recovering content absent from the sidecar while maintaining token/cost efficiency.
• For strong models, full sidecars attain comparable accuracy to PDFs using only 55% of the tokens; further ablation finds that statement-only sidecars retain 88% of the accuracy with 93% token savings.
• Traceability is qualitatively improved: under Knows, per-weakness ID references in reviews span 64–91%, contrasting with 0% in PDF-only settings—a crucial property for automated audit and multi-agent reasoning.
• Deterministic linting detects 100% of structural corruption but 0% of semantic corruption, motivating a future LLM-based verification layer.
• The sidecar effect is primarily driven by statement granularity; denser sidecars eliminate discipline-dependent performance gaps.
• LLM-based sidecar generation: Only Claude-family models reach 100% schema-valid outputs with high accuracy when consumed by other agents; non-Claude models presently lack sufficient competence for structurally-valid extraction in complex domains.

Case Studies & Ecosystem Applications

Knows is operationalized at scale, with a public community hub indexing $>$10,000 publications. Demonstrated use cases include:

• Large-scale literature intelligence: Agent-driven conference analysis, field-wide trend tracking, and survey bootstrapping—where token cost prohibits raw PDF/HTML consumption.
• Human-agent collaborative workflows: Discipline-specific method/evidence extraction, limitations synthesis, and fine-grained comparative analysis across diverse user intents.
• Agent-native research pipelines: Autonomous authoring, reviewing, and meta-research across multi-agent systems, where structured, machine-traversable research objects enable compositional, audit-ready protocol chaining.

Trajectory toward Agent-Native Publishing

Knows establishes a bridge from PDF-centric to agent-native publishing. The forward trajectory envisions:

• Authoring toolchains natively emitting both structured and prose representations.
• Autonomous research agents authoring, reviewing, and transacting research as machine-validated records, with prose as a derived human interface.
• An agent-native scholarly record, where traceability, provenance, and compositional reasoning are primary, and compatibility with legacy formats is preserved.

  Figure 4: The three-phase agent-native publishing trajectory—transitioning from narrative PDF, to sidecar-augmented, to structured-record-centric publishing.

Implications and Future Directions

Practically, Knows enables immediate acceleration of agentic research, systematic error tracing, and large-scale literature audits. Theoretically, it represents a modular, compositional, and verifiable substrate for autonomous scientific ecosystems. Prominent future lines include LLM-based semantic verification, lifecycle audit-trace formalization, and domain-specific profile expansion (e.g., dataset@1, model@1). As agent-native conferences and publishing venues emerge, Knows or similar schema will become central infrastructure.

Conclusion

Knows delivers a minimal, interoperable, and schema-validated format for structured research representation directly consumable by LLM agents. It achieves significant improvements in efficiency, consistency, and traceability across agent-assistive and agent-native settings, and is deployable at scale without disrupting legacy human-centric workflows. As agent-led scientific workflows grow, modular, schema-first representations such as Knows are poised to become the backbone of the scholarly graph, bridging current and future paradigms of research communication.