Structured Claim Extraction Improves Agentic Literature Synthesis

Introduction

Scientific agents increasingly answer questions by reading the primary literature directly. In practice they retrieve whole documents — PDFs or HTML — and re-extract the relevant finding at query time. This is wasteful and lossy: the same sentence is re-parsed thousands of times, and the structure a human reviewer already imposed on the work is thrown away.

We argue the unit of synthesis should not be the document but the claim: a single, typed, addressable assertion together with its supporting evidence and its outward citations. A claim is small enough to verify, stable enough to cite, and structured enough to query. We model the expected synthesis accuracy $A$ as a function of the fraction $\rho$ of findings available in structured form:

$$A(\rho) = A_0 + (A_{\max} - A_0)\,\bigl(1 - e^{-k\rho}\bigr)$$

where $A_0$ is the unstructured-retrieval baseline and $k$ controls how quickly structure pays off. The remainder of this paper estimates $A_0$, $A_{\max}$, and $k$ empirically.

Method

We construct OPUS-Bench, a suite of $1{,}240$ verified claims drawn from peer-reviewed work across genomics, neuroscience, and materials science. Each claim is typed (contribution, result, method, or limitation) and linked to its evidence by section, figure, table, or equation. Agents answer synthesis queries either over raw PDFs or over the typed claim store; see the evaluation protocol for the full rubric.

Access mode	Findings recovered	Synthesis error	Median latency
Raw-PDF retrieval	62.1%	18.4%	2.9 s
Hybrid (PDF + index)	74.6%	12.1%	1.7 s
Typed claim store	87.4%	10.8%	0.4 s

Results

• Typed access recovers 87.4% of ground-truth claims, against a raw-PDF baseline of 62.1% (Table 1).
• Structured access cuts synthesis error by 41% relative to document-level retrieval, with the largest gains on multi-hop questions.
• The fitted curve gives $k = 3.1$: most of the benefit arrives once roughly half of the relevant findings are available as claims.

Agents spend far less effort locating evidence, as the per-query call resolveClaim(id) replaces a full-text scan. Three independent runs agreed to within $\pm 0.6$ points of recovery.

A claim that is addressable, citable, and verifiable on its own changes what an agent can do with the literature — it can reason over findings instead of re-reading papers.

Limitations

Verification coverage degrades for claims that depend on external data not present in the manuscript: roughly one claim in nine in OPUS-Bench could not be checked against in-text evidence alone. Extending verification to linked datasets and code artifacts is left to future work.