Does Recursive Knowledge Construction Help Agents Verify Consensus Claims in Science?

Co-scientists and agentic scientific claim verifiers are increasingly used to assess claims across the sciences. Natural-language scientific claims, written for human experts, are often underspecified and so admit several defensible interpretations. Such verifiers usually pick one, often tacitly, and benchmarks score their verdicts against expert labels without recording the expert's interpretation or the verifier's in a comparable form. This leads to disagreements with expert labels that cannot be traced to the evidence or the interpretation, and to compute spent on unchecked interpretations. We introduce Claim Cards to make such interpretations explicit in the reasoning itself. A Claim Card records a verifier's interpretation as nine fields before evidence gathering, each marked stated, inferred, or default. To test whether an agent's interpretation can change its verdict, we ran a GPT-5.4 agent five times on each of 60 expert-labeled materials-science and biomedical claims with a card, without one, and with one card frozen across runs (900 runs). Over 40% of card fields were not marked stated, and a card added about a tenth to a run's GPT-5.4 tokens and time without measurably changing agreement with expert labels. The agent's own interpretations varied between runs, although within a claim this variation did not measurably track the verdict. A deliberate change to one field of the frozen card, however, making the quantifier universal or the threshold for support stricter, changed the verdict on about a third of the claims, against about one in seven expected by chance (p < 0.001 for each edit). On 24 materials-science claims run with computational tools, the stricter threshold also changed what the agent asked those tools to compute. We therefore argue that verdicts should be reported with their interpretations and that benchmarks could publish the interpretation behind each label.

Scientific discovery increasingly relies on automated systems that generate hypotheses, inspect multimodal evidence, and validate claims at scale. Yet scientific claim verification is not well served by asking a vision-language model for a direct binary judgment: claims often combine numerical results, comparisons, scope qualifiers, and explanatory context, while evidence is encoded in tables and figures with distinct grounding structures. We present SciLens, an evidence-conditioned atomic entailment framework for multimodal scientific claim verification. SciLens decomposes each claim into central empirical atoms and background atoms, grounds the central atoms to modality-specific evidence witnesses, and predicts the final label with an atom-level entailment rule. For tables, atoms are grounded to rows, columns, cells, arithmetic relations, and table scope; for figures, they are grounded through panels, axes, legends, visual encodings, categories, trends, ranks, and qualifier checks. This yields a unified validation procedure in which a claim is supported only if every central empirical atom is entailed by the current evidence. On the SciClaimEval development set, SciLens achieves 79.2% macro-F1 and 63.1% pair accuracy, showing that structured agentic validation improves both evidence sensitivity and interpretability.

Scientific AI agents increasingly retrieve evidence, reason across sources, and synthesize conclusions used in consequential decisions. Yet, their ability to do so in high-stakes domains such as health remains unclear. We introduce SciConBench, a large-scale live benchmark of 9.11K questions and expert-written conclusions from systematic reviews to evaluate open-domain scientific conclusion synthesis. The benchmark draws on an expert-validated automated evaluation pipeline that decomposes conclusions into atomic facts and measures correctness and comprehensiveness via factual precision and recall. To mitigate data leakage, we further introduce SciConHarness, a clean-room evaluation harness that equips agents with controlled web interaction to ensure valid measurement. Evaluating 8 frontier models and deep research agents, we find that factual quality remains low: under clean-room settings, the best agent achieves only a factual F1 of 0.337. Our clean-room setting consistently reduces performance relative to unconstrained evaluation, suggesting that leakage inflates estimates of models' true synthesis capabilities. Finally, we audit consumer-facing agents (e.g., Google AI Overview, OpenEvidence) and find they frequently generate incomplete and sometimes contradictory conclusions, even when the ground-truth answer is available. Overall, our results show that reliable synthesis of scientific conclusions remains an open challenge, and that clean-room evaluation is essential for assessing open-domain AI agents.