Abstract
Agentic LLM systems such as OpenScholar and PaperQA2 read scientific literature and return cited answers. Both they and their benchmarks check whether those citations hold, but neither audits the reliability of that check itself. We demonstrate that this check is unreliable, which is significant. For identical agent outputs, the measured unsupported-citation rate ranges from about 3% to 18%, depending solely on the verifier's strictness. Although verifiers agree on which citations are supported, they disagree on which to flag (negative-specific agreement 0.27 to 0.30), making no single flag set trustworthy, and cross-paper comparison invalid without a named verifier and protocol.
We present a gold-anchored evaluation protocol and a deployable guard that make this behavior measurable and bounded. The protocol validates the verifier, measures re-attribution, and calibrates a guarantee against human gold rather than another model's verdict; the verifier is a swappable instrument chosen on cost (recall 0.94 on the supported class, held out), and re-attribution is a commodity step where a deterministic BM25 matches the best open generator. The guard adds a split-conformal layer placing a distribution-free, finite-sample bound on truly unsupported citations that slip past a chosen flagging rule, guaranteeing catch rate rather than conclusion correctness. The bound holds on held-out gold, and we identify and quantify the condition governing its transfer to deployment, calibration-negative difficulty, with a concrete recalibration recipe, left untested by prior conformal-factuality work.
Validated across four open 27-35B models and three agentic pipelines on public benchmarks (SciFact, QASA, PubMedQA), with confidence intervals on every headline number, the protocol and guard ship as an open single-GPU kit.
Blogger's Review: This paper introduces a novel perspective on the authenticity of scientific citations, which is crucial in today's context of information overload. The proposed methods enhance the transparency of citation checks and provide a standardized framework for future research, ensuring the reliability of scholarly references.