When Verification Fails: How Compositionally Infeasible Claims Escape Rejection

arXiv cs.CL / 4/14/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper studies scientific claim verification under the Closed-World Assumption (CWA), where a claim should be accepted only if all asserted constraints are positively supported by evidence.
It argues that existing verification benchmarks fail to detect models that use a shortcut—salient-constraint checking—by focusing only on the most salient constraint rather than all constraints.
The authors introduce compositionally infeasible claims where the salient constraint is supported while a non-salient constraint is contradicted, revealing that many models over-accept such claims.
Across multiple model families and modalities, the results indicate widespread shortcut reasoning and show via context interventions that models differ mainly in verification thresholds rather than fundamentally in reasoning capability.
The paper concludes that a compositional inference bottleneck is a structural limitation of current verification behavior that cannot be fixed reliably by prompt/strategy guidance alone.

Abstract

Scientific claim verification, the task of determining whether claims are entailed by scientific evidence, is fundamental to establishing discoveries in evidence while preventing misinformation. This process involves evaluating each asserted constraint against validated evidence. Under the Closed-World Assumption (CWA), a claim is accepted if and only if all asserted constraints are positively supported. We show that existing verification benchmarks cannot distinguish models enforcing this standard from models applying a simpler shortcut called salient-constraint checking, which applies CWA's rejection criterion only to the most salient constraint and accepts when that constraint is supported. Because existing benchmarks construct infeasible claims by perturbing a single salient element they are insufficient at distinguishing between rigorous claim verification and simple salient-constraint reliance. To separate the two, we construct compositionally infeasible claims where the salient constraint is supported but a non-salient constraint is contradicted. Across model families and modalities, models that otherwise saturate existing benchmarks consistently over-accept these claims, confirming the prevalence of such shortcut reasoning. Via model context interventions, we show that different models and prompting strategies occupy distinct positions on a shared ROC curve, indicating that the gap between model families reflects differences in verification threshold rather than underlying reasoning ability, and that the compositional inference bottleneck is a structural property of current verification behavior that strategy guidance alone cannot overcome.