Papers/2609.09184
🧪 Test?View on arXiv

Evidence-Order Calibration for Selective Visual Reasoning under Progressive Loss of Question-Critical Evidence

Not provided in the abstract

visual reasoningconfidence calibrationevidence lossmultimodal
2609.09184
Builder Relevance
70%
1h ago

Abstract

This paper investigates the reliability of vision-language models under conditions of degraded visual evidence, revealing significant inconsistencies in confidence levels.

Reality Card

Core Claim

The study demonstrates that adding evidence-order supervision to a binary cross-entropy loss reduces the evidence monotonicity violation rate (EMVR) from 0.330 to 0.303.

Method / Result

The evidence-order supervision approach reduced EMVR by 0.027, indicating improved reliability in visual reasoning.

Limitations

The results on AUROC, Brier, and AURC differences are statistically inconclusive, raising concerns about the reproducibility of the findings.

Paper to code

Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.

No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.
← Back to all papers