Witness Evidence Portfolios: Single-Prefill Risk Detection for Closed Multimodal Answers

2026-07-30

AI Summary

This work addresses the problem of overconfident yet unreliable outputs in closed-ended visual question answering by multimodal large language models (MLLMs). The authors propose Witness Evidence Portfolios, a method that leverages white-box prefilling pathways to quantify, layer by layer, the degree to which visual evidence supports a given prediction—without modifying the decoder or introducing external verifiers. By integrating question-relevant evidence tracing, symbolic concentration metrics, and nested group validation, the approach constructs sparse top-k evidence portfolios to identify high-risk predictions. As the first single-pass prefill risk detection framework that requires neither image perturbations nor backpropagation, it consistently improves performance across three MLLMs and four binary VQA benchmarks, achieving an average 0.134 gain in error detection AP, with strictly positive confidence intervals observed across ten image-cluster-guided groups.

Abstract

Reliable deployment of multimodal large language models (MLLMs) requires deciding whether a confident visual answer should be trusted, reviewed, or routed to a stronger system. Confidence scores capture candidate margins, but not where the estimated signed visual readouts associated with those margins come from or how they are distributed. We study inference-time risk detection for closed visual answers using the same white-box prefill path that produces the answer. Witness Evidence Portfolios (WEP) first estimates, layer by layer, which visual contributions support or contradict the predicted candidate. It summarizes these contributions through two interpretable route families: question-related evidence provenance and signed evidence concentration. Nested grouped validation chooses the more reliable family and a sparse top-k route portfolio, which is fused with candidate confidence. WEP needs no image perturbation, decoding change, backward pass, or external verifier. Across three MLLMs and four binary-answer benchmarks, WEP improves mean error AP by 0.134. All 12 model--dataset gains are positive, and image-cluster bootstrap intervals are strictly positive on 10 pairs. WEP targets white-box closed-answer systems and uses a labeled calibration slice.

Problem

• risk detection
• multimodal large language models
• closed visual answers
• confidence reliability
• trustworthiness assessment

Innovation

• Witness Evidence Portfolios
• risk detection
• multimodal LLMs
• prefill-path analysis
• interpretable evidence