Research map / Vision-language and multimodal / Multimodal LLMs

Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Bias-aware multi-image robustness

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EUQ, a training-free method that decomposes epistemic uncertainty of LVLMs into conflict and ignorance via Dempster-Shafer evidence from the output-head pre-logits, detecting hallucinations, jailbreaks, adversarial inputs and OOD failures in a single forward pass.

Problem

Detecting misbehaviors (hallucinations, jailbreaks, adversarial vulnerabilities, OOD failures) of large vision-language models, where existing uncertainty methods capture only total predictive uncertainty and are costly (sampling) or unreliable (verbalized confidence).

Approach

Treat output-projection pre-logits features as evidence, convert them into positive/negative belief mass per hypothesis via an LCP-optimal affine transform, and fuse with Dempster's rule to get separate conflict (CF) and ignorance (IG) measures from one forward pass.

Key results

• CF and IG beat the best baseline by 10.4% and 7.5% AUROC and 5.3% and 5.5% AUPR on average across four LVLMs and misbehavior types.
• CF reaches average AUROC 0.812 and IG 0.783 across datasets, versus 0.707 for HiddenDetect and 0.626 for self-consistency.
• Hallucinations are associated with higher conflict, while OOD failures are associated with higher ignorance; EUQ needs ~1e-3 to 1e-2 s per example versus ~0.9 s for sampling methods.

Surprising findings

• Prompting models with a 'None of the above' option rarely elicits abstention (Qwen 0.27%, Intern 0.00%), indicating persistent overconfidence.

Limitations

• Requires access to internal representations, so it cannot be applied to closed-source API models such as GPT-4.
• Layer-wise separation of misbehaviors is apparent only in certain decoder layers.
• Detection performance varies strongly across models and datasets (e.g., IG on OOD for some models and jailbreak on Qwen).

Experimental setup

Datasets: POPE R-Bench FigStep Hades VisualAdv ANDA PGN MMBench

Benchmarks: POPE R-Bench FigStep Hades VisualAdv ANDA PGN Misbehavior-Bench

Metrics: AUROC AUPR accuracy

Baselines: self-consistency semantic entropy predictive entropy length-normalized predictive entropy HiddenDetect

Builds on: Dempster-Shafer theory evidential deep learning HiddenDetect semantic entropy predictive entropy

Closest papers by meaning

• FUSE: Quantifying Uncertainty in Vision-Language Models by Bayesian Fusing Epistemic and Aleatoric Uncertainty (ICML 2026 · similarity 0.90)
• Enhancing Vision-Language Model Reliability with Uncertainty-Guided Dropout Decoding (NeurIPS 2025 · similarity 0.89)
• VES-RFT: Rewarding Visual Evidence Sensitivity to Mitigate Hallucinations in Large Vision-Language Models (CVPR 2026 · similarity 0.89)
• On Epistemic Uncertainty of Visual Tokens for Object Hallucinations in Large Vision-Language Models (NeurIPS 2025 · similarity 0.88)
• CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness (ICLR 2025 · similarity 0.88)
• A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models (ICLR 2026 · similarity 0.87)