The paper introduces Evidential Uncertainty Quantification (EUQ), a fine‑grained method for detecting misbehaviors in large vision‑language models (LVLMs). EUQ interprets model output features as supporting or opposing evidence and aggregates them using Evidence Theory to quantify internal conflict and ignorance in a single forward pass. Experiments on hallucinations, jailbreaks, adversarial attacks, and out‑of‑distribution failures show that EUQ outperforms strong baselines, linking hallucinations to high conflict and OOD failures to high ignorance, and provides layer‑wise dynamics for interpreting internal representations.
arXiv:2608. 03817v1 Announce Type: cross Abstract: Large vision--language models (LVLMs) demonstrate strong multimodal reasoning capabilities but remain prone to hallucination, where model predictions are not grounded in visual evidence.
arXiv:2606. 07537v1 Announce Type: cross Abstract: Large language models hallucinate--producing fluent, confident, factually wrong outputs--with a consistency that persists across generations and scales.
arXiv:2607. 14499v1 Announce Type: new Abstract: Multi-modal Large Language Models (MLLMs) have made substantial advances on benchmarks, yet their real-world effectiveness remains uncertain.
The paper introduces Causal-Invariant Masking (CIM) to better quantify epistemic uncertainty in Multimodal Large Language Models (MLLMs) by measuring semantic shift between original predictions and those conditioned on a causally-focused view. It proposes Semantic Divergence as a core metric that converges to the variance of the model’s sensitivity to non‑causal correlations, and introduces Expected Embedding Drift (EED) as a fast geometric proxy that estimates this shift directly in the embedding space. Experiments demonstrate state‑of‑the‑art uncertainty quantification performance and a nearly 50% speedup with EED.
arXiv:2507. 06722v2 Announce Type: replace-cross Abstract: Understanding how large language models (LLMs) internally represent and process their predictions is central to detecting uncertainty and preventing hallucinations.
arXiv:2603. 24967v2 Announce Type: replace Abstract: Understanding why a large language model (LLM) is uncertain about the response is important for their reliable deployment.
CounterCount is a diagnostic dataset that pairs factual and counterfactual images to test whether Vision‑Language Models (VLMs) rely on visual evidence or on world knowledge when counting attributes. Humans achieve near‑perfect accuracy on both types of images, whereas VLMs drop from 91.6% to 49.5% on counterfactuals, often defaulting to canonical counts. The study shows that VLMs exhibit a System‑1‑like bias, underutilize visual evidence, and that targeted input manipulations can mitigate this bias.
arXiv:2608. 10430v1 Announce Type: cross Abstract: Large Language Models (LLMs) deployed as AI agents frequently exhibit user specification-grounding failures, executing hallucinated, undesired actions to force a resolution rather than expressing uncertainty.
arXiv:2608.22232v1 Announce Type: new Abstract: Real-world situation appearances can deviate from their underlying physical states, challenging the reliability of multimodal large language models (ML...
FusionAudit introduces PaRMA, a pathway-conditioned robustness auditing framework for native multimodal vision‑language models. It evaluates two image‑level interventions—pixel perturbation and rendered‑text insertion—to expose architecture‑specific blind spots. The study applies diagnostic‑motivated attacks (MANTIS and Nat‑Evid) to 18 open‑weight VLMs, revealing that legacy single‑attack rankings are unstable and that PaRMA uncovers higher residual risk than standard benchmarks.
arXiv:2603. 03989v2 Announce Type: replace-cross Abstract: When visual evidence is ambiguous, vision models must decide how to interpret face-like patterns.
arXiv:2606. 19353v1 Announce Type: cross Abstract: In-Context Learning (ICL) allows LLMs to adapt to new tasks from a few demonstrations, but its reliability remains a concern: predictions are highly sensitive to both prompt design and the model's ability to understand the context, obscuring whether failures arise from data properties or model limitations.