Don't Blink: Evidence Collapse during Multimodal Reasoning

• URL: http://arxiv.org/abs/2604.04207v1
• Date: Sun, 05 Apr 2026 18:01:00 GMT
• Title: Don't Blink: Evidence Collapse during Multimodal Reasoning
• Abstract summary: Reasoning VLMs can become more accurate while progressively losing visual grounding as they think.<n>This creates task-conditional danger zones where low-entropy predictions are confident but ungrounded.<n>Using this structure, a targeted vision veto reduces selective risk by up to 1.9 percentage points at 90% coverage.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reasoning VLMs can become more accurate while progressively losing visual grounding as they think. This creates task-conditional danger zones where low-entropy predictions are confident but ungrounded, a failure mode text-only monitoring cannot detect. Evaluating three reasoning VLMs on MathVista, HallusionBench, and MMMU_Pro, we find a pervasive evidence-collapse phenomenon: attention to annotated evidence regions drops substantially, often losing over half of evidence mass, as reasoning unfolds. Full-response entropy is the most reliable text-only uncertainty signal under cross-dataset transfer, yet adding vision features with a single global linear rule is brittle and often degrades transfer. An entropy-vision interaction model reveals a task-conditional regime: lowentropy, visually disengaged predictions are hazardous on sustained visual-reference tasks but benign on symbolic tasks. Using this structure, a targeted vision veto reduces selective risk by up to 1.9 percentage points at 90% coverage, while avoiding degradations where disengagement is expected. The results support task-aware multimodal monitoring for safe deployment under distribution shift.

Related papers

• When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs [54.95635713318535]
  Vision-language models (VLMs) are increasingly deployed in high-stakes settings, where a response that is reasonable in general may still be unsafe for a particular user.<n>We study this problem of personalized safety in multimodal systems and introduce MPS-Bench, a benchmark of 5,181 scenarios from 584 real-world images.<n>We find that they almost always respond directly (86-99%) rather than seek missing context, and none exceeds 2.6/5 on personalized safety.<n>Motivated by this mechanism, we propose PRISM, a lightweight input monitor that uses bidirectional cross-modal modulation to predict when a query is likely to require
  arXiv  Detail & Related papers  (2026-09-03T03:58:56Z)
• Geometric Collapse: When Vision Models Fail to Verify Physical Causality [52.43819179934414]
  We propose Scrambled Edges, a controlled counterfactual that injects salient edge-like cues.<n>With energy-matched and structure-matched controls, we isolate the effect of unsupported edge evidence from high-frequency energy and edge sparsity.<n>Across CNN/ViT/SSL depth predictors, Scrambled Edges induce up to 3.2x larger deviation from clean predictions than energy-matched noise.
  arXiv  Detail & Related papers  (2026-07-08T00:04:02Z)
• SPOT-E: Test-Time Entropy Shaping with Visual Spotlights for Frozen VLMs [74.89208610190595]
  We present SPOT-E, a test-time method that produces question-conditioned spotlights, optimized per instance via light-weight tuning.<n> SPOT-E yields consistent gains and improved robustness under visual corruptions.
  arXiv  Detail & Related papers  (2026-06-18T13:56:30Z)
• When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models [7.185483586381928]
  Failures in multi-turn reasoning models are largely invisible to terminal-score evaluation.<n>A model can lock onto an unsafe stance early in a long dialogue, yet its final-turn refusal rate may appear indistinguishable from a robustly aligned baseline.<n>We propose a trace-level diagnostic - the CoT-Output 2x2 safety matrix.
  arXiv  Detail & Related papers  (2026-06-09T11:50:28Z)
• Consistent Yet Wrong: Evidence Insensitivity in Spatial Vision-Language Models [29.66841995436342]
  Modern vision-language models (VLMs) remain unreliable on metric distance queries.<n>We introduce textbfViewDiag, a controlled multi-view evaluation protocol built from Hypersim, ScanNet, and KITTI360.<n>We show that stable predictions may reflect prior-driven collapse rather than evidence-sensitive reasoning.
  arXiv  Detail & Related papers  (2026-06-01T18:06:08Z)
• Detecting Is Not Resolving: The Monitoring Control Gap in Retrieval Augmented LLMs [20.59321114618083]
  We show that single-turn diagnostics systematically overestimate RAG safety, that contradiction is uncorrelated with safe resolution, and that no universal prompt fix exists.<n>The gap between what models recognize and what they do must be measured and closed before retrieval-augmented systems can be trusted in high-stakes settings.
  arXiv  Detail & Related papers  (2026-05-26T15:18:43Z)
• Silent Collapse in Recursive Learning Systems [24.578102943388256]
  We propose a lightweight metacognitive loop that monitors trajectory statistics and estimates a slow-timescale trust variable.<n>MTR provides early warning and actively prevents silent collapse without requiring access to pristine real data.
  arXiv  Detail & Related papers  (2026-05-14T08:59:26Z)
• Neural Uncertainty Principle: A Unified View of Adversarial Fragility and LLM Hallucination [60.197429875410286]
  Adversarial vulnerability in vision and hallucination in large language models are conventionally viewed as separate problems.<n>Formalizing a Neural Uncertainty Principle (NUP) under a loss-induced state, we find that in near-bound regimes, further compression must be accompanied by increased sensitivity dispersion.<n>In vision, masking highly coupled components improves robustness without costly adversarial training.<n>In language, the same prefill-stage probe detects hallucination risk before generating any answer tokens.
  arXiv  Detail & Related papers  (2026-03-20T02:07:10Z)
• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• Revisiting Multivariate Time Series Forecasting with Missing Values [65.30332997607141]
  Missing values are common in real-world time series.<n>Current approaches have developed an imputation-then-prediction framework that uses imputation modules to fill in missing values, followed by forecasting on the imputed data.<n>This framework overlooks a critical issue: there is no ground truth for the missing values, making the imputation process susceptible to errors that can degrade prediction accuracy.<n>We introduce Consistency-Regularized Information Bottleneck (CRIB), a novel framework built on the Information Bottleneck principle.
  arXiv  Detail & Related papers  (2025-09-27T20:57:48Z)
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning [76.82121084745785]
  Believe Your Eyes (BYE) is a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples.<n>It achieves near-zero attack success rates while maintaining clean-task performance.
  arXiv  Detail & Related papers  (2025-05-22T17:11:58Z)
• SafetyNet: Detecting Harmful Outputs in LLMs by Modeling and Monitoring Deceptive Behaviors [2.07180164747172]
  High-risk industries like nuclear and aviation use real-time monitoring to detect dangerous system conditions.<n>We propose a real-time framework to predict harmful AI outputs before they occur by using an unsupervised approach.
  arXiv  Detail & Related papers  (2025-05-20T12:49:58Z)
• Hinge-Wasserstein: Estimating Multimodal Aleatoric Uncertainty in Regression Tasks [9.600416563894658]
  We study regression from images to parameter values and here it is common to detect uncertainty by predicting probability distributions. Traditional loss functions lead to poor probability distribution estimates and severe overconfidence, in the absence of full ground truth distributions. We propose hinge-Wasserstein -- a simple improvement of the Wasserstein loss that reduces the penalty for weak secondary modes during training.
  arXiv  Detail & Related papers  (2023-06-01T11:20:09Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.