# Don't Blink: Evidence Collapse during Multimodal Reasoning

arXiv: https://arxiv.org/abs/2604.04207 · HTML: https://mlsanity.com/paper/2604.04207/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Multimodal reasoning, one of 17 directions in LLM reasoning.
- Builds on 21 of its 23 references that are in the map.
- First posted on arXiv 5 Apr 2026.

## Placed in

- LLM reasoning / [Multimodal reasoning](https://mlsanity.com/lr/multimodal-reasoning/)

## Builds on (most cited first, 21 of 21)

- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [On Calibration of Modern Neural Networks](https://mlsanity.com/paper/1706.04599/) (2017)
- [Qwen3-VL Technical Report](https://mlsanity.com/paper/2511.21631/) (2025)
- [MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts](https://mlsanity.com/paper/2310.02255/) (2023)
- [Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs](https://mlsanity.com/paper/2306.13063/) (2023)
- [Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation](https://mlsanity.com/paper/2302.09664/) (2023)
- [HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models](https://mlsanity.com/paper/2310.14566/) (2023)
- [MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark](https://mlsanity.com/paper/2409.02813/) (2024)
- [GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning](https://mlsanity.com/paper/2507.01006/) (2025)
- [Measuring Calibration in Deep Learning](https://mlsanity.com/paper/1904.01685/) (2019)
- [DeepStack: Deeply Stacking Visual Tokens is Surprisingly Simple and Effective for LMMs](https://mlsanity.com/paper/2406.04334/) (2024)
- [Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs](https://mlsanity.com/paper/2510.17771/) (2025)
- [Prolonged Reasoning Is Not All You Need: Certainty-Based Adaptive Routing for Efficient LLM/MLLM Reasoning](https://mlsanity.com/paper/2505.15154/) (2025)
- [Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LLM Reasoning](https://mlsanity.com/paper/2510.08146/) (2025)
- [Investigating Uncertainty Calibration of Aligned Language Models under the Multiple-Choice Setting](https://mlsanity.com/paper/2310.11732/) (2023)
- [Unveiling Uncertainty: A Deep Dive into Calibration and Performance of Multimodal Large Language Models](https://mlsanity.com/paper/2412.14660/) (2024)
- [Predicting When to Trust Vision-Language Models for Spatial Reasoning](https://mlsanity.com/paper/2601.11644/) (2026)
- [KVSmooth: Mitigating Hallucination in Multi-modal Large Language Models through Key-Value Smoothing](https://mlsanity.com/paper/2602.04268/) (2026)
- [Explicit Abstention Knobs for Predictable Reliability in Video Question Answering](https://mlsanity.com/paper/2601.00138/) (2025)
- [VLM-UQBench: A Benchmark for Modality-Specific and Cross-Modality Uncertainties in Vision Language Models](https://mlsanity.com/paper/2602.09214/) (2026)
- [SAVAA: Mitigating Hallucinations in LVLMs via Step-wise Adaptive Visual Attention Amplification](https://mlsanity.com/paper/2602.13600/) (2026)

## Built on by (most cited first, 2 of 2)

- [The Visual Insensitivity Gap: Diagnosing When Vision-Language Models Fail to Use Visual Evidence](https://mlsanity.com/paper/2609.00868/) (2026)
- [CLAP: Direct VLM-to-VLA Adaptation via Language-Action Grounding](https://mlsanity.com/paper/2607.08974/) (2026)

## Abstract (arXiv)

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.