Don't Blink: Evidence Collapse during Multimodal Reasoning

Suresh Raghu; Satwik Pandey

cs.AIApr 5, 2026

Don't Blink: Evidence Collapse during Multimodal Reasoning

Authors: Suresh Raghu, Satwik Pandey

Organizations: Independent Researcher, New Delhi, India

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.

Figures & tables

Figure 1: Evidence collapse creates confident errors invisible to text-only monitoring. Two HallusionBench exemplars processed by GLM-4.6V-Flash on nearly identical map-comparison questions. Top (Failure): On an adversarially modified map, the model attempts visual grounding early but textual priors override image evidence mid-reasoning (red spans), producing a confident wrong answer with no attention recovery. Bottom (Success): On the unmodified map, textual priors align with the image and visual engagement is sustained, peaking at mid-generation. Entropy-derived confidence differs by only 3 percentage points (87.2% vs. 90.2%), yet the visual attention trajectories, summarized by $V_{\mathrm{thinking\_auc}}$ , are qualitatively different. Text-only confidence cannot separate these cases; cumulative visual attention can.

Figure 2: Entropy discrimination across nine model $\times$ dataset cells. Panel (a) shows answer-time entropy ( $\bar{H}_{\text{ans}}$ ) and panel (b) shows full-generation entropy ( $\bar{H}_{\text{full}}$ ). Values are Cohen’s $d$ (correct vs. incorrect). For panel (a), significant conventional discrimination appears only in GLM $\times$ HallusionBench. For panel (b), all nine cells are negative (incorrect answers have higher entropy), yielding a more consistent text-only baseline.

Figure 3: Architecture-dependent grounding layers. Layerwise AUROC for evidence localization on a held-out calibration set. Qwen models (DeepStack) peak in early layers (0–5, AUROC $\approx 0.75$ – $0.77$ ). GLM (single-stream early-fusion) peaks in layers 15–20. Grounding layer selection must be architecture-specific.

Figure 4: Three trajectory patterns of visual grounding across seven generation positions. Correct (solid) vs. incorrect (dashed) with 95% CI bands. Sustained deficit (HallusionBench): correct maintains higher

V

throughout. Crossover (MathVista): incorrect starts higher, then crosses below during reasoning. Moderate (MMMU_Pro): smaller, less consistent separations. Exact effect sizes and significance tests are reported in Appendix D .

Figure 5: Task-conditional interaction coefficients (

\beta_{EV}

). Forest plot with bootstrap 95% CIs. Negative

\beta_{EV}

= confident-but-blind penalty. MMMU_Pro shows significant penalties for GLM (

-0.41

) and Qwen-8B (

-0.61

); MathVista is near zero or positive. This sign flip explains why global linear fusion fails.

Figure 6: Regime map of confident-but-blind behavior (GLM representative). Predicted correctness surface from the continuous interaction model (Eq. 8 ) over entropy and visual-grounding z-scores for MMMU_Pro and MathVista. The Q2 region (confident + blind) is highlighted at

q=0.20

. On MMMU_Pro, Q2 corresponds to a penalty in predicted correctness; on MathVista, the same region is comparatively safe.

Transfer target	Model	$\Delta$ risk (pp)	Wins
MMMU_Pro	GLM-4.6V-Flash	$-1.94$	10/10
MMMU_Pro	Qwen3-VL-2B	$-0.66$	10/10
HallusionBench	GLM-4.6V-Flash	$-1.11$	10/10
MathVista	GLM-4.6V-Flash	$+1.87$	0/10
MathVista	Qwen3-VL-2B	$+1.86$	0/10
MathVista	Qwen3-VL-8B	$+1.20$	0/10

Table 1: Vision veto results across transfer directions. Negative

\Delta

risk = veto helps. The veto helps exactly where the interaction predicts danger (MMMU_Pro targets) and hurts where it predicts safety (MathVista targets).

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
max_tokens	8 192
temperature	0.7
top_p	0.8
top_k	20
top_logprobs	20
seed	42

Appendix

Table 2: Generation hyperparameters (shared across all models and datasets).

Model	Dataset	$n$	Start	Early	Mid	Late	EndThink	AnsStart	AnsEnd	$V_{\text{thinking\_auc}}$
GLM	HallusionBench	300	$-0.819$ ( $<0.001$ )	$-0.488$ ( $0.006$ )	$-0.729$ ( $<0.001$ )	$-0.638$ ( $<0.001$ )	$-0.561$ ( $<0.001$ )	$-0.510$ ( $<0.001$ )	$-0.561$ ( $<0.001$ )	$-0.888$ ( $<0.001$ )
GLM	MathVista	298	$+0.389$ ( $0.037$ )	$-0.132$ ( $0.392$ )	$+0.061$ ( $0.691$ )	$-0.096$ ( $0.483$ )	$+0.001$ ( $0.994$ )	$+0.038$ ( $0.785$ )	$-0.172$ ( $0.179$ )	$+0.007$ ( $0.958$ )
GLM	MMMU_Pro	287	$-0.032$ ( $0.827$ )	$-0.490$ ( $<0.001$ )	$-0.352$ ( $0.002$ )	$-0.401$ ( $<0.001$ )	$-0.289$ ( $0.025$ )	$-0.214$ ( $0.034$ )	$-0.110$ ( $0.360$ )	$-0.514$ ( $<0.001$ )
Qwen-2B	HallusionBench	251	$-0.362$ ( $0.033$ )	$-0.970$ ( $<0.001$ )	$-1.046$ ( $<0.001$ )	$-0.847$ ( $<0.001$ )	$-1.127$ ( $<0.001$ )	$-1.161$ ( $<0.001$ )	$-1.038$ ( $<0.001$ )	$-1.050$ ( $<0.001$ )
Qwen-2B	MathVista	239	$+0.336$ ( $0.032$ )	$-0.155$ ( $0.239$ )	$-0.186$ ( $0.180$ )	$-0.126$ ( $0.359$ )	$-0.199$ ( $0.145$ )	$-0.236$ ( $0.091$ )	$-0.133$ ( $0.334$ )	$-0.070$ ( $0.613$ )
Qwen-2B	MMMU_Pro	175	$-0.001$ ( $0.995$ )	$-0.222$ ( $0.179$ )	$-0.294$ ( $0.058$ )	$-0.194$ ( $0.241$ )	$-0.261$ ( $0.097$ )	$-0.337$ ( $0.031$ )	$-0.238$ ( $0.133$ )	$-0.204$ ( $0.221$ )

Appendix

Table 3: Trajectory discrimination statistics for total visual attention

V_{t}

. Each entry is Cohen’s

d

(incorrect

-

correct) with Welch

p

-value in parentheses.

Model	Train	Test	AUC (Entropy)	AUC (Entropy+Vision)	$\Delta$ AUC
GLM	HallusionBench	MMMU_Pro	0.712	0.647	-0.064
GLM	HallusionBench	MathVista	0.732	0.588	-0.143
GLM	MMMU_Pro	HallusionBench	0.755	0.770	+0.015
GLM	MMMU_Pro	MathVista	0.732	0.686	-0.045
GLM	MathVista	HallusionBench	0.755	0.676	-0.079
GLM	MathVista	MMMU_Pro	0.712	0.661	-0.051

Appendix

Table 4: All 18 transfer pairs for entropy-only vs. entropy+vision probes (AUC).

\Delta

AUC is computed as (entropy+vision)

-

(entropy-only).

Model	Train	Test	Q1 Error $(n)$	Q2 Error $(n)$	Q2 $-$ Q1	$\Delta>0$ Splits
GLM	HallusionBench	MMMU_Pro	0.058 (52)	0.333 (6)	+0.276	10/10
GLM	HallusionBench	MathVista	0.077 (52)	0.000 (8)	-0.077	0/10
GLM	MMMU_Pro	HallusionBench	0.033 (60)	– (0)	–	0/10
GLM	MMMU_Pro	MathVista	0.077 (52)	0.000 (8)	-0.077	0/10
GLM	MathVista	HallusionBench	0.033 (60)	– (0)	–	0/10
GLM	MathVista	MMMU_Pro	0.058 (52)	0.333 (6)	+0.276	10/10

Appendix

Table 5: Quadrant error-rate summary at

q=0.20

for all 18 directed transfers. Q1 is confident+grounded; Q2 is confident+blind. Entries show mean error rate with mean cell size

n

in parentheses across 10 splits.

Model	Dataset	Total	normal_stop	max_tokens	max_tokens_loop
Qwen3-VL-2B	HallusionBench	300	251	0	49
Qwen3-VL-2B	MathVista	300	239	20	41
Qwen3-VL-2B	MMMU_Pro	300	175	28	97
Qwen3-VL-8B	HallusionBench	300	299	0	1
Qwen3-VL-8B	MathVista	300	285	12	3
Qwen3-VL-8B	MMMU_Pro	300	241	53	6

Appendix

Table 6: Termination classification across all model

\times

dataset cells. Budget-exhausted combines max_tokens and max_tokens_loop (see § A.3 for definitions). Qwen3-VL-2B

\times

MMMU_Pro retains only 58% of samples after filtering.

Model	Dataset	Total Samples	Normal-Stop Samples	Correct (normal-stop)	Incorrect (normal-stop)
Qwen3-VL-2B	HallusionBench	300	251	213	38
Qwen3-VL-2B	MathVista	300	239	188	51
Qwen3-VL-2B	MMMU_Pro	300	175	120	55
Qwen3-VL-8B	HallusionBench	300	299	265	34
Qwen3-VL-8B	MathVista	300	285	223	62
Qwen3-VL-8B	MMMU_Pro	300	241	183	58

Appendix

Table 7: Sample counts after normal-stop filtering.

Model	Dataset	$N$	Mean $\bar{H}_{\text{full}}$ (correct)	Mean $\bar{H}_{\text{full}}$ (incorrect)	Cohen’s $d$	$p$ -value
Qwen3-VL-2B	HallusionBench	251	0.393	0.687	-1.46	$<10^{-12}$
Qwen3-VL-2B	MathVista	239	0.462	0.576	-0.64	$<10^{-3}$
Qwen3-VL-2B	MMMU_Pro	175	0.582	0.717	-0.54	$<10^{-3}$
Qwen3-VL-8B	HallusionBench	299	0.305	0.446	-1.01	$<10^{-5}$
Qwen3-VL-8B	MathVista	285	0.350	0.438	-0.61	$<10^{-3}$
Qwen3-VL-8B	MMMU_Pro	241	0.440	0.586	-0.90	$<10^{-8}$

Appendix

Table 8: Full-response entropy discrimination (

\bar{H}_{\text{full}}

Model	Dataset	$N$	Mean $\bar{H}_{\text{ans}}$ (correct)	Mean $\bar{H}_{\text{ans}}$ (incorrect)	Cohen’s $d$	$p$ -value
Qwen3-VL-2B	HallusionBench	251	0.137	0.249	-0.54	0.02
Qwen3-VL-2B	MathVista	239	0.263	0.278	-0.08	0.7
Qwen3-VL-2B	MMMU_Pro	175	0.289	0.441	-0.57	0.003
Qwen3-VL-8B	HallusionBench	299	0.395	0.389	+0.03	0.9
Qwen3-VL-8B	MathVista	285	0.266	0.285	-0.11	0.4
Qwen3-VL-8B	MMMU_Pro	241	0.321	0.419	-0.49	$<10^{-3}$

Appendix

Table 9: Answer-span entropy discrimination (

\bar{H}_{\text{ans}}

). Strong effects (

|d|>0.5

) are bolded.

Model	Dataset	$N$ (normal-stop)	$V_{\text{start}}$	$V_{\text{ans\_end}}$	$\Delta V$ (absolute)	$\Delta V$ (%)	Samples with $\Delta V>0$
Qwen3-VL-2B	MathVista	239	0.095	0.017	0.078	82.5	239/239 (100.0%)
Qwen3-VL-2B	HallusionBench	251	0.125	0.044	0.081	64.6	251/251 (100.0%)
Qwen3-VL-2B	MMMU_Pro	175	0.082	0.012	0.070	85.4	175/175 (100.0%)
Qwen3-VL-8B	MathVista	285	0.115	0.015	0.100	86.8	285/285 (100.0%)
Qwen3-VL-8B	HallusionBench	299	0.172	0.038	0.134	78.0	299/299 (100.0%)
Qwen3-VL-8B	MMMU_Pro	241	0.099	0.009	0.090	90.9	241/241 (100.0%)

Appendix

Table 10: Visual attention decay (

V

) across all nine model

\times

dataset cells.

N

denotes normal-stop samples.

Model	Dataset	$N$ (normal-stop)	$A_{\text{bbox,start}}$	$A_{\text{bbox,ans\_end}}$	$\Delta A_{\text{bbox}}$ (absolute)	$\Delta A_{\text{bbox}}$ (%)	Samples with $\Delta A_{\text{bbox}}>0$
Qwen3-VL-2B	MathVista	239	0.017	0.003	0.014	82.0	191/191 (100.0%)
Qwen3-VL-2B	HallusionBench	251	0.015	0.007	0.009	56.8	250/251 (99.6%)
Qwen3-VL-2B	MMMU_Pro	175	0.019	0.003	0.016	84.2	116/116 (100.0%)
Qwen3-VL-8B	MathVista	285	0.024	0.003	0.020	85.7	220/223 (98.7%)
Qwen3-VL-8B	HallusionBench	299	0.024	0.005	0.019	77.8	299/299 (100.0%)
Qwen3-VL-8B	MMMU_Pro	241	0.029	0.003	0.026	89.7	164/164 (100.0%)

Appendix

Table 11: Evidence attention decay (

A_{\text{bbox}}

) across all nine model

\times

dataset cells.

N

denotes normal-stop samples.

A_{\text{bbox}}

decay-positive counts use valid bbox samples only (denominator shown in the last column).

Feature	GLM	Qwen-2B	Qwen-8B
$V_{\text{thinking\_auc}}$	$+54/-6$	$+50/-10$	$+58/-2$
$V_{\text{ans\_end}}$ †	$+10/-8$	$+16/-2$	$+18/-0$
$\bar{H}_{\text{full}}$ ‡	$+6/-0$	$+2/-4$	$+0/-6$
$L$ §	$+0/-18$	$+0/-18$	$+0/-18$
$\mathrm{RVAR}_{\text{ans\_end}}$ †	$12/18$	$0/18$	$0/18$
$V_{\text{reengagement}}$	$+22/-38$	$+58/-2$	$+8/-52$

Appendix

Table 12: Coefficient-sign consistency across transfer evaluation. Unless noted otherwise, entries report positive-sign fits out of 60 per model (6 transfers

\times

10 splits).

Metric	Value
$r_{\text{overall}}$	$-0.43$
$r_{\text{GLM}}$	$-0.41$
$r_{\text{Qwen-2B}}$	$-0.59$
$r_{\text{Qwen-8B}}$	$-0.55$
$N_{\mathrm{PI\text{-}dom}}/N_{\mathrm{transfer}}$	$10/18$
$N_{\beta>0}/N_{\mathrm{split}}$	$148/180$

Appendix

Table 13: Length-confound diagnostics for

V_{\text{thinking\_auc}}

Selective Prediction for VLMs

See all 12 papers

Are Reasoning Vision-Language Models Robust to Semantic Visual Distractions?
Jun 8, 2026Yizheng Sun, Mochuan Zhan, Yanan Ma +10VLM Evaluation VLM Robustness
University of Manchester · Marex · Imperial College London
Reasoning Dynamics and the Limits of Monitoring Modality Reliance in Vision-Language Models
Apr 16, 2026Danae Sánchez Villegas, Samuel Lewis-Lim, Nikolaos Aletras +1CoT Faithfulness Vision-Language Models
University of Copenhagen, Department of Computer Science · University of Sheffield, School of Computer Science
Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap
Apr 17, 2026Yige Xu, Yongjie Wang, Zizhuo Wu +3Vision-Language Models VLM Adaptation
College of Computing and Data Science, Nanyang Technological University, Singapore · Alibaba-NTU Global e-Sustainability CorpLab (ANGEL) · Tongyi Lab, Alibaba Group, China

Jun 8, 2026J/K move · Enter open · S save

Are Reasoning Vision-Language Models Robust to Semantic Visual Distractions?

Yizheng Sun, Mochuan Zhan, Yanan Ma, Jia Tong See+9

University of Manchester · Marex · Imperial College London

VLM Evaluation VLM Robustness VLM Reasoning Visually Grounded Reasoning

Open paper AskPDF arXiv

Reasoning Vision-Language Models (VLMs) achieve strong performance on complex multimodal tasks, but reliable real-world application requires handling visual inputs that are messier than clean, curated benchmarks. Existing works mainly evaluate such reliability of VLMs through input corruptions, such as noise, blur and weather effects, which make visual evidence harder to perceive. This leaves a critical reliability failure mode underexplored: a model may perceive the evidence correctly, yet reason from plausible but irrelevant and distracting evidence and propagate this mistake to its final answer. To address this gap, we introduce Distract-Bench, a benchmark for evaluating VLM robustness to semantic visual distractions, defined as meaningful but task-irrelevant visual cues added to inputs while preserving the ground-truth answer. We comprehensively evaluate eight leading open-source and two closed-source VLMs across conventional vision corruptions and Distract-Bench. Our results show that Distract-Bench exposes a robustness failure distinct from vision corruptions: reasoning VLMs largely track their non-reasoning base models under perceptual degradation, but show consistently lower robustness to semantic distractions. Further analysis shows that these distractions often enter the reasoning process of VLMs, are treated as evidence, and lead to incorrect answers. Together, these findings reframe robustness evaluation for reasoning VLMs, shifting the focus from degraded perception to distractions for reliable real-world visual reasoning. Our data and code are available at https://github.com/Yizheng-Sun/Distract-Bench.

Abstract

Selective Prediction for VLMs

Explore similar work

Are Reasoning Vision-Language Models Robust to Semantic Visual Distractions?

Reasoning Dynamics and the Limits of Monitoring Modality Reliance in Vision-Language Models

Do Vision-Language Models Truly Perform Vision Reasoning? A Rigorous Study of the Modality Gap