Benchmarking Deflection and Hallucination in Large Vision-Language Models

Nicholas Moratelli; Christopher Davis; Leonardo F. R. Ribeiro; Bill Byrne; Gonzalo Iglesias

cs.CLApr 13, 2026

Benchmarking Deflection and Hallucination in Large Vision-Language Models

Authors: Nicholas Moratelli, Christopher Davis, Leonardo F. R. Ribeiro, Bill Byrne, Gonzalo Iglesias

Organizations: University of Modena and Reggio Emilia · Amazon AGI · University of Cambridge

Abstract

Large Vision-Language Models (LVLMs) increasingly rely on retrieval to answer knowledge-intensive multimodal questions. Existing benchmarks overlook conflicts between visual and textual evidence and the importance of generating deflections (e.g., Sorry, I cannot answer...) when retrieved knowledge is incomplete. These benchmarks also suffer from rapid obsolescence, as growing LVLM training sets allow models to answer many questions without retrieval. We address these gaps with three contributions. First, we propose a dynamic data curation pipeline that preserves benchmark difficulty over time by filtering for genuinely retrieval-dependent samples. Second, we introduce VLM-DeflectionBench, a benchmark of 2,775 samples spanning diverse multimodal retrieval settings, designed to probe model behaviour under conflicting or insufficient evidence. Third, we define a fine-grained evaluation protocol with four scenarios that disentangle parametric memorization from retrieval robustness. Experiments across 20 state-of-the-art LVLMs indicate that models usually fail to deflect in the presence of noisy or misleading evidence. Our results highlight the need to evaluate not only what models know, but how they behave when they do not, and serve as a reusable and extensible benchmark for reliable KB-VQA evaluation. All resources will be publicly available upon publication.

Figures & tables

Figure 1: Overview of VLM-DeflectionBench . Top: LVLMs often hallucinate instead of abstaining when context is misleading. Bottom: VLM-DeflectionBench evaluates calibration across four scenarios- Parametric , Oracle , Realistic , and Adversarial to test whether models align their behavior with available knowledge.

Benchmark	Evaluation Scope				Evidence Configuration
Benchmark	Hallucination	Deflection	Scenarios	Dynamic	MM-Query	MM-Context	Pre-retrieved	Negatives
OK-VQA Marino et al. (2019)	$\times$	$\times$	1	$\times$	$\checkmark$	$\times$	$\times$	$\times$
A-OKVQA Schwenk et al. (2022)	$\times$	$\times$	1	$\times$	$\checkmark$	$\times$	$\times$	$\times$
E-VQA Mensink et al. (2023)	$\times$	$\times$	1	$\times$	$\checkmark$	$\times$	$\times$	$\times$
InfoSeek Chen et al. (2023)	$\times$	$\times$	1	$\times$	$\checkmark$	$\times$	$\times$	$\times$
WebQA Chang et al. (2022)	$\times$	$\times$	1	$\times$	$\times$	$\checkmark$	$\checkmark$	$\checkmark$

Table 1: Comparison of VLM-DeflectionBench with current KB-VQA benchmarks. Our benchmark is the only one that explicitly evaluates hallucination and deflection across four retrieval scenarios: parametric (no additional context), oracle (gold evidence only), realistic (mixed positive and negative evidence), and adversarial (only negative contexts).

Figure 2: Pipeline of VLM-DeflectionBench . Starting from 6 benchmarks, we apply parametric filtering to remove query-solvable samples ( stage I ), retrieve negative (multimodal) contexts via different indices ( stage II ), then perform oracle filtering to eliminate false positive contexts (where gating models fail) and unreliable negative contexts (where gating models succeed) ( stage III ).

Source	Samples	Filtered (%)	Query	Context
InfoSeek	1646	5.6%	Multi	Textual
WebQA	826	16.6%	Textual	Multi
E-VQA	185	4.9%	Multi	Textual
MMDocRAG	58	1.5%	Textual	Multi
MRAG-Bench	45	3.3%	Multi	Visual
ViQuAE	15	1.2%	Multi	Textual

Table 2: Per-source breakdown of our dataset. We report the number of retained samples, filtering rate, and dominant modality for each source. The “Filtered” column indicates the percentage of samples retained from the original source after applying our filtering criteria.

Statistic	Value
Total samples	2,775
Unique questions	1,246
Unique query images	2,717
Avg. question length (words)	12.5
Avg. gold contexts per sample	1.9
Avg. negative contexts per sample	11.8

Table 3: Global statistics of VLM-DeflectionBench . The dataset combines gold and negative contexts across text and vision, ensuring multimodal coverage and non-triviality.

Model	Parametric			Oracle			Realistic			Adversarial
Model	Acc ↑	Defl ↑	Hall ↓	Acc ↑	Defl ↓	Hall ↓	Acc	Defl ↓	Hall ↓	Acc ↑	Defl ↑	Hall ↓
Open Source Models
LLaVA-OneVision	4.7	4.0	91.3	54.9	3.5	41.6	34.0	10.8	55.2	4.5	14.5	81.0
MiniCPM-V-4.5-8B	4.6	3.2	92.2	65.3	8.9	25.8	44.3	16.3	39.4	1.9	46.2	51.9
Keye-1.5-8B	3.0	23.9	73.1	63.4	9.3	27.3	47.6	11.9	40.5	2.8	43.4	53.8
InternVL3-9B	3.1	9.6	87.3	61.2	7.7	31.1	45.1	10.3	44.6	2.3	35.9	61.8

Table 4: Performance of LVLMs across evaluation scenarios. We report Accuracy (Acc), Deflection Rate (Defl), and Hallucination Rate (Hall). The number of negatives is fixed at 2 to match the average number of gold contexts per sample, ensuring a balanced ratio of positive and negative evidence. ⋆ marks selected gating models.

Model	Context		Acc ↑	Defl ↓	Hall ↓
Model	Positive	Negative	Acc ↑	Defl ↓	Hall ↓
Keye-1.5-8B	Textual	MM	80.1	3.3	16.6
		Visual	82.1	2.7	15.2
		Textual	82.5	3.3	14.2
	Visual	MM	1.0	82.5	16.5
		Visual	9.4	19.2	71.4

Table 5: Effect of negative-context modality in the Realistic scenario. Results reveal a language-over-vision bias: models handle visual distractors well when positives are textual, but collapse when textual distractors interfere with visual positives, demonstrating that misleading text systematically overrides correct visual evidence.

Figure 3: Effect of distractor quantity in the Realistic scenario, including Oracle results (0 distractors) for reference. Accuracy declines and hallucination rises as negatives increase, while deflection shows limited improvement.

Model	Severity	Realistic			Adversarial
Model	Severity	Acc ↑	Defl ↓	Hall ↓	Acc ↑	Defl ↑	Hall ↓
Keye-1.5-8B	None	58.7	4.7	36.6	4.1	21.0	74.9
	Soft	56.9	8.8	34.3	3.2	37.2	59.7
	Moderate	56.0	9.8	34.2	3.1	43.8	53.1
	Severe	36.2	42.9	20.9	1.5	77.2	21.3
Ovis2-16B	None	57.9	1.0	41.1	4.6	5.7	89.7

Table 6: Effect of prompt strictness across Realistic and Adversarial scenarios for four representative LVLMs. Higher strictness increases deflection rates and reduces hallucination while decreasing accuracy in realistic scenarios.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Comparison of category distributions before and after curation. (Left) Aggregate distribution of the six original source benchmarks. (Middle) Final distribution in VLM-DeflectionBench . (Right) Overlay showing relative shifts across categories.

Figure 5: Performance on Parametric Scenario and dataset size as models are incrementally added to the filtering pipeline. Convergence is reached after four models, with over 90% of samples removed.

Model	Agreement (%)	Cohen’s $\kappa$
Keye-1.5-8B	91.1	0.90
Ovis2-16B	92.4	0.91
Mistral-3.1-24B	94.0	0.92
Claude-Opus-4	93.0	0.92
Average	92.6	0.91

Appendix

Table 7: Human validation of GPT-4o judgments (

100

samples per model, balanced across scenarios). We report raw agreement (%) and Cohen’s

\kappa

. Averages are computed across the four models.

Model	Source	# Samples	Parametric			Oracle			Realistic			Adversarial
Model	Source	# Samples	Acc ↑	Defl ↑	Hall ↓	Acc ↑	Defl ↓	Hall ↓	Acc ↑	Defl ↓	Hall ↓	Acc ↑	Defl ↑	Hall ↓
Keye-1.5-8B	Infoseek	1,646	1.6	17.1	81.3	56.7	11.7	31.6	47.1	10.8	42.1	0.9	38.4	60.7
	WebQA	826	5.6	28.6	65.9	79.5	4.4	16.1	70.7	8.6	20.7	6.5	52.7	40.8
	Encyclopedic-VQA	185	3.8	27.0	69.2	77.3	11.9	10.8	73.5	9.7	16.8	2.2	66.5	31.4
	MMDocRAG	58	0	31.0	69.0	69.0	0	31.0	60.3	0	39.7	15.5	12.1	72.4
	MRAG-Bench	45	11.1	4.4	84.4	37.8	8.9	53.3	35.6	6.7	57.8	8.9	26.7	64.4

Appendix

Table 8: Breakdown of performance across individual sources for four representative LVLMs. Results are reported in terms of accuracy, deflection, and hallucination under parametric, oracle, realistic, and adversarial settings. The table highlights heterogeneous dataset difficulty, systematic effects of external grounding on hallucination, and consistent relative model ranking across sources.

Model	Severity	Parametric			Oracle
Model	Severity	Acc ↑	Defl ↓	Hall ↓	Acc ↑	Defl ↑	Hall ↓
Keye-1.5-8B	None	4.6	4.5	90.9	68.0	2.9	29.1
	Soft	4.1	11.4	84.5	65.8	6.9	27.3
	Moderate	3.1	21.4	75.5	64.8	9.4	25.8
	Severe	0.6	82.3	17.1	37.5	47.0	15.5
Ovis2-16B	None	5.3	9.6	85.1	63.0	1.2	35.8

Appendix

Table 9: Deflection prompt strictness effects in the Parametric and Oracle scenarios across four LVLMs. Increasing strictness successfully raises deflection rates and reduces hallucination in strong models, but comes at the cost of significant accuracy degradation, revealing fundamental trade-offs in reliability tuning.

Panel of Figure 6 — Figure 6: System prompt used in our RAG experiments showing the initial instructions given to the model.

Panel of Figure 7 — Figure 7: Soft prompt variant encouraging models to answer based on available information with optional deflection when necessary.

Panel of Figure 8 — Figure 8: Moderate prompt variant instructing models to answer only when sufficiently confident in accuracy, with deflection preferred over uncertain responses.

Panel of Figure 9 — Figure 9: Severe prompt variant requiring models to answer only when completely certain, mandating deflection for any uncertainty or doubt.

Figure 10: Example from Encyclopedic-VQA ( Oracle ). All models are given gold textual evidence about the coral snake’s diet. Ovis2-16B and Claude-Opus-4 ground their answers correctly, while Mistral-3.1-24B deflects despite sufficient information, illustrating conservative behavior under clear evidence.

Figure 11: Example from Encyclopedic-VQA ( Oracle ). Gold text specifies the habitat; Ovis2-16B answers correctly while Mistral-3.1-24B and Claude-Opus-4 over-deflect under gold evidence.

Panel of Figure 12 — Figure 12: Example from MRAG-Bench ( Oracle , visual gold). Ovis2-16B hallucinates a wrong car model; Mistral-3.1-24B and Claude-Opus-4 abstain. A grounding failure with gold images.

Panel of Figure 13 — Figure 13: Example from MRAG-Bench ( Oracle , visual gold). Mistral-3.1-24B identifies the breed correctly; Ovis2-16B hallucinates; Claude-Opus-4 over-deflects. Calibration varies despite gold images.

Panel of Figure 14 — Figure 14: Example from MMDocRAG ( Oracle ). Gold text supports exact numeric extraction: Ovis2-16B is correct; Mistral-3.1-24B over-deflects; Claude-Opus-4 hallucinates outdated figures.

Figure 15: Example from WebQA ( Oracle , visual gold). Counting letters from the gold image: Ovis2-16B and Claude-Opus-4 hallucinate, Mistral-3.1-24B over-deflects.

Panel of Figure 16 — Figure 16: Example from WebQA ( Oracle , visual gold). With two gold images, Ovis2-16B correctly infers the taller object; Mistral-3.1-24B and Claude-Opus-4 over-deflect.

VLM Evaluation

Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.

See all 989 papers

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models
May 14, 2026Xiyu Ren, Zhaowei Wang, Yiming Du +11VLM Evaluation Memory-Augmented VLMs
CSE Deparment, HKUST · CUHK · OmniMemory (Shenzhen) Intelligent Technology Co., Ltd. +1
MMGist: A Comprehensive Multimodal Benchmark for 2027
Jun 21, 2026Wenzhen Yuan, Jiacheng Ruan, Wutao Xiong +3VLM Evaluation Large Vision-Language Models
1Shanghai Jiao Tong University · 2Sichuan University
Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?
May 21, 2026Zixuan Lan, Luzhe Sun, Matthew R. Walter +1VLM Evaluation VLM Interpretability
University of Chicago · Toyota Technological Institute at Chicago · Stony Brook University

May 14, 2026J/K move · Enter open · S save

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models

Xiyu Ren, Zhaowei Wang, Yiming Du, Zhongwei Xie+10

CSE Deparment, HKUST · CUHK · OmniMemory (Shenzhen) Intelligent Technology Co., Ltd.+1

VLM Evaluation Memory-Augmented VLMs Multimodal Memory Vision-Language Models Persistent Memory for Language Models Long-Context VLMs+2

Open paper AskPDF arXiv

Memory is essential for large vision-language models (LVLMs) to handle long, multimodal interactions, with two method directions providing this capability: long-context LVLMs and memory-augmented agents. However, no existing benchmark conducts a systematic comparison of the two on questions that genuinely require multimodal evidence. To close this gap, we introduce MEMLENS, a comprehensive benchmark for memory in multimodal multi-session conversations, comprising 789 questions across five memory abilities (information extraction, multi-session reasoning, temporal reasoning, knowledge update, and answer refusal) at four standard context lengths (32K-256K tokens) under a cross-modal token-counting scheme. An image-ablation study confirms that solving MEMLENS requires visual evidence: removing evidence images drops two frontier LVLMs below 2% accuracy on the 80.4% of questions whose evidence includes images. Evaluating 27 LVLMs and 7 memory-augmented agents, we find that long-context LVLMs achieve high short-context accuracy through direct visual grounding but degrade as conversations grow, whereas memory agents are length-stable but lose visual fidelity under storage-time compression. Multi-session reasoning caps most systems below 30%, and neither approach alone solves the task. These results motivate hybrid architectures that combine long-context attention with structured multimodal retrieval. Our code is available at https://github.com/xrenaf/MEMLENS.

Abstract

VLM Evaluation

Explore similar work

MemLens: Benchmarking Multimodal Long-Term Memory in Large Vision-Language Models

MMGist: A Comprehensive Multimodal Benchmark for 2027

Seeing without Looking: Do Vision-Language Benchmarks Really Test Vision?