PaperlayerView exact revision on arXiv

Source-linked AI summary

Reliability Challenges in Diffusion Vision-Language Models

Md. Atabuzzaman, Chris Thomas

arXiv:2609.01318v1cs.CVcs.CL

TL;DR

The paper addresses the limited understanding of reliability in diffusion-based vision-language models. It systematically benchmarks six dLVLMs against autoregressive baselines across four reliability dimensions and finds distinct, sometimes more severe, reliability patterns, including diffusion-specific hallucination signals and strong selection bias.

  • Problem

    Reliability properties of diffusion-based LVLMs, particularly their hallucination and bias patterns, remain largely uncharacterized.

  • Method

    The study benchmarks six dLVLMs against competitive autoregressive baselines across object hallucination, open-ended hallucination, demographic bias, and selection bias.

  • Results

    dLVLMs show distinct reliability profiles: reversed yes-bias, competitive hallucination with degraded linguistic quality, near-zero accuracy for some underrepresented racial groups, and larger multiple-choice length-bias gaps.

  • Takeaways & Limitations

    Reliability evaluation for dLVLMs must separately assess hallucination, linguistic quality, demographic bias, option length, and diffusion-specific denoising signals.

  • Takeaways & Limitations

    Isolating the generative mechanism from vision encoders and model scale requires more extensive ablations, while linguistic-quality judgments lack human-annotation validation.

Abstract

from arXiv · show

Diffusion-based Large Vision-Language Models (dLVLMs) have recently emerged as a compelling alternative to autoregressive (AR) LVLMs, offering advantages in parallel decoding, bidirectional context, and controllable generation. Despite rapid progress, their reliability properties remain largely uncharacterized. We present the first systematic reliability evaluation of hallucination and bias in dLVLMs, benchmarking six diffusion models against competitive AR baselines across four dimensions. Our key findings are: (1) dLVLMs reverse the yes-bias of AR models in binary visual queries; (2) they achieve competitive hallucination rates yet exhibit degraded linguistic quality; (3) they collapse to near-zero accuracy on underrepresented racial groups with opposite-polarity gender bias; and (4) they exhibit accuracy collapse in multiple-choice settings when the correct option is shorter than its distractors, associated with a length prior that emerges at the first denoising step. Tokens committed at late denoising steps with low confidence further correlate with hallucinated content, pointing to a mechanistic signal unique to diffusion generation. These patterns vary across model families, suggesting reliability is shaped by the generative paradigm together with training data.

1 Introduction

This paper asks whether diffusion-based generation changes the reliability failures known in autoregressive vision-language models. It introduces the first systematic evaluation of dLVLM reliability across hallucination, demographic bias, and selection bias.

  • 1 Introduction: Diffusion-based LVLMs generate through iterative denoising, enabling parallel decoding, bidirectional context, and iterative refinement.This contrasts with autoregressive left-to-right token generation.
  • 1 Introduction: The study examines whether diffusion generation alters hallucination and bias patterns documented in autoregressive models.The evaluation targets object and open-ended hallucination, demographic disparities, and multiple-choice selection bias.
  • 1 Introduction: Six dLVLMs are benchmarked against competitive autoregressive baselines across four reliability dimensions.The dimensions are object hallucination, open-ended hallucination, demographic bias, and option-length selection bias.
  • 1 Introduction: dLVLMs show reliability profiles distinct from autoregressive models, including more severe length bias and separable linguistic-quality and hallucination failures.The patterns vary systematically across model families.

2 Related Work

Prior work established discrete diffusion language models and their multimodal extensions, while hallucination and bias evaluations had primarily focused on autoregressive LVLMs. This paper connects those lines of work through diffusion-specific reliability analysis.

  • Diffusion Language Models: Discrete diffusion language models generate by iteratively denoising masked token sequences rather than decoding strictly left to right.LLaDA and Dream are described as important diffusion-language backbones.
  • Diffusion-based LVLMs: Recent multimodal diffusion models extend these backbones with vision encoders, multimodal instruction tuning, reasoning, and diffusion-specific training strategies.Examples include LaViDa, LLaDA-V, MMaDA, Dimple, and Dream-VL.
  • Hallucination and Bias in LVLMs: Existing reliability research has extensively studied hallucination and bias in autoregressive LVLMs.The cited work covers object hallucination, demographic bias, and other systematic output preferences.
  • Hallucination and Bias in LVLMs: Qualitative diffusion-model analysis links hallucinated objects to late denoising commitments and low token confidence.In the living-room example, clock, vase, and posters are committed around steps 126–128, unlike the grounded lantern at step 40.

3 Preliminary

dLVLMs generate responses by iteratively unmasking tokens from a fully masked sequence. At each denoising step, they predict masked positions, assign confidence scores, and commit selected tokens to final values.

  • Diffusion-Based Large Vision-Language Models: The forward process corrupts a clean response into a fully masked sequence, while inference reverses this through K denoising steps.Generation begins from an entirely masked response.
  • Diffusion-Based Large Vision-Language Models: At each step, the denoiser predicts all masked positions simultaneously using full bidirectional attention.Every position can attend to all other positions during prediction.
  • Diffusion-Based Large Vision-Language Models: The model assigns each token a confidence score based on its maximum vocabulary softmax probability and selects positions to unmask.These scores determine which masked tokens are updated at each step.
  • Diffusion-Based Large Vision-Language Models: A token’s commit step is the first denoising step at which it receives its final non-mask value.Commit step and confidence form the basis of the paper’s mechanistic hallucination analysis.

4 Experiments

Across hallucination, linguistic quality, demographic bias, and selection bias, dLVLMs show reliability patterns that differ substantially across model families and from AR baselines. The experiments identify competitive hallucination resistance alongside linguistic, demographic, and length-related vulnerabilities, plus diffusion-specific signals associated with hallucination.

  • Hallucination: dLVLMs achieve competitive object and open-ended hallucination resistance, with some models surpassing older AR baselines.LaViDa-Dream surpasses InstructBLIP by 7.03 and 8.23 percentage points on Popular and Adversarial POPE, while Dream-VL records CHAIRI of 11.69% and CHAIRS of 23.21%.
  • Hallucination: dLVLMs reverse the pronounced yes-bias of older AR models, typically producing affirmative responses in only 35–45% of POPE cases.Dream-VL reaches up to 100% precision on Random, but reduced recall accompanies the lower Yes% values; MMaDA-MixCoT is an exception on Adversarial.
  • Linguistic quality: Reducing denoising steps primarily damages linguistic quality, with LaViDa-LLaDA’s overall error rising from 3.0% to 87.6% and Dream-VL’s from 1.2% to 43.8%.The reported degradation includes LaViDa-LLaDA grammatical error increasing from 1.0% to 60.0% and repetition from 2.4% to 64.6%.
  • Linguistic quality: 13.8% to 2.0%: AR-style decoding sharply reduces Dimple’s linguistic error, while CHAIR improves only modestly from 19.59% to 16.23%.LaViDa-LLaDA shows the same dissociation, improving linguistic error from 3.0% to 1.2% while CHAIR changes from 18.43% to 17.50%.
  • Mechanistic hallucination analysis: Commit step and confidence distinguish hallucinated from grounded object tokens, with commit-step ROC-AUCs of 0.699 and 0.667 across the two models.Hallucinated objects commit 22–26 steps later on a 128-step budget and at lower confidence; commit step is the stronger single feature and has no AR analogue.
  • Demographic bias: Underrepresented racial groups show severe, backbone-dependent failures, while gender bias reverses polarity across dLVLM families and is highly sensitive to image padding.LaViDa-Dream routes 63–84% of Latino Hispanic and Southeast Asian faces into East Asian, whereas LaViDa-LLaDA routes them into Indian and White; gender gaps include +13.48, +10.26, and −23.34.
  • Selection bias: In Shorter Correct MCQs, accuracy collapses to 2.80% for LaViDa-Dream and 6.40% for LaViDa-LLaDA on CUB without class names, while Longer Correct accuracy reaches 68–99%.At least 90.9% of examples have a step-0 prediction selecting the longer distractor and never revised during denoising; LLaDA-V and Dimple show partial mitigation.
  • Selection bias: AR models remain more robust to misleading option length, typically achieving 19–51% on CUB and 16–59% on Dogs in Shorter Correct conditions with class names.Qwen2.5-VL-7B reaches 50.70% on CUB Shorter Correct with class names, while class names provide only modest or negligible gains for dLVLMs.

5 Conclusion

This study provides the first systematic reliability evaluation of dLVLMs against competitive AR baselines across four dimensions. It finds distinct, sometimes more severe, reliability challenges whose patterns vary across diffusion backbones.

  • The evaluation benchmarks six diffusion models against competitive AR baselines across hallucination, demographic bias, and selection bias dimensions.
  • dLVLMs show no-bias tendencies in binary visual queries, competitive open-ended hallucination, degraded linguistic quality, severe demographic disparities, and larger MCQA length-bias gaps than AR models.The length-bias gap is associated with a one-shot prior established at the first denoising step.
  • Late-committed, low-confidence tokens correlate with hallucinated content, achieving ROC-AUC up to 0.699 on CHAIR object tokens across two backbones.This signal is unique to diffusion generation.
  • Reliability patterns vary across diffusion backbones, with the generative paradigm contributing alongside training data and scale.

Limitations

The study’s conclusions are bounded by incomplete causal isolation, judge-based linguistic evaluation, and coverage of only four reliability dimensions.

  • More extensive ablations are needed to isolate the generative mechanism from confounds such as vision encoders and model scale.The study includes controlled comparisons but does not fully separate these factors.
  • Linguistic quality evaluation relies on GPT-4o-mini as a judge and lacks validation against human annotations.
  • The evaluation covers four reliability dimensions and excludes failure modes such as toxicity, sycophancy, and temporal reasoning.

A Appendix

The appendix provides additional hallucination examples and lists analyses of attention, demographic bias, selection bias, and linguistic quality. Its qualitative examples contrast fluent captions containing absent objects with grounded scene content.

  • Qualitative hallucination examples: In the kitchen example, the caption includes bowl, cup, and vase even though these objects are absent from the image.The description otherwise depicts a person stirring noodles in a kitchen.
  • Qualitative hallucination examples: The kitchen caption remains fluent and detailed despite introducing objects not present in the visual scene.The description also mentions a green shirt, pots and pans, and a low-angle view.
  • Qualitative hallucination examples: In the cat-and-laptop example, the caption describes an office scene containing a whiteboard and clock alongside the cat, laptop, monitor, and mouse.The figure identifies clock and unexpected as hallucinated tokens.

A.1 Additional Qualitative Hallucination Analysis

Additional qualitative examples across kitchen and indoor-desk scenes show hallucinated tokens being committed late and with low confidence, while grounded tokens are committed earlier with high confidence.

  • LaViDa-Dream: In the kitchen scene, bowl, cup, and vase are hallucinated at steps 127–128 with near-zero confidence, unlike grounded tokens committed by step 31.Grounded examples include person at step 10, kitchen at step 12, stirring at step 17, and noodles at step 31.
  • LaViDa-LLaDA: In the cat-and-laptop scene, clock is hallucinated at step 96 with low confidence, whereas cat, laptop, and keyboard are committed earlier with high confidence.The examples span LaViDa-LLaDA and LaViDa-Dream across distinct image types and scene categories.

A.2 Attention-Based Analysis

Attention-based analysis finds little separation between hallucinated and grounded tokens, while commit steps remain strongly discriminative in the examined setting.

  • Attention comparison: Mean image-patch attention is identical for hallucinated and grounded tokens at the examined layers, measuring 0.0005 versus 0.0005.Attention entropy was also similar, at 4.41 versus 4.40, with peak attention uninformative.
  • Commit-step comparison: Grounded sofa is committed at step 19, while hallucinated clock and vase are committed at steps 127 and 126, respectively.The passage presents these commit-step differences as consistent with the commit-step hypothesis.

A.3 Demographic Bias Evaluation Protocol

The demographic-bias protocol uses completion-style gender classification and letter-based seven-class race recognition, with deterministic mappings and conservative exact-match scoring.

  • Evaluation protocol: Gender outputs are exact-matched to male or female, while race answers select one of seven FairFace labels through lettered options A–G.Non-matching outputs count as incorrect, and both mappings are deterministic.

A.4 Demographic Bias Evaluation Results

Table 7 reports FairFace gender and race recognition results across models, racial groups, genders, and two face-crop padding settings. It also includes female–male accuracy gaps and per-model accuracy changes between padding settings.

  • Table 7 evaluates gender and race recognition across all models, seven racial groups, two genders, and two face-crop padding settings.
  • The table reports the female–male accuracy gap and each model’s accuracy change across padding settings.
  • A detailed analysis of these FairFace results appears in Section 4.2.

A.5 Demographic Bias Confusion Summary

The section summarizes racial confusion patterns across FairFace models and contrasts them with option-length effects in multiple-choice evaluation. It also documents the structured procedure used to judge caption linguistic quality.

  • Demographic Bias Confusion: LaViDa-Dream absorbs 63% of Latino Hispanic faces and 84% of Southeast Asian faces into East Asian, while LLaVA-1.5-7B redirects 73% of Southeast Asian faces there.
  • Demographic Bias Confusion: LaViDa-Dream and LaViDa-LLaDA show different dominant top-2 racial confusion targets despite sharing the same training pipeline.
  • Length-Controlled Multiple Choice: On CUB with class names, Qwen2.5-VL-7B reaches 50.70% under Shorter Correct versus 11.50% for LaViDa-LLaDA, while LLaDA-V performs comparably to InternVL2.5-8B.
  • Length-Controlled Multiple Choice: Without class names, LaViDa-Dream falls to 2.80% on CUB and 4.00% on Dogs, whereas InternVL2.5-8B retains 37.80% and 38.17%.
  • Length-Controlled Multiple Choice: Most dLVLMs exceed 82% accuracy under Longer Correct, indicating substantially better performance when the correct option is longer than distractors.
  • Linguistic Quality Evaluation: Caption quality is judged by GPT-4o-mini using deterministic decoding, image-independent text evaluation, five explicit error types, and exactly five binary outputs.
Loading 2609.01318v1…