PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Language Priors: Diagnosing and Fixing Visual-Origin Hallucinations in Multimodal LLM

Peiyang Xu, Xiaopei Zhu, Jun Zhu, Xiaolin Hu

arXiv:2609.00231v1cs.CV

TL;DR

Object hallucination in MLLMs is commonly attributed to language priors, but this paper identifies visual-origin hallucination from visual feature extraction errors and image–text misalignment. It diagnoses this mechanism with alignment and attention analyses, then introduces AHAF and ACFT, which achieve state-of-the-art hallucination mitigation using 0.9% of COCO data and zero inference overhead.

  • Problem

    Object hallucination remains a critical obstacle, and existing language-prior explanations do not fully account for hallucinations rooted in visual feature extraction and image–text misalignment.

  • Method

    ACFT uses AHAF to apply targeted adversarial perturbations that flip hallucination attributes and create aligned positive–negative pairs for contrastive fine-tuning.

  • Results

    ACFT achieves state-of-the-art hallucination mitigation across multiple models and benchmarks with only 0.9% of COCO data and zero inference overhead.

  • Takeaways & Limitations

    Visual-origin hallucination is a distinct mechanism supported by lower cross-modal alignment and inverted visual attention, while AHAF also probes fragile MLLM visual representations.

  • Takeaways & Limitations

    The method includes empirically determined hyperparameters, including λ.

Abstract

from arXiv · show

Existing research on object hallucination in multimodal large language models (MLLMs) predominantly attributes the problem to language priors such as over-reliance on textual co-occurrence statistics. We challenge this view by presenting quantitative evidence for a complementary, under-explored cause: visual-origin hallucination, where hallucinations arise from incorrect visual feature extraction and misalignment between image and text embeddings. Through cosine similarity analysis and Smooth Grad-CAM entropy measurements, we show that hallucinated samples exhibit systematically lower image-text similarity (average 0.158 vs. -0.122) and inverted attention patterns, where attention is dispersed when the target object is present but wrongly concentrated when it is absent. Guided by this diagnosis, we propose Adversarial Contrastive Fine-Tuning (ACFT). ACFT uses an Adversarial Hallucination Attribute Flipping (AHAF) procedure, involving minimal, targeted adversarial perturbations that flip an image's hallucination attribute, to construct perfectly aligned positive-negative pairs, which are then used for contrastive fine-tuning. AHAF simultaneously serves as a diagnostic probe, revealing that MLLM visual representations lie dangerously close to hallucination decision boundaries. Requiring only 0.9% of the COCO dataset and adding zero inference overhead, ACFT achieves state-of-the-art performance on POPE, MME, and four description-level hallucination benchmarks across LLaVA, MiniGPT-4, and Qwen2.5-VL. Code is available at https://github.com/zxp555/ACFT_MM

1 Introduction

The paper identifies visual-origin hallucination as a complementary mechanism to language-prior explanations, linking hallucinations to visual feature errors and image–text misalignment. It diagnoses this mechanism quantitatively and proposes ACFT, which uses targeted adversarial pairs for data-efficient mitigation.

  • Object hallucination occurs when MLLMs falsely perceive absent objects or ignore objects present in images.
  • Short-output questions can expose visual-origin hallucination, which arises from incorrect visual feature extraction rather than only linguistic biases.
  • 0.158 for correct cases versus −0.122 for hallucinated cases: hallucinated samples show lower image–text embedding cosine similarity.
  • 5.1% higher entropy when objects are present and 6.2% lower entropy when absent indicate inverted attention patterns in hallucinated cases.Attention is too dispersed when the target is present and wrongly concentrated on an irrelevant region when it is absent.
  • Visual-encoder interventions change POPE accuracy from 0.842 to 0.739–0.822 with degradations and to 0.864 with SigLIP-SO400M, supporting a link to visual feature quality.
  • ACFT uses AHAF-generated aligned positive–negative pairs whose controlled perturbation isolates hallucination-triggering visual features for contrastive fine-tuning.AHAF also reveals that minimal perturbations can flip answers, indicating fragile visual representations near hallucination decision boundaries.
  • 0.9% of COCO data and zero inference overhead: ACFT achieves state-of-the-art hallucination mitigation across models and benchmarks while preserving general visual comprehension.Reported evaluations include POPE and MME, with results spanning LLaVA, MiniGPT-4, and Qwen2.5-VL.

2 Related Work

Related work primarily addresses MLLM hallucination through language-prior interventions, especially for long textual outputs. The paper complements these methods by targeting visual-origin hallucination with adversarial contrastive fine-tuning.

  • Most MLLMs combine a visual encoder with a large language model to support cross-modal alignment and reasoning.
  • Input-level, post-processing, latent-space, and post-training methods intervene at different stages to mitigate hallucination.
  • These approaches mainly target textual biases, and their effectiveness degrades in short-output settings where models rely more on visual features.
  • The paper identifies visual-origin hallucination as a distinct mechanism and proposes a targeted fix from the image-prior perspective.

3 Methods

The method uses AHAF to generate targeted, aligned positive–negative pairs and diagnose fragile visual representations, then applies ACFT to improve visual–text alignment during fine-tuning. ACFT combines adversarial contrastive and generation objectives while adding no inference overhead.

  • Adversarial Hallucination Attribute Flipping: Ordinary contrastive fine-tuning uses unrelated negatives with uncontrolled feature differences, making target-object hallucination triggers difficult to learn.AHAF addresses this by selectively altering key visual features associated with the target object.
  • Adversarial Hallucination Attribute Flipping: Small ℓ∞-bounded perturbations can flip the model from correct to hallucinated answers, indicating that many visual features lie near hallucination decision boundaries.AHAF therefore serves both as a training-data generator and as a diagnostic probe.
  • Adversarial Hallucination Attribute Flipping: AHAF applies subtle PGD perturbations to flip an image’s hallucination attribute, creating positive–negative pairs that differ only by a controlled perturbation.The original image is non-hallucinating, while the perturbed image induces object hallucination.
  • Adversarial Contrastive Fine-Tuning: ACFT maximizes text-anchor similarity with positive images while minimizing similarity with adversarial negative images to correct visual–textual misalignment.The contrastive objective is based on image and text representations extracted by visual and text encoders, with temperature controlling softmax sharpness.
  • Adversarial Contrastive Fine-Tuning: ACFT combines adversarial contrastive loss with cross-entropy generation loss, preserving the model’s visual-language generation capability during fine-tuning.The total objective is L_total = L_gen + λL_contra, where λ is empirically determined.

4 Experiments

Experiments evaluate ACFT across multiple MLLMs and hallucination benchmarks, comparing it with inference-time, post-processing, latent-space, and post-training baselines. Results show consistent gains, including strong improvements over OCFT, preserved general visual comprehension, and corrected visual-grounding patterns.

  • Target MLLMs: ACFT is evaluated on LLaVA v1.5-7B, MiniGPT-4 13B, Qwen2.5-VL-7B, and InternVL-3.5-4B.The evaluation spans three primary target models and an additional architecture to test generalization.
  • Baselines and benchmarks: The study compares ACFT with inference-time, post-processing, latent-space, standard SFT, RLHF, and DPO-based baselines on POPE and MME.POPE measures object hallucination through image-object existence questions, while MME includes an Existence subset with the same binary-response format.
  • OCFT comparison: 35.8%, 7.4%, and 17.6% accuracy improvements over OCFT are reported on the three POPE subsets, respectively.OCFT reaches only 0.483 accuracy on the Adversarial subset, below the original LLaVA model.
  • POPE results: ACFT surpasses the second-best baseline on all three POPE subsets for LLaVA and MiniGPT-4, while improving Qwen2.5-VL from 0.864 to 0.877, 0.875 to 0.900, and 0.884 to 0.916.These results are reported across the Random, Popular, and Adversarial subsets.
  • MME results: On MME-Existence, ACFT improves LLaVA accuracy by 3.3% and MiniGPT-4 accuracy by 1.7%, while Qwen2.5-VL and ACFT both achieve 1.000.Some baselines, including OPERA and Woodpecker, degrade LLaVA performance on this subset.
  • Analysis and generalization: ACFT achieves the best overall performance under a comparable ∼6k-sample budget, and it does not degrade full-MME performance while correcting visual-grounding signatures.After ACFT, image-text similarity improves, attention becomes more appropriate, and entropy changes by −8.7% and −2.6% for present objects and +7.4% and +6.4% for absent objects.

5 Conclusion

The paper identifies visual-origin hallucination as a distinct mechanism and proposes AHAF and ACFT to diagnose and mitigate it. Experiments report state-of-the-art performance with limited data and no inference overhead.

  • Visual-origin hallucination is driven by incorrect visual feature extraction and image-text embedding misalignment.
  • AHAF diagnoses fragile visual representations and generates aligned contrastive training pairs for ACFT.
  • ACFT requires only 0.9% of COCO data and adds zero inference overhead.
  • Experiments on POPE, MME, and four description-level benchmarks demonstrate state-of-the-art performance.

Supplementary Material

The supplementary analysis examines object hallucination from the perspective of visual features, using LLaVA v1.5 7B as the representative MLLM.

  • The analysis investigates object hallucination through the perspective of visual features.
  • LLaVA v1.5 7B serves as the representative MLLM for the analysis.
  • The selected backbone provides the basis for examining visual-feature contributions to hallucination.

A.1 Quantitative Analysis

The quantitative and qualitative analyses link hallucination to cross-modal misalignment, incorrect attention, and visual signal quality. Causal interventions further support visual feature quality as a driver of object-existence hallucination.

  • A.1 Quantitative Analysis: Cosine similarity measures alignment between image embeddings and corresponding ground-truth text embeddings.
  • A.1 Quantitative Analysis: 500 correct and 500 hallucinated COCO cases are sampled after evaluating ten POPE-style object-existence questions per image.
  • A.2 Qualitative Analysis: Smooth Grad-CAM visualizes attention maps, with entropy quantifying whether attention concentrates on present objects or disperses when they are absent.
  • A.2 Qualitative Analysis: Hallucinated cases attend to visually similar distractors or irrelevant regions instead of correctly identifying the target object.
  • A.3 Visual Feature Analysis: Lower text-image similarity and misaligned attention distributions characterize hallucinated instances relative to correct cases.

B Implementation Details for Baseline Methods

The implementation details specify baseline settings and the principal adversarial fine-tuning configuration used for AHAF and ACFT.

  • Baseline Methods: VCD uses a noise step of 500, cd-alpha of 1, and cd-beta of 0.1.
  • Baseline Methods: OPERA uses a scale factor of 50.0, threshold 15, five attention candidates, penalty weights of 1.0, and five-beam search.
  • AHAF and ACFT: AHAF uses PGD with 100 iterations, attack budget 16 / 255, and step size 1, while ACFT fine-tunes with LoRA.

D Experimental Setup and Results for Comparison between OCFT and ACFT

The comparison evaluates ACFT against OCFT using aligned adversarial negatives, showing that ACFT better separates target from non-target objects and primarily changes visual representations.

  • Experimental setup: ACFT constructs negative samples by applying PGD to positive COCO images, whereas OCFT selects negatives randomly.Both methods use 50 sampled positive images; ACFT’s negatives are corresponding adversarial samples.
  • Experimental setup: ACFT measures image–text similarity gaps for target and irrelevant non-target objects using extracted image embeddings and anchor text embeddings.The target-object gap is compared with the non-target gap to assess whether perturbations selectively alter target perception.
  • Comparison results: ACFT produces a target-object similarity gap of 0.738 versus 0.276 for non-target objects, while OCFT’s target gap is 0.121 and nearly indistinguishable.These results indicate stronger target-specific manipulation by ACFT and minimal influence on non-target objects.
  • Counterfactual analysis: A counterfactual experiment replaces AHAF visual-origin negatives with GPT-5.5-generated language-prior negatives while keeping the ACFT pipeline fixed.The experiment uses the same 6,000 COCO images and identical fine-tuning setup.
  • Adversarial method: PGD is selected for AHAF because it offers a favorable effectiveness–efficiency trade-off compared with FGSM and CW.FGSM is less robust in challenging scenarios, whereas CW is computationally expensive and complex.
  • Perturbation locus: The visual representation changes by 0.057 versus 0.026 for text-token hidden states, indicating that AHAF perturbations are concentrated in the visual pathway.The visual-level change is over 2× larger than the text-token change across 500 successfully flipped pairs.

F Computational Cost Analysis

AHAF requires substantial offline computation to generate adversarial samples, whereas ACFT adds only negligible training overhead after those samples are available.

  • AHAF cost: AHAF constructs 3,000 contrastive and 3,000 normal samples, requiring roughly 12 GPU-hours on one NVIDIA A100.Each contrastive image uses a 100-iteration PGD attack averaging about 15 seconds.
  • ACFT cost: ACFT reuses final image and text representations without extra backbone forward passes, adding about 90 seconds to standard SFT.On LLaVA-v1.5-7B with batch size 16, training takes 30 min 57s for SFT and 32 min 21s for ACFT.

G Comparison with Training-based Baselines

Under identical backbone, data, and budget, ACFT outperforms standard SFT, adversarial training, DPO, and CHiP-DPO on average POPE accuracy.

  • Controlled comparison: ACFT reaches 0.881 average POPE accuracy, exceeding Base at 0.842, Standard SFT at 0.851, adversarial training at 0.848, Standard DPO at 0.849, and CHiP-DPO at 0.856.The comparison uses LLaVA-v1.5-7B, 6,000 COCO samples with generated negatives, and the same training budget.
  • Interpretation: The authors attribute ACFT’s advantage to aligned pairs that isolate hallucination-triggering visual features and contrastive loss that targets the cross-modal gap.The other baselines supervise outputs, preferences, or perturbation robustness rather than directly targeting image–text alignment.

H Generalization to Additional Architecture

ACFT generalizes across architectures and beyond yes/no object-existence questions, improving POPE performance on InternVL and counting accuracy on TallyQA.

  • InternVL generalization: On InternVL-3.5-4B, ACFT improves POPE accuracy by 1.3%, 1.3%, and 0.4% across the three subsets.On MME Existence, both the original model and ACFT reach 1.000 accuracy.
  • Counting generalization: ACFT improves LLaVA counting accuracy from 0.673 to 0.705 on a 1,000-example TallyQA test set.The model was not specifically trained on counting tasks, supporting transfer beyond yes/no formats.
Loading 2609.00231v1…