Source-linked AI summary
ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large language Models
Hao Yin, Guangzong Si, Zilei Wang
TL;DR
Object hallucinations in MLLMs arise when language priors outweigh visual evidence, while contrastive decoding can harm content quality and slow inference. The paper proposes Visual Amplification Fusion, which strengthens visual attention during middle-layer fusion without additional training. Across benchmarks, VAF reduces hallucinations while preserving coherence, accuracy, and inference speed.
Problem
Contrastive decoding can reduce hallucinations but may compromise generated-content coherence and accuracy while increasing inference time.
Method
Visual Amplification Fusion amplifies visual signals in middle layers, where modality fusion primarily occurs, without additional training or contrastive-sample processing.
Results
VAF significantly improves hallucination-mitigation performance across multiple benchmarks while maintaining content quality and inference speed.
Takeaways & Limitations
Enhancing visual attention during modality fusion reduces over-reliance on language priors while preserving the beneficial influence of language for generation.
Takeaways & Limitations
Contrastive decoding’s content-quality degradation is more pronounced in complex tasks such as multiple-choice question answering and image captioning than in binary object-hallucination tasks.
Abstract
from arXiv · showhide
Contrastive decoding strategies are widely used to mitigate object hallucinations in multimodal large language models (MLLMs). By reducing over-reliance on language priors, these strategies ensure that generated content remains closely grounded in visual inputs, producing contextually accurate outputs. Since contrastive decoding requires no additional training or external tools, it offers both computational efficiency and versatility, making it highly attractive. However, these methods present two main limitations: (1) bluntly suppressing language priors can compromise coherence and accuracy of generated content, and (2) processing contrastive inputs adds computational load, significantly slowing inference speed. To address these challenges, we propose Visual Amplification Fusion (VAF), a plug-and-play technique that enhances attention to visual signals within the model's middle layers, where modality fusion predominantly occurs. This approach enables more effective capture of visual features, reducing the model's bias toward language modality. Experimental results demonstrate that VAF significantly reduces hallucinations across various MLLMs without affecting inference speed, while maintaining coherence and accuracy in generated outputs.
1. Introduction
MLLMs often hallucinate objects because generated descriptions can over-rely on language priors rather than visual evidence. VAF addresses contrastive decoding’s quality and speed drawbacks by amplifying visual signals during middle-layer modality fusion.
- Motivation: Object hallucination occurs when textual descriptions do not match image objects, reflecting over-reliance on unimodal priors and posing risks in high-precision applications.The paper highlights medical diagnosis and autonomous driving as examples of such applications.
- Limitations of Contrastive Decoding: Contrastive decoding can compromise generated-content coherence and accuracy by suppressing language priors, despite reducing hallucinations.Its additional processing also increases computational load, creating a trade-off between visual grounding, content quality, and speed.
- Method: VAF amplifies visual signals in middle layers, where modality fusion primarily occurs, to capture distinctive visual features and reduce false descriptions.The method strengthens visual representations while retaining beneficial language-prior influence.
- Results: Across multiple object hallucination benchmarks, VAF improved performance by approximately 3% on POPE and 7% on MME while preserving coherence and accuracy.These results support the method’s effectiveness without compromising generated-content quality.
- Motivation and Results: VAF avoids contrastive-sample processing, maintaining inference speed while contrastive decoding can substantially slow generation.Contrastive decoding requires separate processing of original and contrastive inputs, increasing inference time.
2. Related work
MLLM development has improved multimodal alignment through increasingly capable language-based architectures and mechanisms such as Q-Former and task-specific instructions. Contrastive decoding separately grounds generation by contrasting outputs from original versus distorted visual inputs or disrupted instructions.
- MLLM Architectures: Q-Former mechanisms in BLIP-2 and miniGPT-4 improve alignment between visual and textual inputs, while InstructBLIP adds task-specific instructions.These developments support more precise interactions across modalities.
- Contrastive Decoding: Visual Contrastive Decoding contrasts output distributions from original and distorted visual inputs to reduce reliance on statistical bias and unimodal priors.Instruction Contrastive Decoding applies a similar comparison to standard and disrupted instructions.
- Contrastive Decoding: Contrastive methods aim to ground generated content closely to visual inputs, producing contextually accurate outputs while addressing hallucinated concepts.The related approaches include visual-input and instruction-based perturbations.
3. Preliminary and Motivation
Visual Contrastive Decoding reduces hallucinations by contrasting outputs from original and distorted visual inputs, but its extra processing and suppression of language priors can reduce speed and content quality.
- 3.1. Contrastive Decoding: VCD contrasts output distributions conditioned on original and distorted visual inputs to mitigate hallucinations by countering language-prior bias.The distorted input is produced using predefined distortions such as a Gaussian noise mask, and α = 0 reduces to regular decoding.
- 3.2. Limitations of Contrastive Decoding: Although contrastive decoding requires no additional training or external tools, it trades this efficiency and versatility for reduced content quality and slower inference.These two drawbacks motivate alternatives that preserve both generation quality and speed.
- 3.2. Limitations of Contrastive Decoding: Subtracting language-prior influence can undermine coherence and accuracy, especially in multiple-choice question answering and image captioning.The issue is less pronounced for binary yes/no object-hallucination tasks.
- 3.2. Limitations of Contrastive Decoding: VCD reduced performance by 5% on ScienceQA and 45% on NoCaps, indicating substantial content-quality impairment in nuanced generation tasks.The evaluation used LLaVA-v1.5-7B and LLaVA-v1.5-13B models; the supplied figure reports CIDEr for NoCaps and Accuracy for ScienceQA.
- 3.2. Limitations of Contrastive Decoding: VCD's additional contrastive processing nearly doubles inference time compared with vanilla decoding.Each visual input requires separate computation of logits for the original and distorted inputs.
4. Visual Neglect in Modal Fusion
Saliency and attention analyses identify the middle layers as the main site of visual-textual fusion, yet visual information receives insufficient attention there, encouraging reliance on language priors.
- 4.1. Mid-layer: Visual-Language Fusion: Saliency analysis examines attention-matrix token interactions using Taylor expansion, averaging saliency across attention heads to quantify information flow.The analysis defines intra-visual flow among image tokens and visual-textual flow from image tokens to instruction tokens.
- 4.1. Mid-layer: Visual-Language Fusion: 8th–15th layers show especially important visual-textual information flow that substantially influences prediction outcomes.Experiments use LLaVA-v1.5-7B on 500 MS COCO examples under POPE.
- 4.2. Attention Imbalance Across Modalities: In middle layers, visual attention is lower than attention to system prompts and user instructions, indicating insufficient visual focus during modality fusion.The resulting underutilization skews output distributions toward language priors.
- 4.1. Mid-layer: Visual-Language Fusion: Middle layers integrate visual and textual inputs into cross-modal semantic representations that drive final predictions.
- 4.3. Insights: Insufficient visual attention during critical fusion leads to incomplete use of visual information and subsequent dependence on language priors associated with hallucinations.
5. Visual Amplification Fusion
Visual Amplification Fusion (VAF) enhances visual attention during middle-layer modality fusion, selectively reallocating attention to visual features while reducing unnecessary system-prompt attention.
- 5.1. Attention Redistribution: The method amplifies visual features and suppresses superfluous attention to system prompts, reducing over-reliance on language priors.These changes are intended to improve visual-information integration and reduce false descriptions.
- 5.1. Attention Redistribution: VAF adjusts attention in the middle layers, where visual–textual fusion occurs, because visual attention there remains insufficient.The method targets layers 8 < l < 15 to improve balance during modality integration.
- 5.1. Attention Redistribution: VAF modifies each selected attention score using enhancement and suppression coefficients, with α controlling visual amplification and β controlling system-prompt suppression.The enhancement coefficient satisfies α > 0, while the suppression coefficient satisfies 0 < β < 1.
- 5.1. Attention Redistribution: Enhanced visual attention in the fusion layers reduces hallucinations by improving the model’s focus on visual information.Figure 6 illustrates this effect through correction of a grape-color prediction from “green” to “red”.
- 5.1. Attention Redistribution: Rather than enhancing every head, VAF restricts visual-attention reallocation to visual perception heads identified among those with the top 50% visual attention allocation.Other attention heads remain unchanged, limiting the potentially harmful effects of overly aggressive enhancement.
6. Experiment
Experiments across multiple MLLMs and benchmarks evaluate VAF for hallucination suppression, content quality, inference speed, and component effectiveness. VAF consistently outperforms contrastive baselines while preserving response quality and imposing minimal speed costs.
- 6.2.1. Hallucination Mitigation: VAF consistently surpasses VCD and ICD on POPE across LLaVA-v1.5 and Qwen-VL models, while MME results show suppression of object- and attribute-level hallucinations.POPE results average performance across MS-COCO, A-OKVQA, and GQA.
- 6.2.2. Content Quality: VCD and ICD reduce NoCaps CIDEr scores by 18% and 27%, respectively, whereas VAF has minimal negative impact on generated-content coherence and accuracy.The comparison attributes contrastive-decoding degradation to crude disruption of language priors.
- 6.2.3. Inference Speed: VCD and ICD nearly double inference time, while VAF has minimal impact on multimodal-model inference speed.Figure 7 summarizes the speed comparison on NoCaps.
- 6.3. Ablation Study: When 0 < α < 0.25, hallucinations are effectively suppressed on POPE, but performance degrades once α exceeds 0.25.The authors associate excessive α with over-focusing on visual features and disrupted language-information integration.
- 6.3. Ablation Study: Restricting redistribution to visual perception heads reduces hallucinations while better preserving content quality than unrestricted visual-attention enhancement.The ablation evaluates this mechanism on POPE and NoCaps.
7. conclusion
The paper identifies quality and speed drawbacks in contrastive decoding and proposes Visual Amplification Fusion to mitigate hallucinations while preserving both. Across multiple benchmarks and MLLMs, VAF improves grounding between generated content and visual inputs.
- 7. conclusion: VAF addresses contrastive decoding’s reduced content quality and slower inference by enhancing visual attention during modality fusion.The method is designed to reduce language-prior over-reliance while preserving inference speed and generation quality.
- 7. conclusion: Experiments across multiple benchmarks and MLLMs demonstrate a clear advantage for VAF in hallucination mitigation.The conclusion links this advantage to greater consistency between generated content and visual inputs.
8. Additional Experimental Results
Across MME, POPE, and LLaVA-Bench evaluations, VAF consistently improves hallucination suppression and task performance over VCD while preserving inference efficiency. The results attribute these gains to stronger visual attention during modality fusion without separately processing contrastive inputs.
- 8.1. Detailed Experimental Results on MME: VAF consistently outperforms VCD on perception-related MME tasks for both LLaVA-v1.5-7B and 13B models.The 13B model achieves leadership across all reported perception tasks.
- 8.1. Detailed Experimental Results on MME: VAF slightly improves cognition-task performance, whereas VCD degrades it by disrupting effective linguistic priors.VAF instead integrates visual information while preserving useful linguistic information.
- 8.2. Detailed Experimental Results on POPE: VAF delivers more stable and significantly improved hallucination suppression than VCD across MSCOCO, A-OKVQA, and GQA in POPE.The comparison covers both LLaVA-v1.5-7B and 13B model results.
- 8.3. Inference Speed and Memory Usage: VAF has minimal inference-speed impact and requires no additional GPU memory, unlike VCD, because it avoids processing contrastive inputs.The efficiency comparison covers POPE and Nocaps experiments.
- 8.4. Case Studies on LLaVA-Bench: VAF consistently suppresses hallucinations across numerical perception, image description, and complex reasoning tasks on LLaVA-Bench.The reported examples use the LLaVA-v1.5-7B model.
9. Additional Ablation Studies
Ablations show that VAF works best when visual attention is enhanced in middle layers and when its suppression coefficient remains within a moderate range. It also remains effective across sampling strategies.
- 9.1. Visual Attention Layer Placement: Enhancing visual attention in middle layers significantly reduces hallucinations, while shallow and deep-layer modifications have minimal effect.This supports the model’s middle layers as the primary site of modality integration.
- 9.2. Suppression Coefficient β: VAF significantly improves hallucination suppression when 0 < β < 0.15, but performance deteriorates when β > 0.15.The experiments fix α at 0.15 on COCO-Random within POPE.
- 9.3. Sampling Strategies: VAF significantly mitigates hallucinations across all evaluated sampling strategies.The test uses LLaVA-v1.5-7B on COCO-Random within POPE.
10. Prompts for Different Tasks
The paper uses task-specific prompts for POPE, Nocaps/Flickr30k, and Sci-VQA. These prompts respectively elicit binary object-presence answers, one-sentence captions, and selected-answer responses.
- POPE Dataset: POPE uses a yes-or-no prompt asking whether a cow appears in the image.The template frames the interaction as a helpful assistant conversation.
- Nocaps Datasets: Nocaps and Flickr30k use a prompt requesting a one-sentence caption for the provided image.The same assistant-conversation template is used.
- Sci-VQA Dataset: Sci-VQA asks the model to select the best answer and return the option’s letter directly.The example asks which property three objects share among shiny, slippery, and opaque.