Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Scalpel: Fine-Grained Alignment of Attention Activation Manifolds via Mixture Gaussian Bridges to Mitigate Multimodal Hallucination

Ziqiang Shi, Rujie Liu, Shanshan Yu, Satoshi Munakata, Koichi Shirahata

TL;DR

Scalpel addresses hallucinations in vision-language transformers by explicitly modeling trusted and hallucinated attention activation manifolds with Gaussian Mixtures and computing minimal transport via the Schrödinger bridge. The method performs inference-time, per-token corrections by mapping hallucinated components to trusted ones and applying targeted interventions to the top-$k$ heads, without retraining. Empirical results on POPE and MME benchmarks show Scalpel achieving state-of-the-art performance across multiple LVLM backbones and datasets, with strong robustness to adversarial and complex scenarios. This approach improves reliability in multimodal reasoning while preserving the model's knowledge capacity and incurs no additional decoding steps beyond the proposed corrections.

Abstract

Rapid progress in large vision-language models (LVLMs) has achieved unprecedented performance in vision-language tasks. However, due to the strong prior of large language models (LLMs) and misaligned attention across modalities, LVLMs often generate outputs inconsistent with visual content - termed hallucination. To address this, we propose \textbf{Scalpel}, a method that reduces hallucination by refining attention activation distributions toward more credible regions. Scalpel predicts trusted attention directions for each head in Transformer layers during inference and adjusts activations accordingly. It employs a Gaussian mixture model to capture multi-peak distributions of attention in trust and hallucination manifolds, and uses entropic optimal transport (equivalent to Schrödinger bridge problem) to map Gaussian components precisely. During mitigation, Scalpel dynamically adjusts intervention strength and direction based on component membership and mapping relationships between hallucination and trust activations. Extensive experiments across multiple datasets and benchmarks demonstrate that Scalpel effectively mitigates hallucinations, outperforming previous methods and achieving state-of-the-art performance. Moreover, Scalpel is model- and data-agnostic, requiring no additional computation, only a single decoding step.

Scalpel: Fine-Grained Alignment of Attention Activation Manifolds via Mixture Gaussian Bridges to Mitigate Multimodal Hallucination

TL;DR

Scalpel addresses hallucinations in vision-language transformers by explicitly modeling trusted and hallucinated attention activation manifolds with Gaussian Mixtures and computing minimal transport via the Schrödinger bridge. The method performs inference-time, per-token corrections by mapping hallucinated components to trusted ones and applying targeted interventions to the top- heads, without retraining. Empirical results on POPE and MME benchmarks show Scalpel achieving state-of-the-art performance across multiple LVLM backbones and datasets, with strong robustness to adversarial and complex scenarios. This approach improves reliability in multimodal reasoning while preserving the model's knowledge capacity and incurs no additional decoding steps beyond the proposed corrections.

Abstract

Rapid progress in large vision-language models (LVLMs) has achieved unprecedented performance in vision-language tasks. However, due to the strong prior of large language models (LLMs) and misaligned attention across modalities, LVLMs often generate outputs inconsistent with visual content - termed hallucination. To address this, we propose \textbf{Scalpel}, a method that reduces hallucination by refining attention activation distributions toward more credible regions. Scalpel predicts trusted attention directions for each head in Transformer layers during inference and adjusts activations accordingly. It employs a Gaussian mixture model to capture multi-peak distributions of attention in trust and hallucination manifolds, and uses entropic optimal transport (equivalent to Schrödinger bridge problem) to map Gaussian components precisely. During mitigation, Scalpel dynamically adjusts intervention strength and direction based on component membership and mapping relationships between hallucination and trust activations. Extensive experiments across multiple datasets and benchmarks demonstrate that Scalpel effectively mitigates hallucinations, outperforming previous methods and achieving state-of-the-art performance. Moreover, Scalpel is model- and data-agnostic, requiring no additional computation, only a single decoding step.
Paper Structure (14 sections, 1 theorem, 17 equations, 5 figures, 3 tables)

This paper contains 14 sections, 1 theorem, 17 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Hallucination in LVLMs
  4. Hallucination Mitigation in LVLMs
  5. Methodology
  6. Background and Notation
  7. Modeling Trusted and Hallucinated Manifolds of Attention Activations
  8. Optimal Transport Mapping Between Trusted and Hallucinated GMMs
  9. Hallucination Mitigation Based on Optimal Mapping Between Manifolds
  10. Experiments
  11. Benchmarks and Experimental Setup
  12. Results and Discussion
  13. Ablation study
  14. Conclusion

Key Result

Proposition 3.1

(Optimal GMM bridge policy for hallucination-to-trust transition) For the GMM bridge problem in Eq. (eq:sb_gmm_problem), let ${\mathbf{u}}_{t|ij}$ be the optimal control policy for the Gaussian bridge between hallucinated component $i$ ($\mathcal{N}(\mu_0^i,\Sigma_0^i)$) and trusted component $j$ ($ under optimal transport constraints with $\lambda_{ij}\geq 0, \forall i,j$: The solution $\lambda_

Figures (5)

  • Figure 1: Schematic diagram illustrating the principle of the Scalpel method. First, the trusted and hallucinated attention activations of all heads are obtained by inputting both correct and hallucinated data into the LVLM. These activations serve two purposes. The first is to identify the heads with the highest hallucination discrimination capability using a classifier. The second is to analyze the distribution of credible and hallucinated attention activations across tokens, thereby determining the influence of each individual token on each head.
  • Figure 2: Comparison of the distributions of trusted and hallucinated attention activations at the image and object levels, including the component composition of GMMs and the transition flow from the hallucinated GMMs to the trusted GMMs. For simplicity, we use t-SNE van2008visualizing to map the original attention activations into a two-dimensional space before performing all subsequent operations.
  • Figure 3: On the MME benchmark, Scalpel outperforms prior SOTA methods - ICT, Vanilla LLaVA-1.5 and Qwen2.5-VL. The radar chart highlights improvements across key categories: existence, location, counting, color perception, common sense reasoning, and overall performance.
  • Figure 4: Ablation study of Scalpel under different GMM component settings, evaluated on the POPE benchmark (MS COCO dataset) with LLaVA-1.5-7B.
  • Figure 5: A comparative analysis of Scalpel, ICT, and the original Qwen2.5-VL-7B across selected test cases is presented. Q denotes the question, GT represents the ground truth, and each method's answer follows its name in bold. Extended hallucinated responses are highlighted in red for emphasis. Notably, the original questions included the instruction "Please answer yes or no" immediately after the question mark, which has been omitted to optimize space usage.

Theorems & Definitions (1)

  • Proposition 3.1