Technical Summary: CAST – Mitigating Object Hallucination in Large Vision-Language Models via Caption-Guided Visual Attention Steering

1. Problem Statement

Large Vision-Language Models (LVLMs) have achieved remarkable performance on downstream tasks but frequently suffer from object hallucination, where the model generates content that deviates from or contradicts the visual input. Existing mitigation strategies generally fall into two categories, both of which have significant limitations:

• Training-based methods: These rely on carefully curated datasets and expensive manual annotations or reinforcement learning (e.g., RLHF), incurring high computational and data costs.
• Training-free decoding strategies: Methods like contrastive decoding (e.g., VCD, OPERA) often significantly increase inference latency, making them less practical for real-time applications.

Furthermore, interpretability studies suggest that insufficient attention to visual information is a root cause of hallucination. However, few methods explicitly address this by enhancing visual perception without modifying model parameters or introducing heavy inference overhead.

2. Methodology: Caption-guided Visual Attention Steering (CAST)

The authors propose CAST, a training-free, plug-and-play method that mitigates object hallucination by steering the model's attention mechanisms during inference. The core insight driving CAST is the observation that caption queries (e.g., "Please describe this image in detail") trigger significantly enhanced visual attention in specific attention heads compared to non-caption queries (e.g., "Is there a helmet in the image?").

CAST operates in three distinct stages:

A. Caption-Guided Attention Heads Probe

The method first identifies "caption-guided attention heads"—specific heads within the Transformer layers that exhibit a strong sensitivity to caption queries.

• Mechanism: Using a dataset of image-query pairs, the authors mask attention to textual tokens to isolate visual attention patterns.
• Classification: A binary classifier (e.g., SVM) is trained on the attention outputs of the last input token to distinguish between caption and non-caption queries.
• Selection: The top-$K$ attention heads with the highest classification accuracy are selected as the targets for steering. These heads are found to be concentrated primarily in the middle layers of the model.

B. Estimation of Perception Refined Vectors

Once the target heads are identified, the method computes shift vectors that represent the difference in attention output between caption and non-caption queries.

• Computation: For each selected head, the shift vector $S^{(l,h)}$ is calculated as the average difference between the attention output when fed a caption query ($O^{(l,h)}$) and a non-caption query ($O'^{(l,h)}$):
  $$S^{(l,h)} = \frac{1}{B} \sum_{b=1}^{B} (O^{(l,h)}_b - O'^{(l,h)}_b)$$
• Purpose: These vectors quantify the "fine-grained perception direction," effectively capturing the state of sufficient visual attention induced by caption queries.

C. Steering at Inference Time

During the inference of a non-caption query (the typical scenario where hallucinations occur), the precomputed shift vectors are applied to the selected attention heads.

• Update Rule: The hidden state is updated by adding a scaled version of the shift vector to the attention output of the caption-guided heads:
  $$H_{l+1} = H_l + \sum_{h=1}^{H} (O^{(l,h)} + I^{(l,h)} \alpha S^{(l,h)}) \cdot W^l_o$$
  Where $I^{(l,h)}$ is a gating function (1 for selected heads, 0 otherwise) and $\alpha$ controls the steering intensity.
• Effect: This steers the model's attention from an "insufficient" state (prone to hallucination) to a "sufficient" state (enhanced visual perception), effectively mitigating object hallucination without retraining the model.

3. Key Contributions

1. Novel Insight: The work is the first to explicitly reveal and quantify the impact of caption queries versus non-caption queries on the attention activation patterns of LVLMs, identifying specific attention heads responsible for fine-grained visual perception.
2. CAST Method: A training-free, plug-and-play approach that refines caption-guided attention head outputs during inference. It requires no model parameter updates and adds minimal inference cost.
3. Comprehensive Evaluation: Extensive experiments demonstrate that CAST effectively mitigates hallucination while preserving the model's foundational capabilities (e.g., reasoning, OCR) and generalizing across different model architectures (LLaVA, Qwen-VL, LLaVA-NeXT).

4. Experimental Results

The authors evaluated CAST across five widely used benchmarks (POPE, CHAIR, MMHal-Bench, MME, MHumanEval) and five different LVLMs.

• Hallucination Mitigation:
  • POPE: CAST improved accuracy by an average of 5.64% and F1 score by 5.50% across models.
  • CHAIR: Reduced average hallucination metrics (CS and CI) by 6.43 points.
  • MMHal-Bench: Reduced the hallucination rate (VH.%) by 2.95% and improved the informativeness score.
  • General Performance: On average, CAST reduced object hallucination by 6.03% across five LVLMs and five benchmarks.
• Efficiency: Unlike contrastive decoding methods that increase inference time significantly (e.g., VCD increases TPOT by 2.7x), CAST adds negligible latency (comparable to the baseline, ~1.0x).
• Preservation of Capabilities: CAST improved performance on the MME benchmark across all perception and cognition tasks, demonstrating that it does not degrade the model's other foundational abilities.
• Generalization: The method proved effective on advanced models (Qwen2-VL, InternVL2) and domain-specific tasks (Medical VQA-RAD, OCR), confirming its robustness.

5. Significance and Claims

The paper claims that CAST offers a state-of-the-art (SOTA) solution for mitigating object hallucination in LVLMs by leveraging the inherent attention patterns induced by caption queries. Its significance lies in:

• Cost-Effectiveness: It eliminates the need for expensive manual annotation or retraining.
• Efficiency: It avoids the high inference latency associated with contrastive decoding strategies.
• Mechanistic Insight: It provides a clear, interpretable mechanism for hallucination mitigation by directly targeting the attention heads responsible for visual perception.

The authors conclude that CAST effectively enhances the fine-grained visual perception capability of LVLMs, offering a practical and efficient path to reducing hallucinations without compromising the model's utility.