Technical Summary: MHSA – A Lightweight Framework for Mitigating Hallucinations via Steered Attention in LVLMs

1. Problem Statement

Large Vision-Language Models (LVLMs) have demonstrated exceptional capabilities in multimodal tasks such as visual question answering (VQA) and image captioning. However, they suffer from hallucinations, generating content inconsistent with or absent from the visual input. These manifest as object, attribute, or relational hallucinations, severely undermining model reliability.

Existing mitigation strategies face significant limitations:

• Training-based methods (e.g., RLHF, instruction tuning) require substantial computational resources and high-quality annotated data, and improvements are often bound to specific training distributions.
• Inference-time methods often rely on hand-crafted, fixed heuristic rules (e.g., penalty functions, static re-weighting) that fail to adapt to the specific hallucination signatures of individual input samples.
• Contrastive decoding approaches (e.g., VCD, ICD) typically double inference costs by requiring multiple forward passes.

Recent work, DHCP (Detecting Hallucinations by Cross-modal Attention Pattern), established that hallucinated and non-hallucinated samples exhibit distinguishable cross-modal attention patterns. However, DHCP is limited to detection only and operates primarily at the sentence level, offering no mechanism for mitigation.

2. Methodology: MHSA Framework

The authors propose MHSA (Mitigating Hallucinations via Steered Attention), a lightweight, training-free framework that bridges the gap between detection and mitigation by learning to correct cross-modal attention patterns.

Core Architecture

MHSA operates via a two-stage "detect-then-correct" pipeline without modifying the LVLM backbone parameters:

1. Discriminator ($D$): A pre-trained, lightweight two-layer MLP (originally from DHCP) that acts as a token-level detector. It analyzes cross-modal attention weights to identify hallucinatory patterns.
2. Generator ($G$): A lightweight three-layer MLP that takes the original cross-modal attention tensor ($A$) as input and produces a correction term ($\Delta A$).
3. Steering Mechanism: The corrected attention is computed as $A' = A + \Delta A$. During inference, if $D$ detects a hallucination, $G$ generates $\Delta A$, and the LVLM uses $A'$ instead of $A$ to produce the final output.

Training Objectives

The generator $G$ is trained using a three-component objective to ensure effective and stable correction:

1. Detector-Guided Loss ($\mathcal{L}_{dg}$): An adversarial loss that encourages the generator to transform hallucinatory attention patterns into non-hallucinatory ones, as judged by the discriminator $D$.
2. Regularization Loss ($\mathcal{L}_{reg}$): A constraint on the magnitude of $\Delta A$ ($\|\Delta A\|_2^2$) to ensure corrections are minimal and localized, preserving the original attention semantics and preventing artifacts.
3. LVLM Output Quality Loss ($\mathcal{L}_{LVLM}$): A cross-entropy loss ensuring that the corrected attention maintains or improves the quality of the LVLM's generated output against ground truth.

Adaptation to Generative Tasks

While DHCP originally operated at the sentence level, MHSA extends this to token-level correction for generative tasks (e.g., image captioning).

• The detector and generator are trained on per-token cross-modal attention tensors extracted during image captioning.
• During inference, the system evaluates attention at each decoding step. If a specific token's attention pattern is flagged as hallucinatory, the generator applies a correction before sampling that token.

3. Key Contributions

1. First Data-Driven Mitigation Framework: MHSA is the first framework to mitigate LVLM hallucinations by learning sample-adaptive corrections of cross-modal attention patterns, moving beyond fixed heuristic rules.
2. Token-Level Extension: The authors extend attention-based detection from the sentence level to the token level, enabling mitigation in both discriminative (VQA) and generative (captioning) tasks.
3. Lightweight and Efficient: The framework does not modify LVLM backbone parameters. It trains only a small MLP generator and utilizes a pre-trained detector, introducing marginal computational overhead compared to the LVLM's inference cost.
4. Comprehensive Evaluation: The method is validated across multiple state-of-the-art LVLMs (Qwen2.5-VL, InternVL2, LLaVA-v1.5) and diverse datasets (MSCOCO, Objects365, OpenImages, Flickr30k).

4. Experimental Results

Discriminative Tasks (POPE Benchmark)

• Performance: MHSA achieves consistent and significant improvements in F1-scores across all tested models and datasets. For example, on Qwen2.5-VL-7B with MSCOCO, MHSA improved the F1-score by +7.42 points (from 85.55 to 92.97) and Recall by +15.78 points.
• Generalization: The method demonstrates strong generalization across different model architectures and datasets. Crucially, it shows robust out-of-distribution (OOD) performance; models trained on one dataset (e.g., Objects365) effectively mitigate hallucinations when tested on another (e.g., COCO), suggesting the learned corrections capture fundamental hallucination signatures rather than dataset-specific artifacts.

Generative Tasks (Image Captioning)

• Hallucination Reduction: On the Flickr30k dataset, MHSA reduced both sentence-level ($CHAIR_s$) and instance-level ($CHAIR_i$) hallucination rates by 44% with only a minor drop in Recall (3.9%).
• Trade-off: The method successfully balances hallucination suppression with content preservation, avoiding the severe recall drops seen in aggressive heuristic-based methods.

Efficiency

• Latency: MHSA introduces an amortized latency increase of only 0.43× on the POPE-COCO benchmark. This is because the lightweight detector screens all samples, and the generator is only invoked for the minority of samples flagged as hallucinated (approx. 12.3% in the adversarial POPE setting). This is significantly more efficient than contrastive decoding methods that require a fixed ~1.0× overhead for every sample.

5. Significance and Claims

The paper claims that MHSA offers a novel perspective on hallucination research by extending cross-modal attention mechanisms from mere detection to active mitigation.

• Mechanistic Insight: Visualizations confirm that MHSA functions as a targeted correction mechanism. It concentrates attention on task-relevant regions and reduces spatial entropy, effectively "steering" the model away from hallucinatory patterns without broad perturbation of the attention structure.
• Practicality: By avoiding full model fine-tuning and minimizing inference overhead, MHSA presents a practical, deployable solution for enhancing the reliability of LVLMs in real-world applications.
• Limitations: The authors note that while MHSA significantly improves metrics, it is not a panacea; future work is needed to explore more expressive generator architectures and broader model coverage.

In summary, MHSA leverages the correlation between attention patterns and hallucinations to create a lightweight, data-driven framework that effectively reduces hallucinations in both discriminative and generative LVLM tasks, outperforming existing heuristic and training-based approaches in terms of efficiency and generalizability.