Mitigating Hallucination in Large Vision-Language Models via Adaptive Attention Calibration

1. Problem Statement

Large Vision-Language Models (LVLMs) excel at multimodal tasks but frequently suffer from hallucination, where they confidently describe objects, attributes, or scenes not present in the input image. While existing training-free interventions (e.g., contrastive decoding) show promise in short-form or discriminative tasks, they struggle to maintain accuracy in open-ended and long-form generation scenarios.

The authors identify two fundamental biases causing these hallucinations:

1. Spatial Perception Bias: The model distributes attention disproportionately across image tokens, often over-concentrating on a small subset of tokens (attention sinks) regardless of the actual image content. This leads to the model overlooking relevant visual cues.
2. Modality Bias: As the generation sequence lengthens, the model increasingly shifts its attention from visual inputs to textual history (language priors). This causes the model to "drift" away from the image, resulting in content that is poorly grounded in the visual input, particularly in later tokens of long sequences.

2. Methodology: The CAAC Framework

The authors propose Confidence-Aware Attention Calibration (CAAC), a unified, training-free framework that dynamically recalibrates the LVLM's attention mechanism during inference. It operates via two distinct components targeting the identified biases:

A. Visual-Token Calibration (VTC)

• Goal: Mitigate Spatial Perception Bias.
• Mechanism: VTC smooths the attention distribution across image tokens to prevent over-concentration on specific regions.
• Process:
  1. Reference Construction: A "meaningless" image (e.g., a black canvas) and a generic query are used to extract a baseline attention pattern ($V^{h,l}$) for each attention head.
  2. Inverse Calibration: An inverse calibration vector is computed from this baseline to counteract the inherent skew in attention weights.
  3. Smoothing: For a specific input, the original attention vector ($V$) is combined with the calibrated vector ($V_u$) using a smoothing parameter $\beta$:
     $$V_s = (1 - \beta)V + \beta V_u$$
  4. This ensures a more balanced attention distribution across image tokens while preserving essential visual information.

B. Adaptive Attention Re-Scaling (AAR)

• Goal: Mitigate Modality Bias (the drift toward text over time).
• Mechanism: AAR dynamically amplifies attention to image tokens when the model's confidence in its prediction is low.
• Process:
  1. Confidence Monitoring: During generation, the model calculates the maximum logit probability ($p_t$) for the next token.
  2. Thresholding: If $p_t$ falls below a preset threshold ($p_{thr}$), indicating high uncertainty (and thus high hallucination risk), AAR triggers.
  3. Adaptive Scaling: A scaling factor $\lambda_t$ is calculated based on the confidence level:
     $$\lambda_t = \lambda_{min} \cdot p + \lambda_{max} \cdot (1 - p)$$
     Where $\lambda_{min}=1$ (no scaling if confident) and $\lambda_{max}$ is the maximum boost.
  4. Re-weighting: The attention weights for image tokens are scaled by $\lambda_t$ before the softmax operation, forcing the model to re-ground its generation in the visual input when it is uncertain.

3. Key Contributions

1. Novel Analysis: The paper provides a rigorous analysis using relevancy maps (rather than raw attention weights) to demonstrate that hallucinations are strongly correlated with low relative image relevancy and low generation confidence, particularly in long sequences.
2. Training-Free Framework: CAAC introduces a two-step, training-free approach (VTC + AAR) that can be seamlessly integrated into any LVLM without fine-tuning.
3. Performance Gains: The method significantly outperforms state-of-the-art baselines (OPERA, VCD, M3ID, AvisC) in reducing hallucination rates, specifically in long-form generation tasks where other methods fail.

4. Experimental Results

The framework was evaluated on three major benchmarks: CHAIR, AMBER, and POPE, using models like InstructBLIP, LLaVA-1.5, and LLaVA-NeXT.

• CHAIR Benchmark (Object Hallucination):
  • CAAC achieved the lowest hallucination rates ($C_s$ and $C_i$) across all models.
  • On InstructBLIP, it reduced $C_s$ from 55.6 (base) to 37.4 and $C_i$ from 16.6 to 10.8, outperforming the best baseline (OPERA).
• AMBER Benchmark (Generative & Discriminative):
  • CAAC demonstrated superior performance in long-form generation (Max Tokens = 512). While contrastive decoding methods (VCD, M3ID) saw significant degradation in hallucination metrics as token length increased, CAAC maintained low hallucination rates (e.g., CHAIR score of 7.0 vs. 10.8 for VCD on InstructBLIP at 512 tokens).
  • It achieved a strong balance between Faithfulness (low hallucination) and Completeness (COVER score), whereas methods with high coverage often suffered from high hallucination rates.
• POPE Benchmark (Discriminative):
  • Despite being designed for generative tasks, CAAC performed robustly in discriminative settings, achieving Accuracy and F1 scores within 1% of the top-performing baseline (OPERA) and outperforming other training-free methods.
• Ablation Study:
  • Both VTC and AAR modules contributed individually to reducing hallucinations.
  • The full CAAC framework (VTC + AAR) yielded the best results, confirming that addressing both spatial and modality biases is necessary for optimal performance.

5. Significance and Impact

• Solving the Long-Form Problem: CAAC addresses a critical gap in current research: the failure of existing methods to prevent hallucinations in long, open-ended text generation. By adaptively re-weighting attention based on confidence, it ensures visual grounding persists throughout the generation sequence.
• Safety-Critical Applications: By prioritizing factual accuracy over exhaustive detail (even if it slightly reduces coverage metrics compared to some baselines), CAAC is particularly suitable for high-stakes domains like medical diagnosis, legal document analysis, and autonomous navigation, where hallucination can have severe consequences.
• Efficiency: The method introduces only modest latency. Unlike contrastive decoding which requires two passes for every token, CAAC triggers its adaptive re-scaling only when confidence is low (approx. 14% of tokens in experiments), making it computationally efficient.

In conclusion, CAAC offers a robust, model-agnostic solution to LVLM hallucination by dynamically correcting attention biases, ensuring that generated text remains firmly grounded in visual reality, even during complex, long-form generation tasks.