Region-Aware Attention Recalibration Reduces Object Hallucination in LVLMs

A new training-free inference strategy called region-aware attention recalibration mitigates object hallucination in Large Vision-Language Models (LVLMs). The method computes an outlier-resistant statistical midpoint across attention heads to establish a stable anchor for visual representations, then uses inter-head disagreement mapped across regions to dynamically determine intervention budgets. This approach avoids expensive data-driven fine-tuning, high-latency contrastive decoding, and rigid attention head truncation, which often compromise computational efficiency or feature space continuity. The paper is published on arXiv with ID 2605.24957.