arXiv:2605.23344v1 Announce Type: new 
Abstract: Large Vision-Language Models have shown strong multimodal reasoning capabilities, yet they remain susceptible to object hallucinations when language priors dominate insufficient or misaligned visual evidence. Training-free contrastive decoding methods mitigate this issue by comparing predictions from original and perturbed visual inputs, but existing approaches either apply global perturbations that may alter useful visual evidence or invoke an additional negative branch at every decoding step. In this paper, we observe that hallucination risks are transient and token-specific: visual attention shifts across generated tokens, while some functional tokens are produced with high confidence and do not require contrastive calibration. Based on this observation, we propose Contrastive Hallucination-Aware Step-wise Decoding (CHASD) for Large Vision-Language Models, an inference-time framework for "calibration on demand". CHASD uses an uncertainty-driven confidence gate to activate the contrastive branch only when the maximum probability of the next-token is less than the threshold, and constructs the negative branch through attention-guided localized perturbations of the currently salient visual tokens. This design reduces unnecessary negative-branch forward passes while preserving the original distribution for high-confidence steps. Experiments on POPE, AMBER, MME, MMHal-Bench, and CHAIR show that CHASD improves hallucination-related metrics over strong training-free baselines with competitive inference efficiency.

قدم الباحثون إطار عمل جديد يسمى (CHASD) الذي يهدف إلى تحسين أداء نماذج الرؤية واللغة الكبيرة (LVLMs) من خلال معالجة مشكلة الهلوسة المتعلقة بالأشياء أثناء الاستدلال. تركز هذه الطريقة على ضبط الانتباه البصري بشكل ديناميكي، مما يسمح بإنتاج نتائج أكثر دقة وملاءمة للسياق.

Investigadores han presentado el Decodificador Contrastivo Consciente de Alucinaciones (CHASD), un nuevo marco diseñado para mejorar el rendimiento de los Modelos de Visión-Lenguaje (LVLMs) al abordar el problema de las alucinaciones de objetos durante la inferencia. Este método se centra en calibrar dinámicamente la atención visual, permitiendo resultados más precisos y contextualmente relevantes.

Des chercheurs ont introduit le Décodage Contrastif Sensible aux Hallucinations (CHASD), un nouveau cadre conçu pour améliorer la performance des Modèles de Vision-Langage (LVLMs) en abordant le problème des hallucinations d'objets lors de l'inférence. Cette méthode se concentre sur la calibration dynamique de l'attention visuelle, permettant des résultats plus précis et contextuellement pertinents.

Researchers have introduced Contrastive Hallucination-Aware Step-wise Decoding (CHASD), a novel framework designed to enhance the performance of Large Vision-Language Models (LVLMs) by addressing the issue of object hallucinations during inference. This method focuses on calibrating visual attention dynamically, allowing for more accurate and contextually relevant outputs.