Title: MM-Snowball: Assessing and Alleviating the Cascade of Hallucinations in Multimodal Multi-Turn Conversations

Abstract:

While Multimodal Large Language Models (MLLMs) exhibit impressive capabilities in visual comprehension, their dependability during interactive exchanges is significantly compromised by "hallucination snowballing." This specific issue involves initial mistakes escalating over successive dialogue turns, ultimately resulting in a breakdown of logical consistency. Such failures highlight a critical weakness: as conversations progress, models tend to discard visual evidence in favor of relying too heavily on corrupted textual context. Current evaluation datasets are largely limited to single-turn Visual Question Answering (VQA), meaning they do not adequately reflect the intricate error propagation mechanisms found in extended interactions.

To bridge this gap, we present MM-Snowball, a novel benchmark designed for the detailed diagnosis of hallucination snowballing within conversational frameworks. Our comprehensive assessments indicate that this benchmark presents a substantial hurdle even for leading MLLMs, while also exposing the inadequacy of current mitigation strategies tailored for single-turn VQA tasks. In response to this performance decline, we introduce Conflict-Aware Visual Rectification (CAVR). As a training-free approach, CAVR combats snowballing via a dual mechanism that works synergistically: it updates visual grounding at the representation level and corrects output distributions at the logit level, thereby effectively reconnecting the model with visual realities. Our results show that CAVR delivers state-of-the-art performance, providing a viable route toward more trustworthy interactive AI systems. The associated data and code can be accessed at: https://frenkie-chiang.github.io/MM-Snowball

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC