MM-Snowball: Evaluating and Mitigating Hallucination Snowballing in Multimodal Multi-Turn Dialogue

Multimodal large language models (MLLMs) demonstrate remarkable visual understanding, yet their reliability in interactive settings is severely undermined by hallucination snowballing: a phenomenon where initial errors amplify across conversational turns, leading to a collapse in coherence. This failure reveals a fundamental vulnerability where models progressively neglect visual grounding in favor of over-relying on polluted textual history. Existing benchmarks are predominantly confined to single-turn VQA, which fail to capture the complex dynamics of error propagation in long-horizon interactions. To address this, we introduce MM-Snowball, the first benchmark for fine-grained diagnosis of hallucination snowballing within dialogues. Extensive evaluation shows that our benchmark poses a significant challenge even to advanced MLLMs and reveals the inefficacy of existing mitigation methods designed for single-turn VQA. To counteract this degradation, we propose Conflict-Aware Visual Rectification (CAVR). This training-free method mitigates snowballing through a synergistic dual-mechanism that refreshes visual grounding at the representation level and rectifies output distributions at the logit level, effectively re-anchoring the model to visual facts. Experiments demonstrate that CAVR achieves state-of-the-art performance, offering a promising path toward more reliable interactive AI. Data and code are available at: https://frenkie-chiang.github.io/MM-Snowball

AI systems that understand both images and text are becoming increasingly useful in everyday applications, such as visual assistants, education tools, and interactive search. However, these systems can still make mistakes when answering questions about images. More importantly, in a multi-turn conversation, one early mistake may influence later answers, causing the system to move further and further away from what is actually shown in the image. We call this problem “hallucination snowballing.” In this paper, we study this issue in detail. We introduce MM-Snowball, a new benchmark designed to evaluate how visual-language AI systems behave across longer conversations, rather than only testing them with a single question and answer. Our results show that even strong existing models struggle with this setting, and that methods designed to reduce mistakes in single-question tasks are not enough to prevent errors from accumulating over time. To address this problem, we propose a new method called Conflict-Aware Visual Rectification. The key idea is to help the model return its attention to the image when the conversation history becomes misleading. Our method does not require additional training and can improve the reliability of model responses by better grounding them in the visual content. Experiments show that this approach reduces error accumulation and leads to more trustworthy performance in interactive visual conversations.