Reproduce “MM-Snowball: Evaluating and Mitigating Hallucination Snowballing in Multimodal Multi-Turn Dialogue” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yue Jiang, Xue Jiang, Lihua Zhang, Zhiqiang Wang, Yuhang Lu, Peng Wang, Bo Han, Feng Zheng and 1 more
Researchers at Fudan University and collaborators introduce MM-Snowball, a diagnostic benchmark to evaluate "hallucination snowballing" in long-horizon, multi-turn multimodal dialogues, and propose Conflict-Aware Visual Rectification (CAVR). MM-Snowball reveals MLLMs exhibit a "V-shaped" performance curve where visual information is suppressed by textual history, while CAVR consistently achieves state-of-the-art mitigation by dynamically re-anchoring models to visual facts.
Key insights
- Evaluations on MM-Snowball revealed a consistent "V-shaped" performance curve across 11 advanced MLLMs, indicating an initial accuracy drop, divergence in handling adversarial queries, further degradation, and eventual recovery upon explicit visual re-grounding.
- The phenomenon of "Visual Fading" was identified: original visual information is not permanently forgotten but rather progressively suppressed by the accumulating weight of hallucinated textual history, highlighting a dynamic imbalance between visual evidence and narrative inertia.
- Resistance to hallucination snowballing depends on both the MLLM's single-turn visual understanding capabilities and its capacity for cross-modal conflict detection within the evolving dialogue history.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.