Research map / Vision-language and multimodal / Multimodal LLMs

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

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Decoding-time hallucination mitigation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MM-Snowball, a six-turn multimodal dialogue benchmark for hallucination snowballing built via adversarial trajectory synthesis, and CAVR, a training-free decoding method combining entropy-triggered visual FFN re-injection with JSD-based logit blending.

Problem

MLLMs' hallucinations compound across multi-turn dialogue (hallucination snowballing), but existing hallucination benchmarks are single-turn and existing mitigation methods target single-turn VQA and degrade in multi-turn settings.

Approach

Builds a progressive six-turn adversarial dialogue benchmark via causal state perturbation of visual atomic propositions, and proposes a training-free dual mechanism: entropy-triggered visual retracing inside FFNs (RVR) plus JSD-weighted interpolation between full-context and history-free visual logits (LCR).

Key results

• On LLaVA-1.5-7B, CAVR reaches 2.00 accuracy at Turn-3 versus 1.70 for the base model and 35.25 at Turn-6 versus 33.10, with SRS 0.15 vs 0.00.
• On Qwen2.5-VL-7B, CAVR achieves 47.85 at Turn-6 (base 46.90) and SRS 0.30 vs 0.15, the best among compared training-free methods.
• Among evaluated models, Seed-1.6 and Kimi K2.5 keep SRS of 41.70 and 36.95, while most open models score below 5 SRS.

Surprising findings

• Accuracy recovers sharply at Turn-6 after collapse at Turns 3-5 across nearly all models, suggesting visual information is suppressed rather than forgotten.
• Several open-source models (e.g., LLaVA-1.5-7B) fall below 2% accuracy at middle turns, while some proprietary models hold up much better.

Limitations

• CAVR gains over baselines are small in absolute terms (often below 1 point), and absolute accuracy remains very low for middle turns.
• Benchmark relies on GPT-5 as the judge; human agreement is reported only as a Pearson correlation with three experts.
• Evaluation fixed at six turns; the paper's claimed extensibility to longer dialogues is not empirically tested.

Experimental setup

Benchmarks: MM-Snowball

Metrics: accuracy per turn Visual Fading Ratio Snowballing Resistance Score

Baselines: VCD OPERA DoLA DeCo MemVR RVD

Builds on: VCD OPERA DoLA MemVR Residual Visual Decoding

Closest papers by meaning

• Visual Multi-Agent System: Mitigating Hallucination Snowballing via Visual Flow (ICLR 2026 · similarity 0.90)
• When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding (NeurIPS 2025 · similarity 0.90)
• Look Carefully: Adaptive Visual Reinforcements in Multimodal Large Language Models for Hallucination Mitigation (ICLR 2026 · similarity 0.90)
• Seeing Far and Clearly: Mitigating Hallucinations in MLLMs with Attention Causal Decoding (CVPR 2025 · similarity 0.90)
• HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models (CVPR 2024 · similarity 0.89)
• Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.89)