Research map / Vision-language and multimodal / Multimodal LLMs

Robust Multimodal Large Language Models Against Modality Conflict

ICML 2025 · 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 modality conflict (object, attribute, relationship mismatches between image and text) as a source of MLLM hallucination, builds the MMMC dataset (20K triples), and compares prompt engineering, SFT, and REINFORCE++-based RL for mitigating it.

Problem

MLLMs hallucinate when the textual question presupposes content that conflicts with the image (inputs conflicting with each other), a source not addressed by work focused on response-input conflicts.

Approach

Formally defines modality conflict via object, attribute, and relationship mismatches between visual and textual inputs, builds a counterfactual-question dataset from Visual Genome using LLM-based substitution, and tests three mitigation methods (prompt engineering, LoRA SFT, and RL with an LLM-judged binary reward).

Key results

• RL reduces Hallu-Rate for LLaVA-v1.5-7B from 93.25% to 33.87% (Llama judge) and raises LLM-Judge to 2.58, the best among methods on that model.
• For Qwen2-VL-Instruct-7B, RL gives Hallu-Rate 23.52% (Llama judge) versus 52.35% base, while PE reaches 40.10% and SFT 28.58%.
• Prompt engineering gives little or no gain for InstructBLIP-7B and GPT-4o and even increases Hallu-Rate for Qwen2-VL-Instruct-2B (46.55% to 62.10%).

Measured effects reported in the paper

Surprising findings

• Prompt engineering increases hallucination rate for the Qwen2-VL-Instruct-2B model (46.55% to 62.10% by Llama judge).
• RL on InstructBLIP-7B collapsed into repetitive, long outputs around 300 episodes, and RL training on LLaVA-NeXT-7B plateaued after 10,000 episodes.

Limitations

• Hallucination and quality judged by LLM judges (GPT-4o and Llama-3.3-70B) and a reward model, with binary reward based on LLM semantic consistency.
• RL is unstable and prone to collapse; longer training degrades performance for some models.
• Dataset is synthetic, generated by GPT-4o-mini from Visual Genome annotations and verified by human annotators; limited to object, attribute, and relationship conflicts.

Experimental setup

Datasets: Visual Genome MMMC

Benchmarks: MMMC

Metrics: ROUGE-L Hallu-Rate LLM-Judge score

Baselines: GPT-4o InstructBLIP-7B LLaVA-v1.5-7B LLaVA-NeXT-7B Qwen2-VL-Instruct-2B Qwen2-VL-Instruct-7B

Builds on: REINFORCE++ LoRA OpenRLHF Visual Genome

Closest papers by meaning

• The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio (NeurIPS 2025 · similarity 0.93)
• Mitigating Modality Prior-Induced Hallucinations in Multimodal Large Language Models via Deciphering Attention Causality (ICLR 2025 · similarity 0.93)
• Ground What You See: Hallucination-Resistant MLLMs via Caption Feedback, Diversity-Aware Sampling, and Conflict Regularization (AAAI 2026 · similarity 0.92)
• Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.92)
• AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models (ICLR 2025 · similarity 0.92)
• IRIS: Implicit Reward-Guided Internal Sifting for Mitigating Multimodal Hallucination (ICML 2026 · similarity 0.92)