Research map / Vision-language and multimodal / Multimodal LLMs
MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM
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 MIRAGE, a 1,329-question benchmark that isolates reasoning hallucinations in MLLMs (questions where images are correctly perceived), with accuracy, step/claim factuality, and an LLM hallucination score, plus Logos, a GRPO-based baseline using curriculum RL and collaborative hint inference.
Problem
Existing multimodal hallucination benchmarks conflate perception-induced and reasoning-induced hallucinations and evaluate only answers or intermediate steps, so reasoning failures in MLLMs are hard to diagnose.
Approach
Construct questions whose images are verified as correctly described by MLLMs but still lead to reasoning errors, annotate multi-level ground truth (answers, steps, claims, reasoning chains), and score hallucination at answer, step/claim, and whole-chain levels; then train Logos with curriculum GRPO (difficulty-filtered) and hint-based inference.
Key results
- Logos-7B reaches 37.1 accuracy on MIRAGE versus 28.8 for its base Qwen2.5-VL-7B-Instruct, with step factuality 43.3 and claim factuality 38.3.
- Gemini-2-Flash-Thinking has the highest LHS (0.7517) and O1 has the highest accuracy among black-box models (49.7), while the best open 7B-class result is Logos-7B.
- Manually corrected reasoning chains on about 10% of questions raise GPT-4o accuracy from 12.1 to 68.5 and Qwen2.5-VL-72B from 10.4 to 72.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train Qwen2.5-VL-7B with Logos | Qwen2.5-VL-7B-Instruct · MIRAGE | accuracy: 28.8 -> 37.1 | Helped |
| Increase Qwen2.5-VL model size from 3B to 72B | Qwen2.5-VL · MIRAGE | accuracy: 18.8 to 38.8 | Helped |
| Use Qwen2.5-VL rather than Qwen2-VL pretraining | Qwen-VL · MIRAGE | logical hallucination rate: 59.3% -> 47.7% | Helped |
| Add GRPO training (ablation) | Qwen2.5-VL-7B · MIRAGE | accuracy: 28.8 -> 33.7 | Helped |
| Add CRFT to GRPO (ablation) | Logos-7B · MIRAGE | accuracy: 33.7 -> 35.7 | Helped |
| Add both topic-specific and question-specific hints to the base model without GRPO or CRFT (ablation) | Qwen2.5-VL-7B · MIRAGE | accuracy: 28.8 -> 29.0 | No clear effect |
| Add online reward filtration (ablation) | Logos-7B · MIRAGE | accuracy: 34.2 -> 37.1 | Helped |
| Use GRPO instead of PPO (ablation) | Logos-7B without CHI · MIRAGE | accuracy: 29.9 -> 35.7 | Helped |
| Use DAPO instead of GRPO (ablation) | Logos-7B without CHI · MIRAGE | accuracy: 35.7 -> 30.0 | Hurt |
| Add KL-divergence regularization (ablation) | Logos-7B · MIRAGE | accuracy: 37.1 -> 31.0 (1e-2); 35.0 (1e-3); 36.7 (1e-4) | Hurt |
| Replace CRFT with vanilla RL for the same total training epochs (ablation) | Logos-7B · MIRAGE | accuracy: 37.1 -> 35.5 | Hurt |
| Increase curriculum stage from k=1 to k=2 or k=3 (ablation) | Logos-7B · MIRAGE | accuracy: 37.1 -> 37.2 | No clear effect |
Surprising findings
- Training-free methods (self-reflection, visual inference chain) degrade accuracy and LHS on the base 7B model, and CHI on the base model yields no gain.
- Spatial hallucinations show little improvement across model scale, data scale, or training stage, and RL does not reduce factuality or spatial hallucinations.
Limitations
- MIRAGE contains only single-image questions, leaving temporal and cross-image hallucinations unexplored.
- Theoretical analysis of why MLLMs produce reasoning hallucinations is insufficient.
Experimental setup
Datasets: MIRAGE (constructed from public datasets and internet questions) LIMO MathVista
Benchmarks: MIRAGE MathVista
Metrics: accuracy step factuality F1 claim factuality F1 LLMs Hallucination Score
Baselines: Gemini-2-Flash-Thinking O1 Gemini-2-Flash GPT-4o Qwen2.5-VL-72B-Instruct Virgo-72B QvQ-72B-Preview R1-OneVision-7B Mulberry-Qwen2-VL-7B Llava-CoT-11B
Builds on: GRPO DAPO MathVista
Closest papers by meaning
- Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.94)
- More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models (NeurIPS 2025 · similarity 0.94)
- Fine-Grained Multi Image Object Hallucination Benchmark (CVPR 2026 · similarity 0.93)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.93)
- Understanding and Mitigating Hallucinations in Multimodal Chain-of-Thought Models (CVPR 2026 · similarity 0.92)
- HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models (CVPR 2024 · similarity 0.92)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.