Research map / Vision-language and multimodal / Multimodal LLMs

FREAK: A Fine-grained Hallucination Evaluation Benchmark for Advanced MLLMs

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Unified any-modality language models · 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 FREAK, a 1,799-question benchmark of photorealistic counter-commonsense images (generated then edited with Seedream/SeedEdit) for fine-grained hallucination evaluation of MLLMs, and finds SOTA models reach only ~45% accuracy and that CoT reasoning degrades performance.

Problem

Existing hallucination benchmards (POPE, AMBER, HallusionBench, PhD) are saturated, binary-judged, or too narrow to measure fine-grained visual hallucination in state-of-the-art MLLMs.

Approach

A generate-then-edit pipeline: an LLM proposes counter-commonsense attribute descriptions, an image generator makes a normal image, and an image editor applies a localized counter-commonsense edit, followed by LLM-generated multiple-choice/free-form questions and human verification.

Key results

• Best model Gemini-2.5-Pro reaches 45.49% overall accuracy versus 86.71% human baseline; most mainstream models fall in 30-43% range.
• Enabling CoT reasoning lowers accuracy for most models (e.g., Qwen2.5-VL-72B drops 39.39 to 33.39) and raises hallucination rate.
• On a normal vs CCS image subset, accuracy of Qwen2.5-VL-7B falls from 86.04% to 34.28% and InternVL3-14B from 91.26% to 34.69%.

Surprising findings

• Reasoning/CoT mode generally degrades accuracy and increases hallucination rate; o3 gains only about 1% over GPT-4.1.
• Over 70% of CoT failures involve the model favoring an incorrect option from the start and reinforcing it.

Limitations

• Relatively small scale (about 1,800 questions) due to manual verification costs.
• Reliance on external image generation and editing models may introduce artifacts or biases.
• CoT analysis under fine-grained hallucination leaves room for deeper investigation.

Experimental setup

Datasets: ImageNet-1K

Benchmarks: FREAK

Metrics: accuracy hallucination rate consistency accuracy precision recall F1

Baselines: POPE AMBER HallusionBench PhD WHOOPS VLind-Bench VLMBias

Builds on: POPE AMBER HallusionBench PhD

Closest papers by meaning

• Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.92)
• GHOST: Hallucination-Inducing Image Generation for Multimodal LLMs (ICLR 2026 · similarity 0.92)
• AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models (ICLR 2025 · similarity 0.92)
• MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM (NeurIPS 2025 · similarity 0.91)
• Learning to Decode Against Compositional Hallucination in Video Multimodal Large Language Models (ICML 2026 · similarity 0.91)
• Fine-Grained Multi Image Object Hallucination Benchmark (CVPR 2026 · similarity 0.91)