Research map / Vision-language and multimodal / Multimodal LLMs
Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning
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.
Builds a controlled image-editing pipeline producing 50K+ minimally edited image-caption pairs (MED) and fine-tunes open MLLMs with SFT plus a feature-consistency idea to improve fine-grained difference detection, evaluated on a 200-item Micro Edit Detection benchmark.
Problem
MLLMs hallucinate or miss subtle semantic changes between minimally different images, which is attributed to a lack of controlled minimal-edit training data and misaligned learning objectives.
Approach
Generate minimally edited image pairs with Gemini Flash 2.0 guided by 11 hallucination-prone edit categories, align captions via a multi-step MLLM pipeline, and fine-tune MLLMs on the resulting difference QA data with a feature-level consistency framing.
Key results
- Qwen2.5-VL-7B (Ours) reaches 51.61% on the MED benchmark, up from 39.74% (+11.87 in table; text says +11.37), comparable to GPT-4o at 51.32%.
- Qwen2.5-VL-7B (Ours) averages 58.52 across MMStar, BLINK, POPE, MMVP and Count sub-tasks vs 55.47 for the base model, and reaches MME 1701.87 vs 1685.14.
- On the 35-pair MED-Real Set, Qwen2.5-VL-7B improves from 68.57% to 74.29% and LLaVA-V1.6-7B from 34.29% to 45.71%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune on MED using the proposed method | Qwen2-VL-7B · MED benchmark | accuracy (Avg): 38.48 -> 47.55 | Helped |
| Fine-tune on MED using the proposed method | Qwen2.5-VL-7B · MED benchmark | accuracy (Avg): 39.74 -> 51.61 | Helped |
| Fine-tune on MED using the proposed method | LLaVA-V1.6-7B · MED benchmark | accuracy (Avg): 31.04 -> 40.44 | Helped |
| Fine-tune on MED using the proposed method | LLaMA-3.2-11B · MED benchmark | accuracy (Avg): 34.71 -> 40.92 | Helped |
| Fine-tune on MED using the proposed method | Qwen2-VL-7B · MM benchmarks | Ave: 54.35 -> 56.25 | Helped |
| Fine-tune on MED using the proposed method | Qwen2.5-VL-7B · MM benchmarks | Ave: 55.47 -> 58.52 | Helped |
| Fine-tune on MED using the proposed method | LLaVA-V1.6-7B · MM benchmarks | Ave: 47.56 -> 48.71 | Helped |
| Fine-tune on MED using the proposed method | LLaMA-3.2-11B · MM benchmarks | Ave: 42.13 -> 43.82 | Helped |
| Fine-tune on MED using the proposed method | Qwen2-VL-7B · MED-Real Set | accuracy (%): 82.86 -> 82.86 | No clear effect |
| Fine-tune on MED using the proposed method | Qwen2.5-VL-7B · MED-Real Set | accuracy (%): 68.57 -> 74.29 | Helped |
| Fine-tune on MED using the proposed method | LLaVA-V1.6-7B · MED-Real Set | accuracy (%): 34.29 -> 45.71 | Helped |
| Fine-tune on MED using the proposed method | LLaMA-3.2-11B · MED-Real Set | accuracy (%): 40.00 -> 51.43 | Helped |
Surprising findings
- Qwen2-VL-7B drops on the MED-Real Set after fine-tuning remains unchanged at 82.86%, and fine-grained MMStar scores decline for some models after fine-tuning.
- The stated feature-level consistency loss is not written out in the training objective; the SFT loss in the paper is a captioning loss on the difference between image features.
Limitations
- Focuses on binary edits between image pairs; multi-step and temporal or compositional edits are not covered.
- Full fine-tuning of vision and language components is used; efficient adaptation is left for future work, though LoRA is listed in hyperparameter tables.
- Benchmark is small (165 synthetic plus 35 real items), and training and benchmark share the same synthetic editing pipeline.
Experimental setup
Datasets: DOCCI Visual Genome MMVP
Benchmarks: Micro Edit Detection (MED) benchmark MED-Real Set MMStar BLINK POPE MME MMVP
Metrics: accuracy average score MME perception score
Baselines: GPT-4o GPT-4.1 Claude 3.7 Sonnet Qwen-VL-Plus Doubao-1.5-Vision-Pro
Builds on: MMVP LLaMA-Factory DOCCI Visual Genome
Closest papers by meaning
- ZINA: Multimodal Fine-grained Hallucination Detection and Editing (CVPR 2026 · similarity 0.93)
- Hallucination-aware Intermediate Representation Editing in Large Vision-Lanugage Models (ICLR 2026 · similarity 0.92)
- AFTER: Mitigating the Object Hallucination of LVLM via Adaptive Factual-Guided Activation Editing (ICLR 2026 · similarity 0.92)
- When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding (NeurIPS 2025 · similarity 0.92)
- Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024 · similarity 0.92)
- SAVER: Mitigating Hallucinations in Large Vision-Language Models via Style-Aware Visual Early Revision (AAAI 2026 · similarity 0.91)
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.