Research map / Vision-language and multimodal / Multimodal LLMs
When Diffusion Language Models Hesitate: Detecting and Correcting Visual Hallucinations via Confidence Fluctuation
Wenzheng Song, Pei Chen, Yichen Tan, Zejian Li, Lingyun Sun
Read the paperCheck your idea against 137,217 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.
Studies visual hallucinations in multi-modal diffusion language models (MDLMs) and proposes detecting and correcting them via confidence fluctuation during decoding.
Problem
Multi-modal diffusion language models suffer from visual hallucinations because their visual perception is static; unlike autoregressive models, they lack sequential dependency for dynamically attending to visual input.
Approach
Uses fluctuations in model confidence during the iterative diffusion decoding process as a signal to detect hallucinated tokens and correct them. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add VGR (fine-tuning plus refinement loop) | LLaDA-V · CapArena | CapArena-Auto: -77.17 -> -52.17 | Improved |
| add VGR | LLaDA-V · AMBER-g | Hal: 44.9 -> 37.2 | Improved |
| add VGR | LLaDA-V · DetailCaps-4870 | CAPTURE: 59.62 -> 62.35 | Improved |
| add VGR | LLaDA-V · CapMAS | Avg. score: 58.61 -> 60.84 | Improved |
| add VGR-Instruct fine-tuning (ablation) | LLaDA-V · CapArena | CapArena-Auto: -77.17 -> -61.33 | Improved |
| add the inference-time refinement loop after fine-tuning (ablation) | LLaDA-V (SFT) · AMBER-g | Cover: 63.7 -> 64.5 | Improved |
| localize with high CF instead of low confidence (ablation) | VGR refinement · CapArena | CapArena-Auto: -57.83 -> -52.17 | Improved |
| combine high CF and low confidence in a hybrid (ablation) | VGR refinement · CapArena | CapArena-Auto: -52.17 -> -56.50 | Hurt |
| add local view to global view (dual-view) (ablation) | VGR refinement · CapArena | CapArena-Auto: -53.67 -> -52.17 | Improved |
| use syntax-aware span remasking instead of fixed window (ablation) | VGR refinement · CapArena | CapArena-Auto: -59.50 -> -52.17 | Improved |
| raise visual gate threshold from 0.4 to 0.6 (ablation) | VGR refinement · AMBER | Correction Rate: 97.4% -> 93.3% | Hurt |
| apply VGR to MMaDA | MMaDA · AMBER-g | Hal: 72.8 -> 39.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- LLaDA-V — LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning (CVPR 2026) · better on 9 of 9 shared results
- MMaDA — MMaDA: Multimodal Large Diffusion Language Models (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DEEM: Diffusion models serve as the eyes of large language models for image perception (ICLR 2025 · similarity 0.92)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.91)
- VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding (CVPR 2025 · similarity 0.91)
- SECOND: Mitigating Perceptual Hallucination in Vision-Language Models via Selective and Contrastive Decoding (ICML 2025 · similarity 0.91)
- SAVER: Mitigating Hallucinations in Large Vision-Language Models via Style-Aware Visual Early Revision (AAAI 2026 · similarity 0.91)
- Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy (ICML 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.
