Research map / Vision-language and multimodal / Multimodal LLMs
PerturboLLaVA: Reducing Multimodal Hallucinations with Perturbative Visual Training
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 HalFscore, a triplet/language-graph concept-level metric for hallucination in dense captions, and PerturboLLaVA, which fine-tunes LLaVA-1.5 with GPT-4o-generated misleading perturbation text inserted into the input (no loss on it) to reduce reliance on language priors.
Problem
Hallucinations in MLLMs during dense image captioning, where existing metrics (CHAIR, MMHalBench) are coarse and do not measure completeness, and current mitigations rely on costly RLHF/DPO or inference-time decoding.
Approach
Measure dense-caption quality via a GPT-4o-parsed concept graph (precision for hallucinations, recall for omissions, F-score aggregate), and reduce language-prior over-reliance by inserting adversarial, plausible-but-wrong perturbation text into the instruction-tuning input without computing loss on it.
Key results
- PerturboLLaVA improves HalFscore over LLaVA1.5 from 49.2 to 52.2 Fscore (precision 53.3 to 59.5) and reduces CHAIR_s from 54.2 to 36.1.
- General benchmarks improve: MMBench +1.6 (67.3 to 68.9), SEED +0.3, CCBench +1.2 over LLaVA1.5.
- Combining with OPERA yields the best HalFscore of 52.8 and CCBench 31.0 among compared variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · DCI 1,000-image evaluation | Fscore: 49.2 -> 52.2 | Helped |
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · DCI 1,000-image evaluation | Precision: 53.3 -> 59.5 | Helped |
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · DCI 1,000-image evaluation | Recall: 45.8 -> 46.5 | Helped |
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · HallusionBench | score: +0.6 | Helped |
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · MMBench | score: +1.6 | Helped |
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · SEEDImage | score: +0.3 | Helped |
| Add targeted adversarial perturbation text during training | LLaVA1.5 7B · CCBench | score: +1.2 | Helped |
| Combine PerturboLLaVA training with OPERA decoding | LLaVA1.5 7B · DCI 1,000-image evaluation | Fscore: improving HalFscore by 0.6 | Helped |
| Use image-focused instruction without warning about perturbation (ablation) | PerturboLLaVA · DCI 1,000-image evaluation | Fscore: 52.2 -> 52.0 | Hurt |
| Remove guidance around the perturbation text (ablation) | PerturboLLaVA · DCI 1,000-image evaluation | Fscore: 52.2 -> 52.0 | Hurt |
| Replace targeted perturbations with randomly sampled text (ablation) | PerturboLLaVA · DCI 1,000-image evaluation | Fscore: 52.2 -> 50.7 | Hurt |
| Increase perturbation strength from Version1 to Version3 (ablation) | PerturboLLaVA · MMBench | score: 68.9 -> 66.5 | Hurt |
Surprising findings
- RLAIF-V, despite its reward model, lowers general benchmark scores (MMB 63.7 vs 67.3), while the perturbation method improves them.
- Random (non-relevant) perturbation text still reduces hallucination somewhat, suggesting disruption itself helps training.
Limitations
- HalFscore relies on GPT-4o for triplet extraction and graph matching, and the evaluation uses only 1,000 DCI images.
- Relation hallucinations show limited improvement, and recall does not improve over the baseline.
- Experiments are limited to LLaVA-1.5 7B; the comparison with RLAIF-V is not fully fair because of the LLaVA-Next 34B reward model.
Experimental setup
Datasets: LLaVA-1.5 SFT data LLaVA-Pretrain 558k DCI (Densely Captioned Images) 1,000-image subset MSCOCO2014 val (Object HalBench)
Benchmarks: HalFscore (proposed) Object HalBench (CHAIR) HallusionBench MMBench SEED-Image CCBench
Metrics: precision recall F-score CHAIR_s CHAIR_i accuracy
Baselines: LLaVA1.5 RLAIF-V OPERA VCD Ovis1.6-Gemma2 Qwen2-VL LLaVA-OneVision InternVL2 Idefics3 MiniCPM-2.6
Builds on: LLaVA-1.5 RLAIF-V OPERA VCD CHAIR
Closest papers by meaning
- Toward Robust Hyper-Detailed Image Captioning: A Multiagent Approach and Dual Evaluation Metrics for Factuality and Coverage (ICML 2025 · similarity 0.92)
- Mitigating Multimodal Hallucinations via Gradient-based Self-Reflection (CVPR 2026 · similarity 0.91)
- Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning (NeurIPS 2025 · similarity 0.91)
- Hallucination Augmented Contrastive Learning for Multimodal Large Language Model (CVPR 2024 · similarity 0.91)
- PostAlign: Multimodal Grounding as a Corrective Lens for MLLMs (ICLR 2026 · similarity 0.91)
- Multi-Modal Hallucination Control by Visual Information Grounding (CVPR 2024 · 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.