Research map / Vision-language and multimodal / Multimodal LLMs
Mitigating Object Hallucination in MLLMs via Data-augmented Phrase-level Alignment
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.
Proposes DPA, a phrase-level alignment loss with generative data augmentation and token-wise forward KL regularization, to reduce object hallucination in off-the-shelf MLLMs (LLaVA-v1.5, VILA-v1.5) while preserving general capabilities.
Problem
Object hallucination in multimodal LLMs (generating objects, attributes or actions not present in the image), where existing sequence-level preference-optimization fixes degrade general vision-language performance.
Approach
Construct hallucinated/correct response pairs by altering ground-truth phrases in correct responses, then apply a phrase-level contrastive log-likelihood loss on only the hallucinated phrases plus a token-level forward KL term to a frozen reference model, yielding HALVA.
Key results
- HALVA-7B reduces CHAIR sentence score from 50.0 to 41.4 relative to LLaVA-v1.5-7B and raises MME-Hall from 648.3 to 665.0.
- On AMBER discriminative tasks, HALVA-13B improves F1 from 73.1 to 86.5, matching or exceeding GPT-4V (87.4).
- HALVA preserves general benchmark performance (e.g., VQA-v2 unchanged at 78.5 for 7B), whereas HA-DPO and EOS significantly drop on TextVQA and MME.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune with phrase-level data-augmented alignment | LLaVA-v1.5-7B · AMBER | F1: 74.7 -> 83.4 | Helped |
| Fine-tune with phrase-level data-augmented alignment | LLaVA-v1.5-13B · AMBER | F1: 73.1 -> 86.5 | Helped |
| Fine-tune with phrase-level data-augmented alignment | LLaVA-v1.5-7B · CHAIR | C_s: 50.0 -> 41.4 | Helped |
| Fine-tune with phrase-level data-augmented alignment | LLaVA-v1.5-13B · MME-Hall | score: 643.3 -> 675.0 | Helped |
| Fine-tune with phrase-level data-augmented alignment | LLaVA-v1.5-7B · HallusionBench | overall accuracy: 47.09 -> 48.95 | Helped |
| Fine-tune with phrase-level data-augmented alignment | LLaVA-v1.5-7B · VQA-v2 | accuracy: 78.5 -> 78.5 | No clear effect |
| Use a very low KL coefficient, α=0.01 (ablation) | HALVA-Med · general vision-language tasks | language generation quality | Hurt |
| Set α in the 0.3 to 0.4 range (ablation) | HALVA-Med · MME-Hall | score | Helped |
| Set α in the 0.4 to 0.6 range (ablation) | HALVA-Large · MME-Hall | score | Helped |
| Use open-set hallucinated descriptions instead of closed-set samples (ablation) | HALVA-Med · CHAIR | C_s: 45.0 -> 39.6 | Helped |
| Add Yes/No hallucinated responses to closed-set and open-set augmentation (ablation) | HALVA-Med · AMBER | F1: 74.1 -> 83.4 | Helped |
| Use seen rather than unseen reference samples for KL divergence (ablation) | HALVA-Med · CHAIR | C_s: 47.4 -> 41.4 | Helped |
Surprising findings
- EOS and HA-DPO reduce hallucination rates but reduce object coverage and general benchmark scores, showing sequence-level alignment trade-offs.
- Using seen instruction-tuning data as KL reference samples works better than using the training pairs themselves.
Limitations
- Evaluated only on LLaVA-v1.5 and VILA-v1.5 base models
- Relies on Gemini Vision Pro for generating correct responses and on Visual Genome annotations
- Hallucination gains on some metrics (e.g., AMBER hallucination rate for 7B) are not uniform across benchmarks
Experimental setup
Datasets: Visual Genome VQA-v2 yes/no questions MME-Hall
Benchmarks: CHAIR MME-Hall AMBER MMHal-Bench HallusionBench VQA-v2 MM-Vet TextVQA MME LLaVA-Bench
Metrics: CHAIR instance and sentence scores F1 hallucination rate overall score accuracy yes/no bias coverage
Baselines: HA-DPO EOS LLaVA-RLHF VCD Woodpecker OPERA RLHF-V GPT-4V
Builds on: DPO LLaVA-v1.5 VILA-v1.5 HA-DPO EOS
Closest papers by meaning
- Look Carefully: Adaptive Visual Reinforcements in Multimodal Large Language Models for Hallucination Mitigation (ICLR 2026 · similarity 0.92)
- MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation (ICLR 2025 · similarity 0.92)
- Mitigating Object Hallucination via Concentric Causal Attention (NeurIPS 2024 · similarity 0.92)
- Multi-Modal Hallucination Control by Visual Information Grounding (CVPR 2024 · similarity 0.92)
- One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination (CVPR 2026 · similarity 0.92)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · 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.