Research map / Vision-language and multimodal / Multimodal LLMs

Analyzing and Mitigating Object Hallucination in Large Vision-Language Models

ICLR 2024 · Vision-language and multimodal › Multimodal LLMs · approach: Preference tuning against hallucination · 29 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.

Proposes LURE, a post-hoc hallucination revisor trained on GPT-3.5-generated hallucinated captions, guided by co-occurrence, uncertainty, and position factors, to rectify object hallucination in six LVLMs' descriptions.

Problem

Large vision-language models generate descriptions containing objects not present in the image (object hallucination), harming downstream tasks; existing fixes require costly fine-tuning data curation.

Approach

A lightweight post-hoc revisor, trained to reconstruct correct captions from synthetic hallucinated ones, where hallucinations are induced by inserting co-occurring objects and masking uncertain or late-position objects with an [IDK] tag; the design is motivated by statistical and theoretical analysis of co-occurrence, uncertainty, and position.

Key results

• On MiniGPT-4, LURE reduces CHAIR_S from 26.8 to 19.7 and CHAIR_I from 7.3 to 4.9, versus 24.0/5.7 for the best baseline (Teacher).
• On LLaVa, CHAIR_S drops from 54.0 to 27.1 and on mPLUG-Owl from 71.2 to 18.8.
• Fine-tuning MiniGPT-4 on the same added data gives CHAIR_S 31.0, worse than the original 26.8, indicating gains come from post-hoc revision rather than extra data.

Measured effects reported in the paper

Surprising findings

• Fine-tuning the original LVLM on the revisor's additional data increases CHAIR_S (26.8 to 31.0), suggesting naive fine-tuning hurts hallucination.
• Greedy decoding alone does not consistently reduce hallucination; on LLaVa it raises CHAIR_S from 54.0 to 50.9 only modestly and on MMGPT it is near baseline.

Limitations

• CHAIR is restricted to COCO's 80 object categories, limiting automatic evaluation elsewhere
• Theoretical analysis relies on simplified linear models and Gaussian assumptions
• Relies on GPT-3.5 for synthetic hallucination data generation and GPT-3.5 for evaluation

Experimental setup

Datasets: MSCOCO 2014 LLaVA-150k

Benchmarks: CHAIR MSCOCO 2014 (5000 images) ImageNet CC POPE MME

Metrics: CHAIR_S CHAIR_I BLEU ROUGE-L BERTScore CLIP score METEOR CIDEr SPICE human hallucination ratio

Baselines: Original Teacher Chain-of-Thought Greedy-Decoding GPT-Ensemble GPT-Teacher

Builds on: MiniGPT-4 denoising autoencoders CHAIR

Closest papers by meaning

• Multi-Object Hallucination in Vision Language Models (NeurIPS 2024 · similarity 0.94)
• Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.94)
• Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance (ICML 2025 · similarity 0.93)
• Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens (CVPR 2025 · similarity 0.93)
• Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · similarity 0.93)
• Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · similarity 0.93)