Research map / Vision-language and multimodal / Multimodal LLMs

Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning

ICLR 2024 · Vision-language and multimodal › Multimodal LLMs · approach: Decoding-time hallucination mitigation · 23 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.

Builds LRV-Instruction, a 400k GPT-4-generated visual instruction dataset with positive and negative (hallucination-inducing) instructions, proposes GAVIE GPT-4-based evaluation without ground-truth answers, and finetunes MiniGPT4 and mPLUG-Owl to reduce hallucination.

Problem

Large multi-modal models hallucinate descriptions that conflict with the image and fail to follow instructions, partly because their visual instruction tuning data is largely positive, template-based and lacks negative instructions; existing hallucination evaluations need human-annotated ground trut

Approach

Generate a large, diverse visual instruction dataset with GPT-4 from Visual Genome dense captions, chart and news images, including negative instructions at three semantic levels (nonexistent object, existent object attribute manipulation, knowledge manipulation); evaluate open-ended outputs with GPT-4 scoring of accuracy and relevancy without ground-truth answers.

Key results

• Finetuned mPLUG-Owl-7B reaches GAVIE accuracy 6.58 and relevancy 8.46, versus 5.93 and 7.34 for InstructBLIP-13B, the best baseline.
• On POPE random set, finetuned mPLUG-Owl-7B reaches accuracy 0.86 and F1 0.88, matching or exceeding InstructBLIP-13B (0.86/0.87).
• A 1:1 positive-to-negative training ratio gives the best balance, with Acc_pos 0.92 and Acc_neg 0.85.

Measured effects reported in the paper

Surprising findings

• Existent object manipulation and knowledge manipulation instructions are harder for LMMs than nonexistent object manipulation.
• Training on all-positive data leaves Acc_neg at 0.05, while all-negative data drops Acc_pos to 0.10, showing strong yes-bias trade-offs.

Limitations

• Finetuning was only done on 7B models (MiniGPT4 and mPLUG-Owl) because of compute limits.
• GAVIE relies on GPT-4 judgments and dense captions as image proxy, and GPT-4 generated answers can contain errors (85% acceptable in a 500-sample check).
• Evaluation set is small (1000 instances) and some public benchmark results are reported only for a subset of models.

Experimental setup

Datasets: LRV-Instruction Visual Genome chart images dataset with captions news images with captions

Benchmarks: GAVIE MME POPE AMBER GQA Hallusionbench

Metrics: accuracy relevancy score F1 standard deviation

Baselines: MiniGPT4 LLaVA InstructBLIP mPLUG-Owl MMGPT MiniGPTv2 LLaVA1.5

Builds on: LLaVA MiniGPT-4 mPLUG-Owl Visual Genome POPE

Closest papers by meaning

• Mitigating Object Hallucination via Concentric Causal Attention (NeurIPS 2024 · similarity 0.92)
• Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.92)
• Learning to Decode Against Compositional Hallucination in Video Multimodal Large Language Models (ICML 2026 · similarity 0.92)
• Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models (ICLR 2026 · similarity 0.91)
• Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.91)
• Instruction Lens Score: Your Instruction Contributes a Powerful Object Hallucination Detector for Multimodal Large Language Models (ICML 2026 · similarity 0.91)