Research map / Vision-language and multimodal / Multimodal LLMs
Hallucination Augmented Contrastive Learning for Multimodal Large Language Model
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 HACL, which adds cross-modal contrastive learning with GPT-4-generated hallucinative captions as hard negatives during MLLM pretraining to reduce hallucinations and improve benchmark performance of MiniGPT-4, LLaVA, and LLaVA-1.5.
Problem
MLLMs often generate erroneous or fabricated content inconsistent with visual input (hallucination), and the learned visual-to-text interface leaves a modality gap and entangles hallucinated and correct text representations.
Approach
Introduce image-text contrastive learning into the first-stage pretraining of MLLMs, using GPT-4-generated hallucinated captions as extra hard negatives so that visual representations align with correct text and move away from hallucinated text in the LLM's last-token representation space.
Key results
- On MMHal-Bench, LLaVA-HACL improves overall score from 1.55 to 2.08 (+0.53) and cuts hallucination rate from 0.76 to 0.62; MiniGPT-4-HACL reaches 1.80 from 1.39.
- On POPE, LLaVA-HACL raises average F1 from 66.48 to 78.31 in the ablation table.
- LLaVA1.5-HACL reaches MMHal-Bench score 2.13 (rate 0.50) and MME 1530.10 vs 1510.70 for LLaVA-1.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HACL | LLaVA-7B · MMHal-Bench | overall score: 1.55 -> 2.08 | Helped |
| add HACL | LLaVA-7B · MMHal-Bench | hallucination rate: 0.76 -> 0.62 | Helped |
| add HACL | MiniGPT-4-7B · MMHal-Bench | overall score: 1.39 -> 1.80 | Helped |
| add HACL | LLaVA1.5-7B · MMHal-Bench | overall score: 2.08 -> 2.13 | Helped |
| add HACL | LLaVA · POPE | F1 score: 66.48 -> 78.31 | Helped |
| add HACL | LLaVA · MME | score: 502.82 -> 562.58 | Helped |
| add HACL | LLaVA · VQAv2 | accuracy: 71.3 -> 73.3 | Helped |
| add contrastive learning without hallucinative captions (ablation) | LLaVA · MMHal-Bench | overall score: 1.55 -> 1.67 | Helped |
| add hallucinative captions as contrastive hard negatives (ablation) | LLaVA with contrastive learning · MMHal-Bench | overall score: 1.67 -> 2.08 | Helped |
| add hallucinative captions as contrastive hard negatives (ablation) | MiniGPT-4 with contrastive learning · POPE | F1 score: 53.54 -> 77.54 | Helped |
| add hallucinative captions as contrastive hard negatives (ablation) | LLaVA1.5 with contrastive learning · MMHal-Bench | overall score: 2.09 -> 2.13 | Helped |
| train the visual encoder during Stage 1 (ablation) | LLaVA-HACL · POPE | F1 score: 78.31 -> 63.42 | Hurt |
Surprising findings
- Using only contrastive learning without hallucinative captions yields smaller gains, and hallucinative hard negatives are needed for the largest hallucination reduction.
- Unfreezing the LLM during the first pretraining stage significantly degrades performance, attributed to catastrophic forgetting.
Limitations
- Hallucinative captions for MiniGPT-4 were generated for only about 10% of the pretraining data due to generation cost.
- Gains on MMHal-Bench and POPE rely on GPT-4-based evaluation and generated negatives; the paper does not release code/data explicitly in the text beyond a GitHub link.
- Reported results on some benchmarks show mixed changes, e.g., LLaVA-HACL recall drops on POPE and MM-Vet slightly decreases for LLaVA1.5-HACL.
Experimental setup
Datasets: LAION Conceptual Captions LLaVA pretraining subsets of LAION/CC/SBU COCO val2017
Benchmarks: MMHal-Bench POPE VQAv2 GQA VizWizQA TextVQA SciQA MME MMBench MM-Vet SEED-Bench
Metrics: overall score hallucination rate accuracy F1 yes ratio VQA accuracy benchmark score
Baselines: Kosmos-2 IDEFICS InstructBLIP LLaVA-RLHF Shikra MM-GPT mPLUG-Owl BLIP-2 Qwen-VL-Chat LLaVA MiniGPT-4 LLaVA-1.5
Builds on: ALBEF MoCo LLaVA MiniGPT-4 POPE
Closest papers by meaning
- ConVis: Contrastive Decoding with Hallucination Visualization for Mitigating Hallucinations in Multimodal Large Language Models (AAAI 2025 · similarity 0.93)
- Mitigating Multimodal Hallucinations via Gradient-based Self-Reflection (CVPR 2026 · similarity 0.92)
- Combating Multimodal LLM Hallucination via Bottom-Up Holistic Reasoning (AAAI 2025 · similarity 0.92)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.92)
- CODE: Contrasting Self-generated Description to Combat Hallucination in Large Multi-modal Models (NeurIPS 2024 · similarity 0.92)
- Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024 · 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.