# Importance Sampling for Multi-Negative Multimodal Direct Preference Optimization

arXiv: https://arxiv.org/abs/2509.25717 · HTML: https://mlsanity.com/paper/2509.25717/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 33 of its 38 references that are in the map.
- First posted on arXiv 30 Sep 2025.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)
- LLM alignment / RLHF / preference optimisation / [Diffusion and multimodal model alignment](https://mlsanity.com/alignment/diffusion-and-multimodal-model-alignment/)

## Builds on (most cited first, 30 of 33)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [GPT-4 Technical Report](https://mlsanity.com/paper/2303.08774/) (2023)
- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Visual Instruction Tuning](https://mlsanity.com/paper/2304.08485/) (2023)
- [Qwen2.5-VL Technical Report](https://mlsanity.com/paper/2502.13923/) (2025)
- [Evaluating Object Hallucination in Large Vision-Language Models](https://mlsanity.com/paper/2305.10355/) (2023)
- [Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling](https://mlsanity.com/paper/2412.05271/) (2024)
- [SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension](https://mlsanity.com/paper/2307.16125/) (2023)
- [Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs](https://mlsanity.com/paper/2401.06209/) (2024)
- [Aligning Large Multimodal Models with Factually Augmented RLHF](https://mlsanity.com/paper/2309.14525/) (2023)
- [HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models](https://mlsanity.com/paper/2310.14566/) (2023)
- [VLMEvalKit: An Open-Source Toolkit for Evaluating Large Multi-Modality Models](https://mlsanity.com/paper/2407.11691/) (2024)
- [RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback](https://mlsanity.com/paper/2312.00849/) (2023)
- [MM-LLMs: Recent Advances in MultiModal Large Language Models](https://mlsanity.com/paper/2401.13601/) (2024)
- [SugarCrepe: Fixing Hackable Benchmarks for Vision-Language Compositionality](https://mlsanity.com/paper/2306.14610/) (2023)
- [Direct Preference Optimization with an Offset](https://mlsanity.com/paper/2402.10571/) (2024)
- [Disentangled Representation Learning](https://mlsanity.com/paper/2211.11695/) (2022)
- [mDPO: Conditional Preference Optimization for Multimodal Large Language Models](https://mlsanity.com/paper/2406.11839/) (2024)
- [WildVision: Evaluating Vision-Language Models in the Wild with Human Preferences](https://mlsanity.com/paper/2406.11069/) (2024)
- [Learning Concise and Descriptive Attributes for Visual Recognition](https://mlsanity.com/paper/2308.03685/) (2023)
- [Enhancing Large Vision Language Models with Self-Training on Image Comprehension](https://mlsanity.com/paper/2405.19716/) (2024)
- [On Softmax Direct Preference Optimization for Recommendation](https://mlsanity.com/paper/2406.09215/) (2024)
- [Direct Multi-Turn Preference Optimization for Language Agents](https://mlsanity.com/paper/2406.14868/) (2024)
- [Text encoders bottleneck compositionality in contrastive vision-language models](https://mlsanity.com/paper/2305.14897/) (2023)
- [CHiP: Cross-modal Hierarchical Direct Preference Optimization for Multimodal LLMs](https://mlsanity.com/paper/2501.16629/) (2025)
- [Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization](https://mlsanity.com/paper/2502.13146/) (2025)
- [Modality-Fair Preference Optimization for Trustworthy MLLM Alignment](https://mlsanity.com/paper/2410.15334/) (2024)
- [Enhancing Multimodal Compositional Reasoning of Visual Language Models with Generative Negative Mining](https://mlsanity.com/paper/2311.03964/) (2023)
- [The Hard Positive Truth about Vision-Language Compositionality](https://mlsanity.com/paper/2409.17958/) (2024)

## Built on by (most cited first, 5 of 5)

- [Skill-R1: Agent Skill Evolution via Reinforcement Learning](https://mlsanity.com/paper/2605.09359/) (2026)
- [WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning](https://mlsanity.com/paper/2602.17025/) (2026)
- [Generate, Filter, Control, Replay: A Comprehensive Survey of Rollout Strategies for LLM Reinforcement Learning](https://mlsanity.com/paper/2605.02913/) (2026)
- [MASS-DPO: Multi-negative Active Sample Selection for Direct Policy Optimization](https://mlsanity.com/paper/2605.10784/) (2026)
- [F-GRPO: Factorized Group-Relative Policy Optimization for Unified Candidate Generation and Ranking](https://mlsanity.com/paper/2605.12995/) (2026)

## Abstract (arXiv)

Direct Preference Optimization (DPO) has recently been extended from text-only models to vision-language models. However, existing methods rely on oversimplified pairwise comparisons, generating a single negative image via basic perturbations or similarity-based retrieval, which fail to capture the complex nature of multimodal preferences, inducing optimization bias and hallucinations. To address this issue, we propose MISP-DPO, the first framework to incorporate multiple, semantically diverse negative images in multimodal DPO via the Plackett-Luce model. Our method embeds prompts and candidate images in CLIP (Contrastive Language-Image Pretraining) space and applies a sparse autoencoder to uncover semantic deviations into interpretable factors. Negative samples are selected based on reconstruction difficulty, semantic deviation from the positive, and mutual diversity, yielding broader and more informative supervision. To handle multi-negative comparisons, we adopt a Plackett-Luce objective and introduce an importance sampling strategy that improves training efficiency. Experiments across five diverse benchmarks demonstrate that MISP-DPO consistently improves multimodal alignment over prior methods, validating the effectiveness of semantic-aware, multi-negative sampling in preference-based learning.