PLUME: Latent Reasoning Based Universal Multimodal Embedding

🏡 Project Page | 📄 Paper | 🤗 Model | 🤗 Training Data | 🤗 Eval Data

PLUME is a latent reasoning framework for universal multimodal embedding (UME). It replaces explicit chain-of-thought (CoT) generation with a short autoregressive rollout of continuous latent states, combined with a semantic-anchor-guided transition adapter (Latent MoE) and a progressive explicit-to-latent curriculum. Built on Qwen2-VL-2B, PLUME achieves 61.6 on the 78-task MMEB-v2 benchmark while delivering over 30x faster inference compared to explicit-CoT methods.

This repository is the official implementation of the paper PLUME: Latent Reasoning Based Universal Multimodal Embedding.

📰 News

• [2026/04] Paper released on arXiv.
• [2025] Code and model weights released.

📝 TODO

• Code Released: Training and evaluation pipeline.
• Model Released: Pre-trained PLUME-Qwen2-VL-2B weights.
• Paper: arXiv.

💡 Highlights

• Replaces hundreds of explicit reasoning tokens with only 8 latent steps, delivering 30.3x faster inference
• 61.6 overall on the 78-task MMEB-v2 benchmark, surpassing UME-R1 (60.1) and VLM2Vec-V2 (58.0)
• Curriculum-based latent reasoning -- gradually replaces chain-of-thought text with continuous thought (<ct>) tokens across training stages
• Contrastive learning -- cross-device bidirectional contrastive loss for query/positive embedding alignment
• Latent MoE -- Mixture-of-Experts transition layer with 4 routed experts + shared expert in the latent reasoning loop
• Multi-modal evaluation -- MMEB image / video / visdoc evaluation with latent-MoE support

🏗️ Method

Overview of PLUME. The bottom panel illustrates the latent rollout process. The top-left panel expands the semantic-anchor-guided transition adapter with shared and specialized experts. The top-right panel shows the progressive explicit-to-latent curriculum.

🎞️ Results on MMEB-v2

All methods share the same Qwen2-VL-2B backbone.

Per-task performance comparison on MMEB-v2. PLUME consistently outperforms UME-R1 and single-pass baselines across most sub-tasks.

📦 Model & Data

🔧 Getting Started

• Environment Setup
• Training
• Evaluation

📁 Directory Structure

PLUME/
├── README.md
├── plume/                              # Python package
│   ├── train/                          #   Training
│   │   ├── train_plume.py              #     Core trainer (PlumeTrainer)
│   │   ├── train_plume_gc.py           #     Gradient-checkpointing variant
│   │   ├── latent_moe.py              #     Latent MoE transition module
│   │   └── argument.py                #     Dataclass argument definitions
│   ├── data/                           #   Data processing
│   │   ├── data_plume.py              #     Dataset, collator, curriculum sampler
│   │   └── rope2d.py                  #     2D/3D RoPE index computation
│   └── eval/                           #   Evaluation analysis tools
│       ├── compare_eval_results.py
│       ├── compute_mmeb_image_hit1_avg.py
│       ├── compute_mmeb_video_hit1_avg.py
│       └── analyze_max_tokens.py
├── configs/
│   ├── deepspeed/                      #   DeepSpeed configs (zero2/zero3/offload)
│   └── eval/                           #   MMEB dataset configs (image/video/visdoc)
├── scripts/                            #   Shell launchers
│   ├── train_singlenode.sh
│   ├── train_multinode.sh
│   ├── launch_multinode.sh
│   └── eval_plume_moe.sh
├── VLM2Vec/                            #   Bundled MMEB eval engine with eval_twomode.py
├── tools/
│   └── check_image.py
└── docs/                               #   Documentation
    ├── environment.md
    ├── training.md
    └── evaluation.md

🫡 Acknowledgements

Many thanks to the code bases from VLM2Vec and Qwen2-VL.

Citation

If you use this code for your research or project, please cite:

@misc{he2026plumelatentreasoningbased,
      title={PLUME: Latent Reasoning Based Universal Multimodal Embedding},
      author={Chenwei He and Xiangzhao Hao and Tianyu Yang and Yuxiang Ma and Yuheng Jia and Lingxiang Wu and Chaoyang Zhao and Haiyun Guo and Jinqiao Wang},
      year={2026},
      eprint={2604.02073},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.02073},
}

License

See the repository root for license information.