MNEMO

MNEMO is the model introduced in the ICML 2026 paper Interactive Person Retrieval via Multi-Turn Multimodal Conversation. It is designed for multimodal interactive person retrieval, where retrieval results are progressively refined through multi-turn conversations and visual feedback.

MNEMO uses InternVL2.5-1B as its backbone. It represents each dialogue turn as an atomic multimodal unit and aggregates dialogue memory to model fine-grained dependencies across turns.

Model Overview

Checkpoints

byougert/MNEMO
├── model.safetensors      # Final model finetuned on MInterPEDES
├── config.json
├── tokenizer files
└── pretrain/
    ├── model.safetensors  # Model pretrained on MALS before finetuning
    ├── config.json
    └── tokenizer files

The files in the repository root correspond to the final MNEMO model finetuned on MInterPEDES. The pretrain/ subfolder contains the checkpoint pretrained on MALS before finetuning.

Resources

• Dataset: MInterPEDES
• Code

Download

Download the complete repository with:

hf download byougert/MNEMO --local-dir MNEMO

The root checkpoint is the recommended model for evaluation and inference. Please refer to the official code repository for model loading and evaluation instructions.

Citation

@inproceedings{bai2026interactive,
  title={Interactive Person Retrieval via Multi-Turn Multimodal Conversation},
  author={Bai, Yang and Wang, Tingfeng and Yang, Bin and Cao, Min and Wang, Jinqiao and Ye, Mang},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026}
}