MARS / README.md
nielsr's picture
nielsr HF Staff
Add dataset card, task category, and paper/code links
2e592b5 verified
|
Raw
History Blame
1.94 kB
metadata
license: apache-2.0
task_categories:
  - image-text-to-text
tags:
  - visual-reasoning
  - reinforcement-learning
  - multi-source
  - rlvr

MARS: Mono-Anchored Multi-Source Visual Reasoning

This repository contains the dataset for the paper "Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning".

Introduction

MARS is a novel mono-anchored multi-source reasoning framework designed for Reinforcement Learning with Verifiable Rewards (RLVR). It treats each visual modality as an independent information source and uses mono-source rewards as dynamic anchors to explicitly incorporate information gain from multi-source fusion while suppressing potential noise or conflicts.

Dataset Summary

The dataset provides reinforcement fine-tuning (RL) data for several multi-source visual reasoning domains:

Domain Task Type
Infrared Grounding tasks (RGB & IR)
Depth Spatial reasoning using depth information
Multi-view Reasoning across multiple perspectives
Text-rich Visual reasoning in document-intensive environments

For instructions on how to download the original images and prepare the data for training using the MARS framework, please refer to the official GitHub repository.

Citation

If you find this work useful, please cite our paper:

@article{zeng2026does,
  title={Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning},
  author={Zeng, Fanhu and Luo, Zhicong and Wang, Zefan and Li, You and Chen, Chi and Sun, Maosong},
  journal={arXiv preprint arXiv:2605.25437},
  year={2026}
}