ReasonMap-Plus Dataset

This repository hosts the ReasonMap-Plus dataset, an extended dataset introduced in the paper RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning.

Paper Abstract

Fine-grained visual reasoning remains a core challenge for multimodal large language models (MLLMs). The recently introduced ReasonMap highlights this gap by showing that even advanced MLLMs struggle with spatial reasoning in structured and information-rich settings such as transit maps, a task of clear practical and scientific importance. However, standard reinforcement learning (RL) on such tasks is impeded by sparse rewards and unstable optimization. To address this, we first construct ReasonMap-Plus, an extended dataset that introduces dense reward signals through Visual Question Answering (VQA) tasks, enabling effective cold-start training of fine-grained visual understanding skills. Next, we propose RewardMap, a multi-stage RL framework designed to improve both visual understanding and reasoning capabilities of MLLMs. RewardMap incorporates two key designs. First, we introduce a difficulty-aware reward design that incorporates detail rewards, directly tackling the sparse rewards while providing richer supervision. Second, we propose a multi-stage RL scheme that bootstraps training from simple perception to complex reasoning tasks, offering a more effective cold-start strategy than conventional Supervised Fine-Tuning (SFT). Experiments on ReasonMap and ReasonMap-Plus demonstrate that each component of RewardMap contributes to consistent performance gains, while their combination yields the best results. Moreover, models trained with RewardMap achieve an average improvement of 3.47% across 6 benchmarks spanning spatial reasoning, fine-grained visual reasoning, and general tasks beyond transit maps, underscoring enhanced visual understanding and reasoning capabilities.

Dataset Overview

ReasonMap-Plus addresses the core challenge of fine-grained visual reasoning for multimodal large language models (MLLMs). It extends the original ReasonMap dataset by introducing dense reward signals through Visual Question Answering (VQA) tasks, enabling effective cold-start training of fine-grained visual understanding skills. This dataset is crucial for the RewardMap framework, which aims to improve both visual understanding and reasoning capabilities of MLLMs in structured and information-rich settings like transit maps.

The dataset includes ReasonMap-Plus for evaluation and ReasonMap-Train for RewardMap training.

Links

• Project Page: https://fscdc.github.io/RewardMap
• Code Repository: https://github.com/fscdc/RewardMap

Sample Usage

To get started with the RewardMap project and utilize the ReasonMap-Plus dataset, follow the steps below.

1. Install dependencies

If you face any issues with the installation, please feel free to open an issue. We will try our best to help you.

pip install -r requirements.txt

2. Download the dataset

You can download ReasonMap-Plus for evaluation and ReasonMap-Train for RewardMap Training from HuggingFace or by running the following command:

python utils/download_dataset.py

Then, put the data under the folder data.

3. Data Format Example

The data will be transferred into a format like:

  {
    "conversations": [
      {
        "from": "human",
        "value": "<image> Please solve the multiple choice problem and put your answer (one of ABCD) in one \"\\boxed{}\". According to the subway map, how many intermediate stops are there between Danube Station and lbn Battuta Station (except for this two stops)? \
A) 8 \
B) 1 \
C) 25 \
D) 12 \
"
      },
      {
        "from": "gpt",
        "value": "B"
      }
    ],
    "images": [
      "./maps/united_arab_emirates/dubai.png"
    ]
  },

Citation

If you find this paper useful in your research, please consider citing our paper:

@article{feng2025rewardmap,
  title={RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning},
  author={Feng, Sicheng and Tuo, Kaiwen and Wang, Song and Kong, Lingdong and Zhu, Jianke and Wang, Huan},
  journal={arXiv preprint arXiv:2510.02240},
  year={2025}
}