RewardMap was jointly proposed by research teams from Westlake University, Tongji University, and other universities in October 2025. The relevant research results were published in the paper "...".RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning".

RewardMap is a multi-stage reinforcement learning (RL) framework designed to enhance the visual understanding and reasoning capabilities of multimodal large language models (MLLMs). The framework incorporates two key design features: First, it introduces a difficulty-aware reward design that includes detailed rewards, directly addressing the sparse reward problem while providing richer supervision. Second, the researchers propose a multi-stage reinforcement learning scheme that progressively transitions from simple perceptual tasks to complex reasoning tasks, offering a more effective cold-start strategy than traditional supervised fine-tuning (SFT).