What is MM-Eureka
MM-Eureka is a multimodal reasoning model jointly developed by researchers from Shanghai Artificial Intelligence Laboratory, Shanghai Innovation Institute, Shanghai Jiao Tong University, and The University of Hong Kong. Through rule-based large-scale reinforcement learning (RL), the model extends key characteristics from unimodal reasoning (such as stable response length growth, accuracy rewards, and visual aha moments) to multimodal scenarios.
MM-Eureka introduces two core models: MM-Eureka-8B and MM-Eureka-Zero-38B, based on InternVL2.5-Instruct-8B and InternVL2.5-Pretrained-38B, respectively. Using only 54K image-text data for rule-based reinforcement learning training, its average performance surpasses the MPO model trained with 1M data. MM-Eureka-Zero-38B, using only 8K image-text mathematical reasoning data, surpasses the instruct model by 8.2% on the self-built K12 benchmark and performs comparably on MathVerse.
Main Features of MM-Eureka
- Multimodal reasoning capability: Extends large-scale rule-based reinforcement learning (RL) to the field of multimodal reasoning, capable of processing text and visual information.
- Reproduction of key characteristics: Reproduces key characteristics of text RL systems (such as DeepSeek-R1) in the multimodal space, including accuracy rewards and steady improvement in response length, as well as the emergence of reflective behavior.
- Data efficiency: Using only 54K image-text data for rule-based RL training, its average performance surpasses the MPO model trained with 1M data, and its overall benchmark accuracy is comparable to models trained with 12M data for CoT SFT.
Technical Principles of MM-Eureka
- Rule-based large-scale reinforcement learning framework: MM-Eureka developed an efficient and scalable multimodal large-scale reinforcement learning framework based on OpenRLHF, supporting multiple models such as InternVL and RL algorithms. It enables the model to train effectively in multimodal environments, successfully reproducing key characteristics of DeepSeek-R1, such as accuracy rewards and steady improvement in response length.
- Data filtering and stable training: The research team found that data selection is crucial for stable RL training. Difficulty-based data filtering strategies played a key role in the stability of RL training.
- Visual aha-moment: MM-Eureka exhibits visual aha moments similar to DeepSeek-R1 during training. Specifically, the model learns to reflect and backtrack, re-examining key information in the image.
- Minimalist reinforcement learning design: The minimalist RL design has proven effective in MM-Eureka. When experimenting on the instruct model, adding KL divergence restricts model exploration, resulting in no observable increase in response length. MM-Eureka adopts a simple reward function (such as accuracy reward and format reward), and achieves stable training through a difficulty-based data filtering strategy.
- Efficient data utilization: MM-Eureka demonstrates extremely high data efficiency. Using only 54K image-text data for rule-based reinforcement learning training, its average performance surpasses the MPO model trained with 1M data. MM-Eureka-Zero, using only 8K image-text mathematical reasoning data (0.05% of the instruct model), surpasses the instruct model by 8.2% on the self-built K12 benchmark and performs comparably on MathVerse. This indicates that in the field of multimodal reasoning, a simple rule-based reinforcement learning design can significantly improve training effectiveness, achieving performance comparable to large-scale training even with less data.
MM-Eureka Project Address
- Github repository：https://github.com/ModalMinds/MM-EUREKA
- arXiv Technical Paper：https://arxiv.org/pdf/2503.07365
Application Scenarios of MM-Eureka
- Education: Through its powerful reasoning capabilities and reflection mechanism, MM-Eureka can help students better understand and solve complex mathematical problems.
- Augmented Reality (AR) and Virtual Reality (VR): In AR and VR scenarios, by combining visual and linguistic information, MM-Eureka can provide users with a more immersive and interactive experience.
- Data Analysis and Decision Support: MM-Eureka's multimodal reasoning capability gives it a significant advantage in data analysis and decision support. It can handle complex image-text data, helping users extract key information from large amounts of information and make wiser decisions.
- Automation and Intelligent Assistants: MM-Eureka can serve as the core technology of intelligent assistants, providing users with a more intelligent and natural interaction experience.
- Games and Entertainment: In the fields of games and entertainment, MM-Eureka's multimodal reasoning capability can be used to develop more intelligent non-player characters (NPCs) and interactive storylines.