Athena: Enhancing Multimodal Reasoning with Data-efficient Process Reward Models

Researchers introduce Athena-PRM, a multimodal process reward model that evaluates reasoning steps in complex problem-solving with remarkable data efficiency, requiring only 5,000 samples. The model leverages prediction consistency between weak and strong AI completers to generate high-quality training labels, achieving state-of-the-art results across multiple benchmarks including WeMath, MathVista, and VisualProcessBench.

Athena-PRM represents a significant advancement in how AI systems can be trained to evaluate intermediate reasoning steps, addressing a critical bottleneck in developing high-performance reward models. Traditional approaches require extensive step-level human annotations and computationally expensive Monte Carlo methods that produce noisy labels, making them costly and time-consuming. This research demonstrates that prediction consistency between weak and strong model completers serves as a reliable signal for identifying correct reasoning steps without heavy annotation burdens.

The breakthrough's efficiency gains matter substantially for the broader AI training landscape. As large language models become more sophisticated, the need for accurate process-level evaluation grows increasingly important for applications requiring complex reasoning—particularly in mathematics, science, and visual reasoning tasks. The ability to achieve strong results with only 5,000 samples, combined with strategic improvements like output reward model initialization and negative data up-sampling, suggests a path toward more accessible reward model development for research teams with limited resources.

For the AI development community, this work lowers barriers to entry for creating specialized reasoning evaluators. Organizations can now develop effective process reward models without proportionally massive computational and human annotation investments. The consistent performance improvements across diverse benchmarks indicate the approach generalizes well beyond specific domains. The development of Athena-7B using these reward models further validates the practical impact, showing measurable improvements across five benchmarks when deployed in actual reasoning scenarios.

• →Athena-PRM achieves state-of-the-art results using only 5,000 training samples through prediction consistency between weak and strong model completers.
• →The model improves multimodal reasoning performance by 10.2 points on WeMath and 7.1 points on MathVista benchmarks.
• →Process reward models can now be developed with significantly lower computational and annotation overhead, democratizing access to advanced reasoning evaluation.
• →Athena-PRM demonstrates robust capability in three distinct applications: test-time scaling verification, step correctness evaluation, and reward-based fine-tuning.
• →The research establishes new state-of-the-art results on VisualProcessBench with 3.9 F1-score improvement over previous leaders.