Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI

Designing environments for AI agents—rather than rigid workflows—unlocks superior creativity and scientific problem-solving capabilities. By providing infrastructure, incentives, and guardrails, environments allow agents to collaborate and compete effectively. The Einstein Arena demonstrates this by enabling agents to discover breakthrough solutions to long-standing mathematical challenges, such as the 11-dimensional kissing number problem, while also optimizing production-level machine learning kernels. Similarly, the Data Science Gym (DSGym) provides a robust, shortcut-free platform for training and evaluating data science agents across diverse scientific domains. These environments move beyond simple prompt-based instructions, fostering collective intelligence that surpasses the performance of individual models. This paradigm shift from task-specific workflows to agent-native environments represents a critical evolution in building powerful, autonomous AI systems capable of genuine scientific discovery.