Will Brown leads applied research at Prime Intellect and created Verifiers, an open-source framework for training and evaluating language-model agents. He builds tools that let engineers improve models through realistic tasks, feedback, and repeated interaction instead of relying exclusively on general-purpose models.
Brown studied computer science, data science, and philosophy at the University of Pennsylvania and completed a PhD in algorithmic game theory at Columbia University, advised by Christos Papadimitriou and Tim Roughgarden. His academic research examined learning amid adaptive behavior, including multiplayer stochastic games, recommendation systems, and differential privacy.
He held research or engineering roles at AWS, Two Sigma, MongoDB, and AmFam before joining Morgan Stanley’s Machine Learning Research group. His independent projects include mlx_parallm, which accelerates parallel language-model inference with Apple’s MLX framework, and mcp-client-server, which combines an MCP client and server to simplify tool development.
At Morgan Stanley, Brown experimented with training a small language model on mathematics problems, rewarding both correct answers and useful response structure. That compact experiment grew into Verifiers, which he continued developing at Prime Intellect. He also contributed to the team behind INTELLECT-3, an openly documented model trained using Verifiers, the Environments Hub, and asynchronous reinforcement-learning infrastructure.
- Verifiers and reusable agent environments. Brown treats an environment as a combination of tasks, an agent harness, tools, and scoring rules. One environment can support evaluation, synthetic-data generation, supervised fine-tuning, distillation, and reinforcement learning. A Wordle-playing agent, for example, can earn rewards for solving the puzzle and using fewer guesses.
- Reasoning and agency as one learning problem. Mathematical reasoning and tool-using agents both improve by exploring possible trajectories and receiving feedback; their environments differ chiefly in complexity. Brown’s work on agentic reasoners translates reinforcement learning into familiar engineering components: tasks, harnesses, evaluations, and repeated interaction.
- Rubric engineering and reward hacking. Open-ended tasks rarely provide clean answer keys. Brown develops supervision from production traces, completed code changes, simulated tools, model-based judges, and detailed rubrics while testing whether agents can exploit scoring shortcuts. His approach to unverifiable rewards also emphasizes calibrating tasks so they are difficult enough to teach useful behavior without becoming impossible.
- Asynchronous reinforcement learning at scale. The team-built PRIME-RL framework separates rollout orchestration, inference, and training so long-running agent tasks do not leave expensive compute waiting. The Environments Hub makes those environments reusable across evaluation and post-training workflows.
Brown’s longer-term goal is continual learning from real-world feedback: turning an agent’s production successes and failures into new training environments while keeping human experts responsible for defining what improvement means.
