Researches reasoning, reinforcement learning and multi-agent systems, with earlier coauthored work on poker search and language-guided negotiation.
Why read Noam Brown?
Read Brown to understand why an AI system might benefit from searching or spending more time on a problem. His ReBeL paper, coauthored with three colleagues, combines learning and search in games with hidden information. Its guarantees depend on a specific two-player setting. That gives you a concrete example of what a search method establishes before applying the intuition to language models.
His team's CICERO paper shows another connection: a language model generates messages conditioned on a strategy, rather than choosing every action through free-form text alone. In his September 2026 interview, Brown also explains why large agent-count demonstrations need smaller comparison runs to isolate their benefit. Start with those answers, then the CICERO architecture to distinguish an impressive outcome from evidence about which component caused it.
Start with the original
Selected work
Interview transcript ·
Noam Brown on parallel agents and reasoning
Original interview with a written transcript. In the opening discussion, Brown explains the need for smaller agent-count comparisons. No mathematical background is needed for that passage.
Research paper ·
Human-level play in the game of Diplomacy by combining language models with strategic reasoning
Team research, with Brown as a corresponding author. Begin with the CICERO overview and Figure 1: the planner supplies intents, and the language model turns them into negotiable messages.
Research paper ·
Combining Deep Reinforcement Learning and Search for Imperfect-Information Games
Coauthored research introducing ReBeL. The opening game example explains why hidden information changes search. Section 9 confines the guarantees to two-player zero-sum games, rather than general language-model reasoning.
Projects & roles
Community rating
Sources & further reading
- His current biography identifies OpenAI, describes his research and links @polynoamial.
- Current X metadata names Noam Brown, identifies OpenAI and links noambrown.com.
- Read for the hidden-information example, method and restricted theoretical guarantees.
- The dated interview transcript separates Brown's answers from the host's framing.