Will Brown

Will Brown is an AI researcher at Prime Intellect and the original creator of Verifiers, a library for creating environments to train and evaluate language models. His work includes reinforcement learning for agents that make successive tool calls, with particular attention to credit assignment: determining which intermediate decisions deserve reward, rather than judging only the final answer. He publishes research under the name William Brown and uses the handle willccbb.[1][5][8][7]

Brown developed Verifiers alongside research at Morgan Stanley. In his May 2025 interview, he recalled beginning work on the repository in earnest in January after a GRPO demonstration attracted attention. He described the multi-turn research project, undertaken with his intern, as one of his last major projects at Morgan Stanley. The repository extended the demonstration into reinforcement learning for multi-turn tool use.[1]

A concrete difficulty was getting models to use tools meaningfully. Brown reported that small models in the experiments sometimes answered directly instead of calling tools; malformed calls and parser errors could also derail their responses. Rewarding tool use introduced another failure: models learned to make the same dummy search and ignore its result. As he put it, “they're using it to get the reward.” The experiments therefore checked whether retrieved Wikipedia results contained information relevant to the ground-truth answer, using string matching to supply an intermediate evaluation.[1]

Brown coauthored Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Credit Assignment, submitted on May 17, 2025, with Siliang Zeng, Quan Wei, Oana Frunza, Yuriy Nevmyvaka, and Mingyi Hong. The collaborators modeled multi-turn tool use as a Markov decision process and proposed fine-grained turn-level advantage estimation. Brown explained the framing by treating each turn as an action and the tool response as the ensuing state. This allows intermediate evaluations to influence how credit is assigned across an interaction.[8][1]

The collaboration continued in an August 21, 2026 revision, with Brown among twelve authors and an expanded title: Reinforcing Multi-Turn Reasoning in LLM Agents via Fine-Grained Reward Structure and Credit Assignment. That version distinguishes terminal, delayed, and per-turn reward structures and derives corresponding GRPO and PPO algorithms. It broadens the framework beyond the original May 2025 version.[9][8]

Before his agentic reinforcement-learning work, Brown studied how learning systems interact with adaptive behavior. He completed a PhD in algorithmic game theory at Columbia, co-advised by Christos Papadimitriou and Tim Roughgarden; his research index lists his thesis, Learning in the Presence of Adaptive Behavior, in 2024. Earlier, he studied computer science, data science, and philosophy at Penn. He subsequently worked on LLM projects in Morgan Stanley’s Machine Learning Research group before joining Prime Intellect.[6][7][1]

Brown’s account of his theoretical research supplies context for his later work on interacting agents. In May 2025, he recalled expecting multi-agent systems to learn continually in parallel and finding difficult open questions in their mathematical foundations. He described a later decision in direct terms: “And at some point I kind of got tired of proving theorems and was like, okay, let's just go build the thing.” He also recalled betting, between the releases of o1 and R1, that reinforcement learning would intersect with tool-using agents.[1]

In May 2025, Brown argued that reasoning models were a step toward practical agents that could act through tools. For coding agents, he wanted models to complete the requested change without adding unnecessary files, helper functions, or comments. He also supported exposing thinking-budget controls to developers as a way to balance quality, cost, and latency: “You do need a knob to control costs and also latency.”[1]

Brown favored exploring model-based rewards beyond deterministic answer checking. He pointed to the difficulty of writing parsers that recognize equivalent mathematical expressions across LaTeX, code, and prose. An LLM judge, he suggested, could evaluate more flexible criteria—including whether an intermediate search query was useful—and supply rewards at the turn level. He presented this as a promising research direction rather than a demonstrated guarantee of reliable judging.[1]

Brown also argued that academia could be a strong source of model evaluations. His advice to PhD students was to turn impressions of model quality into precise scientific questions: clever evaluations could depend more on intellectual effort than on the capital needed to pretrain foundation models. He saw academic incentives to produce interesting findings about the broader field as useful for evaluation research.[1]