Anastasios Angelopoulos

Anastasios Angelopoulos is a co-founder of Chatbot Arena, the AI evaluation platform he led with Wei-Lin Chiang. His work on its statistical methods helps turn users’ comparisons of model responses into rankings and examine how response length and formatting affect those preferences. He also helped separate Arena from the Berkeley student research group LMSYS and explained the subsequent move toward a company as a way to obtain resources beyond an academic setting.[3][5]

Angelopoulos joined Arena a couple of months into the project to work on its statistical aspects. In his November 2024 interview, he recalled that usage had declined after the initial launch and the project’s prospects were uncertain. Chiang described the earlier origins: experiments with an open chatbot led to a comparison interface, then anonymous model battles in which users chose the better response. Angelopoulos’s entry connected that growing evaluation project with his doctoral work in theoretical statistics and the foundations of model evaluation and testing.[3]

By November 2024, Arena was becoming independent of LMSYS. Angelopoulos explained that Arena had become its own project, while LMSYS creators Lianmin and Ying had moved toward SGLANG and other work. Separating the names would give those projects room to develop without LMSYS being treated as synonymous with Arena; Chiang added that they wanted to support new projects. Angelopoulos also invited contributions spanning statistical proofs, backend development, code execution and red teaming, promising credit to contributors.[3]

In May 2025, Angelopoulos and Chiang were recent Berkeley doctoral graduates and postdoctoral fellows, beginning to build a company out of Arena. Angelopoulos said further development required more resources than they could assemble academically. He committed to keeping public use free, supporting open research and working closely with Berkeley, and credited Ion Stoica and the Sky Computing Lab with helping the project get started.[5]

Angelopoulos advocated organic use: let people ask their own questions and offer free direct chat as well as battles. He argued that a team-written prompt set could not capture the diversity of people’s interactions, while acknowledging that the initial choice of user-generated prompts had been somewhat random. He attached a population limit to the resulting rankings: “We represent the people that come and vote.” He also rejected the idea that Arena was uniformly better than other benchmarks; it measured a different, useful kind of performance.[3][5]

In explaining style control, Angelopoulos described Arena scores as coefficients from a logistic regression relating model identity to human preference. The team could add the difference in response lengths and formatting features—such as Markdown headers and bulleted lists—to estimate their predictive effects alongside model identity. This gave readers a way to examine rankings adjusted for those features. He qualified the interpretation: separating their effects formally depends on conditions that are not always verifiable, and establishing causation requires more than the regression alone.[3]

In November 2024, making style control the default remained under consideration. Angelopoulos treated that as a substantive choice about which preferences should count: removing style effects would impose the team’s judgment on users who might value those features. As he put it, “That's kind of a community choice.”[3]

Angelopoulos also addressed the risk that a lab could privately test several variants and release only the highest-ranked one. In his November 2024 defense, he acknowledged that Arena’s confidence intervals were not adjusted for multiple comparisons and identified Bonferroni correction as a possible remedy. He said simulations suggested that selection bias from five variants was small relative to voter variability, and argued that continued fresh voting would reduce the initial winner’s curse. These were his reasons for the policy at that time.[3]

Discussing RouteLLM, whose framework and goals Chiang outlined, Angelopoulos argued that even simple routing could improve the performance–cost tradeoff: send harder questions to a larger model and easier ones to a smaller model, using signals such as code, mathematics or prompt length. He distinguished this from the harder task of estimating which model excels at each query type, where estimation errors could introduce enough variability to undermine the combined pipeline.[3]

In the May 2025 Berkeley interview, Angelopoulos framed AI safety and reliability as engineering responsibilities: deployed software should have its intended effect rather than give wrong answers or harmful recommendations. He described his perspective as apolitical and focused on useful tools rather than existential risks. He welcomed strong open models as beneficial to consumers and hoped AI would relieve repetitive paperwork. Asked about job redundancy, he emphasized increased individual productivity rather than predicting that positions would disappear.[5]

Before his full-time focus on Arena, Angelopoulos worked on statistical foundations that provide context for his role in model evaluation. He earned an undergraduate degree in electrical engineering at Stanford and a Berkeley EECS doctorate advised by Michael I. Jordan and Jitendra Malik. He subsequently worked with Ion Stoica as a Berkeley postdoctoral scholar.[4][3][5]

With Emmanuel Candes and Ryan J. Tibshirani, he co-authored Conformal PID Control for Time Series Prediction, published at NeurIPS 2023. The work combines conformal prediction and control theory to quantify uncertainty in online time-series predictions and adapt to systematic errors arising from seasonality, trends and distribution shifts. The authors reported improved coverage in experiments on four-week-ahead statewide U.S. COVID-19 death forecasts compared with the ensemble forecaster used in official CDC communications—a result about uncertainty coverage rather than a general claim of more accurate point forecasts.[6]