Papers I have written or contributed to, mostly on multi-agent systems, AI safety and language models. The full list is also on Google Scholar.
Using incomplete contract theory, proves that when contracts cannot anticipate every contingency, no realistic mechanism eliminates the welfare loss between self-interested agents. Prosocial agents, which weigh others’ welfare alongside their own, close the gap, as LLM-agent experiments on resource allocation and social dilemmas confirm.
A benchmark of 1,535 high-stakes multi-agent scenarios built on game-theoretic structures such as the Prisoner’s Dilemma, Stag Hunt and Chicken. Across 15 frontier models, agents fail to pick the socially beneficial action in 38% of cases; game-theoretic interventions improve outcomes by up to 18%.
Measures the complexity of a Theory-of-Mind task as the number of states needed to solve it, and uses it to audit five widely used ToM benchmarks. Introduces Discrete World Models (DWM), a prompting technique that describes how the environment changes as agents act, improving LLM performance on ToM tasks.