Research

A scalable LLM-agent simulator for X and Reddit that pushes agent counts to one million, replicates real-world information spreading, and reveals that herd effects require more than 100 agents to emerge.

A sign-restricted SVAR across 14 currency pairs shows that delayed overshooting in response to US monetary shocks is concentrated in the Volcker era (1979–87) and largely absent from post-1987 data.

The NeurIPS 2023 paper that introduced inception prompting and a two-agent role-playing protocol, generating the first large-scale LLM cooperative conversation datasets.

A controlled study of when and why LLM-agent groups fail to reach agreement in a Byzantine consensus game.

A Tsinghua FIB Lab platform that drives 10,000 LLM agents through a real urban, economic, and social environment and reproduces four field social experiments on a single stack.

A Stanford team injects two-hour qualitative interviews of 1,052 real individuals into an LLM and predicts their attitudes and behaviors at the individual level.

A Markovian classification of DSGE models that puts Partially Observable MDPs at the generalization of linear rational expectations, and derives an equilibrium as a fixed point mapping observables to beliefs.

A close reading of the 2023 UIST paper that fused memory streams, reflection, and recursive planning into an LLM loop, with the numbers and failure modes kept intact.