Can societies of LLM agents replicate classic findings in experimental
psychology? This release covers eleven paradigms. Nine are single-agent
(the "group" is described in the prompt text, matching how Asch's
confederates were scripted rather than sampled); two -- risky shift and
groupthink -- are the first paradigms run on a genuine multi-agent
deliberation harness, where multiple real model calls interact through
an actual shared transcript.
Single-agent… See the full description on the dataset page:
https://huggingface.co/datasets/joyboseroy/PsyAgentBench.
