Bertrand_AI

Description

This dataset contains the complete simulation records from a repeated Bertrand oligopoly experiment conducted with large language model (LLM) agents. The experiment replicates and extends the design of Lee and Hoffman (2025, Theory and Decision) by substituting LLM agents for human participants across six communication frequency conditions. Data collection procedure. All data were generated through controlled computational simulation conducted between March and April 2025. Three LLM agents — Claude Sonnet 4.6 (Anthropic), GPT-4o (OpenAI), and Gemini 2.5 Flash (Google) — were instantiated as autonomous pricing agents and subjected to the experimental protocol via each provider's official API. Two agent configurations were implemented: a heterogeneous configuration pairing one agent from each provider, and a homogeneous configuration using three independent sessions of Gemini 2.5 Flash. Each of the twelve treatment-configuration cells (6 treatments × 2 configurations) was replicated across 20 independent groups, yielding 240 groups, 7,200 group-round observations, and 21,600 agent-level observations in total. Dataset contents. Each observation records the group identifier, treatment condition, agent configuration, round number, individual submitted prices for all three sellers, the resulting market price, per-period profits, the full text of any chat messages exchanged (JSON-encoded), the randomized chat order, and a one-sentence contemporaneous pricing rationale generated by each agent alongside its price submission. The reasoning trace field is unique to the LLM setting and has no equivalent in the original human experiment. File structure. The dataset is provided as twelve CSV files, one per treatment-configuration cell, named results_{treatment}_{config}_all.csv. All sessions used fixed random seeds for chat order randomization to ensure reproducibility.

Institutions

• Pusan National University

  Busan, Busan

Categories

Licence