Beyond Mediation: An Evolutionary Benchmark for Emotionally and Normatively Competent AI

An agent-based simulation framework measures three key traits for AI mediators: emotional regulation, willingness to compromise, and post-conflict trust repair. The benchmark compares five mediator policies across five canonical negotiation tasks using Monte Carlo episodes, with results revealing stable trade-offs between efficiency, fairness, and stability. Authored by Kohei Oshio and released under CC-BY-4.0, the dataset is a 162.0 KB PDF file last updated on 2026-04-23.

Use Cases

• Benchmarking AI mediator policies based on the described composite score that makes efficiency-equity-stability trade-offs explicit.
• Analyzing policy performance across heterogeneous negotiation tasks based on described variables like psychological reactivity and interaction-style asymmetry.
• Studying evolutionary policy selection in long-run simulations based on the described analysis of cross-task robustness.
• Evaluating the impact of transparent, weak interventions versus strong control on agreement probability and inequality of outcomes.

Strengths

• Defines five canonical negotiation tasks (S1, H1, H2, H3, C1) with varied parameters like psychological reactivity and escalation.
• Compares five distinct mediator policies under identical stochastic conditions using independent Monte Carlo episodes.
• Employs a composite score with pre-specified weights to evaluate trade-offs explicitly.
• Released under a permissive CC-BY-4.0 license, facilitating reuse and adaptation.

Limitations

• Dataset size is 162.0 KB, indicating a very limited scope focused on the framework description rather than large-scale simulation results.
• Row count and column-level documentation are absent; the primary data format is a PDF, which may require extraction of underlying data.
• The description notes this is an 'emotion-light v1 benchmark' with a baseline calibration that intentionally conserves emotional amplitude.

Provenance

Source

figshare, authored by Kohei Oshio.

Collection Method

Agent-based simulation framework with Monte Carlo episodes.

Freshness

Last updated 2026-04-23 05:26:37.