Abstract
The present subject matter relates to a systematic study of emergent bias and fairness dynamics in multi-agent LLMdecision making. In particular, the present disclosure compares single-agent performance with memory-based and collective refinement multi-agent discussion paradigms across a suite of group fairness metrics on tabular datasets with sensitive attributes. Across settings, fairness changes from multi-agent collaboration relative to single-agent baselines are typically concentrated around small negative values or near zero, indicating that bias is most often only marginally reduced or remains unchanged, while a thin but long positive tail shows rare cases where multi-agent interaction can substantially worsen bias. These findings motivate systematic investigation into the conditions under which multi-agent collaboration mechanisms mitigate or amplify bias.
Publication Date
2026-01-05
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Madigan, Maeve Una; Burrell, Stuart; Parameswaran, Kamalaruban; Moynihan, Glenn; and Sutton, David, "Emergent Bias and Fairness in Multi-Agent Decision Systems", Technical Disclosure Commons, (January 05, 2026)
https://www.tdcommons.org/dpubs_series/9127