Collective Bias Mitigation via Model Routing and Collaboration
The paper proposes routing and coordinating multiple LLMs to reduce bias more than a single model can on its own.
Collective Bias Mitigation learns fine-grained model behavior, then uses collaboration among diverse models to produce fairer responses. The authors say this is the first systematic study of how to select and organize distinct LLMs for that purpose. In experiments, CBM outperformed standalone baselines; in one top-7 Committee setting, the age-bias score fell from 0.25 to 0.10. The Debating and Committee setups drove the strongest reductions, with Committee presented as the better balance of mitigation and inference cost. HF Daily Papers' note
Collective Bias Mitigation learns fine-grained model behavior, then uses collaboration among diverse models to produce fairer responses. The authors say this is the first systematic study of how to select and organize distinct LLMs for that purpose. In experiments, CBM outperformed standalone baselines; in one top-7 Committee setting, the age-bias score fell from 0.25 to 0.10. The Debating and Committee setups drove the strongest reductions, with Committee presented as the better balance of mitigation and inference cost. HF Daily Papers' note
score 4