Collective Bias Mitigation via Model Routing and Collaboration

Collective Bias Mitigation via Model Routing and Collaboration. By Mingzhe Du and nine others, from Nanyang Technological University, the National University of Singapore, and two other institutions. Published on the second of October. No code repository is linked to it.

Large language models are now deployed in public health, in finance and in governance, settings that ask for accuracy and also for alignment with societal values. Despite recent advances, these models often perpetuate or amplify the bias embedded in their training data, which is a direct challenge to fairness. One existing response is self-debiasing: the model is encouraged to identify and correct its own biases. The limitation the authors point to is that this leans entirely on a single model's intrinsic knowledge. When a stereotype is deeply ingrained, the one model asked to catch it may be the very model that holds it.

Collective Bias Mitigation, or CBM, moves the job from one model to several. It is a framework that alleviates bias by learning fine-grained model behaviour and by fostering knowledge sharing among diverse LLMs. The authors describe it as the first work to systematically explore how distinct models should be selected, and how they should be organised, so that together they produce fairer responses.

Those two questions, selection and organisation, are what the title's routing and collaboration refer to. The organisation takes the form of topologies, and two are named: Debating and Committee. The abstract does not describe how either one runs, so the picture to hold is a group of different models arranged in a structure, rather than one model left to correct itself.

Experiments show CBM substantially outperforming standalone baselines. One example is given: in the top seven setting, the Committee topology lowers the age bias score from zero point two five to zero point one. Both Debating and Committee achieve substantial bias reduction, and of the two, Committee is the one that balances how effective the mitigation is against what it costs at inference time.

The argument rests on a simple asymmetry. A bias one model cannot see in itself may be visible to another model trained differently, so the fix is sought in a well chosen group rather than in more introspection.