Multigroup Social Debiasing of LLMs via Neuron Discovery

Mitigating social biases in language models is a prerequisite for their responsible deployment. Many benchmarks and debiasing methods operationalize bias as a difference between two groups, a target group and a single alternative, so a model can reduce this pairwise gap while its preferences over the remaining groups stay biased. We define social bias as the divergence from uniform of the distribution a model induces over a set of demographic groups, with the completion held fixed so that any preference is attributable to the group. To this end, we introduce $\text{Multigroup-StereoSet}$, an extension of $\text{StereoSet}$ in which each target group is complemented by distractor groups selected for maximal dispersion over real-world taxonomies, while stereotypical and anti-stereotypical completions are held fixed. We then propose DAID, a two-stage method that first identifies the MLP neurons whose suppression moves the group distribution toward uniform and then fine-tunes only these neurons' parameters. Across four models on English text, DAID reduces the median $\max /\min$ group-probability ratio from $23$-$35$ to $1.3$-$2.3$ on *nationality* and from $97$-$138$ to below $1.5$ on *occupation*, and increases $\text{StereoSet}$ ICAT by $6$–$16$ points, at a cost of $1$-$5.4$ points on $\text{MMLU}$. Model-editing and activation-intervention baselines raise ICAT by up to $15$ points but leave the ratio on *nationality* between $14$ and $121$. We find that suppressing the identified neurons without training also yields near-uniform distributions but degrades language modelling capabilities.