Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations
Pere Martra, Eugenio Martínez Cámara, Alfonso Ureña López
cs.CL, cs.CY, cs.LG
2026-07-30
Contrastive prompt pairs flag 40 demographic-sensitive neurons in Llama-3.2-1B (<0.03% of MLP width); zeroing them retains 99.5% of reasoning ability, but bias scores swing both ways.
LLMs inherit and sometimes amplify demographic stereotypes from pretraining data. The standard fixes (data filtering, fine-tuning, post-hoc processing) all act on weights indiscriminately: they cannot separate the weights that encode bias from the weights that carry capability, so they tend to dent performance before they dent bias. This paper steps back from proposing yet another debiasing trick and asks a prior question: can we locate the circuits responsible for demographic attributes cheaply, without retraining?
Fairness Pruning is a four-step, training-free pipeline:
Cost is a forward pass per contrastive pair.
Tested on Llama-3.2-1B, Llama-3.2-3B, and the Spanish Salamandra-2B. Capability barely moves:
| Metric | Base | Retention |
| MMLU | 32.0% | 101.34% |
| ARC-Challenge | 37.2% | 97.93% |
| HellaSwag EN | 64.2% | 99.78% |
| WikiText | 11.99 | 99.60% |
All 20 controls fall in 97.93%-101.34%, averaging 99.49% retention. Llama-1B zeroes at most 40 neurons out of 131,072 (16 layers x 8,192), about 0.03% of MLP width; Llama-3B zeroes 20 out of 229,376.
Bias does not move cleanly. In six of eight Llama-1B religion experiments, the ambiguous and disambiguated scores shift in opposite directions. On Llama-3B religion with 20 neurons zeroed, ambiguous bias drops from 6.00% to 1.50% (-4.5pp) while disambiguated bias flips from +3.84% to -1.67%. The paper calls this bidirectional bias destabilization.
The diagnostic value is clear and cheap: a handful of neurons can be tipped to shift demographic response without harming capability, which empirically separates the bias circuit from the capability circuit. That opens a path from blind zeroing toward directional modulation, tuning bias up or down rather than deleting it.
The catch is just as important. Zeroing does not reliably reduce bias. The reason is baked into BiasScore: because it is unsigned, the top-K set mixes neurons that push toward the stereotype with neurons that push against it. Zeroing pulls both, and the net effect depends on which sign dominates. The authors' counterintuitive read is that these neurons act as regulators of demographic expression, not repositories of stereotypes.
For practitioners: as a bias localization tool it works today at near-zero cost. As a one-switch debiaser it does not, and is missing a signed score.
The authors list hard ones:
One gap the paper leaves closed: since the net bias effect depends on the sign balance in the candidate set, the clean 99.49% capability retention may partly reflect bias shifts cancelling each other out, an intervention that did not really change where bias lands. Capability intact and bias addressed are not the same thing here, and the paper does not separate them.