In a nutshell: Fairness Pruning localizes demographic bias in GLU-MLP layers by analyzing differential neuron activation and removes it with minimal capability loss.
Researchers present a method for targeted identification and manipulation of neurons in language models that process demographic bias. In models of the 3-billion-parameter class, such biases can be reduced by deactivating fewer than 40 neurons without impairing domain expertise.
Fairness Pruning is a structural intervention to localize and reduce demographic bias in large language models. The method uses minimally contrastive prompt pairs and captures activations during inference to identify neurons that respond differentially when processing demographic attributes. The analysis focuses on the down_proj input in GLU architectures. The empirical evaluation covered models up to 3 billion parameters, including the Llama-3.2 family and Salamandra-2B, combined with standard benchmarks and qualitative text generation tests.
The central finding is: deactivating identified neurons systematically changes how the model responds to demographic variables. The effect is highly surgical: in Llama-3.2-1B, a maximum of 40 neurons—less than 0.031% of MLP width—suffice to achieve bias reduction. At the same time, the models retain 99.49% of their performance in reasoning and general knowledge. However, the intervention does not result in simple bias reduction, but rather in bidirectional destabilization: because the BiasScore is unsigned, the identified neuron sets contain both those that amplify stereotypes and those that counteract them. The net effect depends on which influence predominates.
These results empirically demonstrate that demographic bias processing and the model’s cognitive capabilities operate on separate circuits. This creates methodological foundations for the transition from blind neuron deletion to directed behavior modulation—a path toward more precise bias mitigation strategies without capability loss.
Source: arxiv.org · Published July 29, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.