Infoblock alpha
cs-ml/interpretability IB-arxiv-2607.28319-neuron-zeroing-retention measured

In Llama-3.2-1B, zeroing at most 40 neurons (less than 0.031% of total MLP width) identified by Fairness Pruning retained a mean 99.49% of reasoning and general-knowledge benchmark capability.

Active Evidence state is recorded explicitly and may be revised by a later event.

Observed quantities

model_params_b 1 · neurons_zeroed_max 40 · pct_of_mlp_width 0.031 · mean_retention_pct 99.49

01

Primary source

  1. arxiv · 2607.28319 · v1 · abstract
Title
Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations
Authors
Pere Martra, Eugenio Martínez Cámara, Alfonso Ureña López
Version
v1
Check
Support quote found word for word in the arXiv abstract v1, 2026-10-08
Supporting quote from the source
zeroing at most 40 neurons in Llama-3.2-1B (less than 0.031% of total MLP width) achieves a mean retention of 99.49% in reasoning and general knowledge capabilities

02

Boundary of this block

No additional public limitation was recorded.

03

Connections

Outgoing

No outgoing public edges.

Incoming

No incoming public edges.