Instruction data selection for fine-tuning large language models often relies on model-internal signals such as neuron activations and gradients. A new preprint on arXiv asks a more fundamental question: how should neuron relevance be characterized in the first place?

The paper's title contrasts two possibilities: "support" (presumably how consistently a neuron is engaged) and "magnitude" (presumably the strength of its activation). The abstract notes that this distinction remains underexplored, and that the field lacks a clear answer to whether selection should prioritize one over the other.

The available abstract stops before presenting results, so this summary only covers the framing. Readers should treat any reported findings as preliminary until the full paper is available.