Neuron-Aware Active Few-Shot Learning for LLMs

NeuFS doesn't just apply active learning to LLMs; it relocates the acquisition signal from where we can easily measure it (model outputs) to where the actual learning happens (internal neuron activations). This is a methodological inversion that only becomes possible once we have tools to read neuron patterns reliably.

This work sits directly alongside the Neuron-OPSD paper from the same day, which also grounds data curation in neuron-level signals for self-distillation. Both papers reflect a maturing consensus that treating LLMs as black boxes wastes information. The broader context comes from the Understanding Large Language Models survey (July 1), which synthesized mechanistic findings on how transformers actually process information. NeuFS applies that mechanistic understanding to a concrete production problem: reducing annotation cost while maintaining domain performance. Unlike the clinical NLP piece from July 1, which found that learned gating rules fail at scale and forced practitioners back to static rules, NeuFS is betting that neuron signals are stable enough to guide sample selection reliably.

If NeuFS maintains its performance gains when tested on out-of-domain tasks (different from the domain it was fine-tuned on), that confirms neuron activation patterns capture generalizable knowledge gaps rather than task-specific artifacts. If performance degrades significantly on held-out domains, the method may only work within narrow task distributions, limiting its production applicability.