NeuFS uses neuron activation patterns for active few-shot learning

Researchers propose NeuFS, a Neuron-Aware Active Few-Shot Learning framework that shifts sample selection from output-level proxies to internal model dynamics. The method utilizes neuron activation patterns to represent samples and employs a dual-criteria strategy to ensure diversity and identify challenging examples prone to hallucination.

• Replaces predictive entropy and external embeddings with neuron activation patterns for sample representation.
• Uses neuron consensus to prioritize informative and challenging few-shot samples that LLMs tend to hallucinate.
• Ensures few-shot sample diversity through neuron patterns for broader example coverage.
• Outperforms existing AFSL baselines on reasoning and text classification tasks across three datasets.

Ablation studies validate that internal neuron activations provide a more principled selection signal than external embeddings, confirming the framework's effectiveness.