The authors propose NeuFS, a Neuron-Aware Active Few-Shot Learning framework that shifts sample selection from output-level proxies to internal model dynamics. By utilizing neuron activation patterns, the method aims to identify specific knowledge gaps and reduce human annotation costs.
- NeuFS represents samples directly using neuron activation patterns rather than external embeddings or predictive entropy.
- It employs a dual-criteria strategy ensuring diversity through neuron patterns and prioritizing informative samples by quantifying neuron consensus.
- Experiments on three datasets show NeuFS outperforms existing Active Few-Shot Learning baselines in reasoning and text classification tasks.
- Ablation studies validate that internal neuron activations provide a more principled selection signal than external embeddings.
The framework offers a more effective approach to selecting few-shot demonstrations by leveraging the model's internal state to pinpoint challenging examples.