agentsfactbullishNeuron activation patterns can effectively represent samples and identify valuable few-shot demonstrations for LLM adaptation to specialized domainsMachine Learning27 Jul 2026http://arxiv.org/abs/2607.02423v1