Seeking arXiv endorsement (cs.LG)

Hello! I’m Dung, an independent researcher with a background in accessibility. I’m seeking your endorsement for cs.LG. My paper introduces a new metric tracking how attention heads bind multi-token terms during training. I find this signal precedes behavioral competence across Pythia checkpoints, revealing a binding-behavior decoupling at intermediate scales where internal structure and external capability diverge. This contributes a diagnostic for detecting latent knowledge before reliable performance emerges, with implications for monitoring when internal representations may suppress capabilities. I believe the paper is a strong fit for cs.LG since it focuses on mechanistic interpretability of training dynamics and emergence in language models. You can endorse me here: https://arxiv.org/auth/endorse?x=TYSM86 If the link doesn’t work, please visit arXiv endorsement page - Code: TYSM86 Thank you! Dung.