scalingfactbullishAt 340M parameters trained on 15B SlimPajama tokens, HOLA lowers perplexity on WikiTextArtificial Intelligence27 Jul 2026http://arxiv.org/abs/2607.02303v1