interpretabilityfactbullishPrototype language models remain within 2.5 percentage points on average downstream accuracy compared to dense baselines across 130M to 1.6B parametersMachine Learning (Statistics)27 Jul 2026http://arxiv.org/abs/2607.00510v1