Make MoE experts from different models talk to each other: a learned translator reads one frozen MoE's expert activations and feeds them to another. Toy-scale experiments, leak-audited.
research pytorch moe interpretability mixture-of-experts llm cross-attention model-merging representation-alignment model-stitching
-
Updated
Nov 4, 2026 - Python