EMO: Pretraining mixture of experts for emergent modularity

The key distinction EMO makes is between imposed modularity (where routing is explicitly designed) and emergent modularity (where specialization arises from training dynamics). Most MoE implementations to date belong to the first category, so the claim here is that structure can be discovered rather than prescribed, which has different implications for interpretability.

This connects directly to the HyCOP paper covered in early May, which took a similar modularity-first stance in scientific ML by replacing monolithic mappings with composable operators. Both works are converging on the same architectural intuition from different directions: that decomposability during training, not just at inference, produces more robust and inspectable systems. The MIT scaling study from May 3rd is also relevant background, since superposition as a mechanistic driver of scaling is precisely what emergent modularity tries to counteract by encouraging cleaner separation of representations across experts.

Watch whether Hugging Face releases downstream fine-tuning benchmarks showing that EMO-pretrained models require fewer examples to adapt to new tasks than standard MoE baselines. That would be the concrete signal that emergent modularity is doing real work, not just producing a tidier routing diagram.