Model research
NemoH Routing Research
Representation-engineering experiments investigating how activation-level interventions influence expert routing in a hybrid Mamba/MoE model.
The question
Can representation-engineering control vectors change routing behavior in a large MoE model, and which layers provide the most leverage?
What I investigated
I adapted a research workflow around a 52-layer hybrid Mamba/MoE architecture with 128 experts, conducting offline routing analyses and a bounded pilot that compared activation-level intervention with LoRA-style adaptation.
What the briefing will show
A future figure will state exactly what moved, compared with which baseline, and under what pilot conditions. The project will keep observed effects, interpretations, and GPU-gated next experiments separate.
Important distinction
This is mechanistic routing research, not the separate NVIDIA Nemotron Model Reasoning Challenge. The two projects have related naming but different methods and evidence.