A research method called EPnG can fine-tune large mixture-of-experts AI models by updating less than 1% of their parameters.
Mixture-of-experts (MoE) models route different inputs to different specialist sub-networks, which makes them efficient to run but awkward to adapt. Standard fine-tuning tools like LoRA treat all experts equally, wasting compute on rarely-used ones. EPnG takes a different approach: it reads the router's own probability scores to figure out which experts are pulling weight, prunes the idle ones, and redistributes that saved capacity to the busy ones using a technique called rank growth with orthogonal initialization. The total parameter budget stays fixed throughout. Tested on OLMoE and Qwen1.5-MoE, EPnG matched full fine-tuning performance while touching only 0.55% to 0.72% of parameters — 140x to 180x fewer than a full update.
This matters because MoE architectures are increasingly the default for frontier models — Mistral, Google, and others have shipped MoE-based systems — yet the tooling for adapting them has lagged behind. If EPnG's gains hold outside the paper's benchmarks, organizations that can't afford to fully retrain a large MoE model gain a credible path to customization. The router-aware approach also surfaces a broader insight: fine-tuning tools designed for dense models are a poor fit for sparse architectures.
The results are compelling on paper, but academic fine-tuning benchmarks rarely capture the messiness of real deployment data — so treat the 140x-180x figure as a ceiling until someone stress-tests it in production.