Dynamic Routing Mixture-of-Experts Architecture MoEP Deep Dive: AI Models Adjust Compute Paths by Input Complexity for the First Time

Tsinghua University Institute of Artificial Intelligence and Zhipu AI published a joint research paper in late October on the MoEP (Mixture-of-Experts Pathways) dynamic routing architecture. For the first time, this architecture enables large language models to automatically adjust their compute pathways based on the complexity of input content.

Existing Mixture-of-Experts (MoE) models reduce inference costs through sparse activation, but their routing strategy is static. Whether the input is a simple factual query or a complex reasoning task, the model goes through the same routing decision process. MoEP changes this paradigm.

MoEPs core innovation is the introduction of a lightweight difficulty evaluator. Before tokens enter the main model, the evaluator determines the cognitive load level of the current input within 0.3 milliseconds and decides which expert layers to activate and the compute depth for each layer.

Zhipu AI chief scientist Tang Jie described a three-tier routing strategy in the paper. For simple factual queries and format conversion tasks, MoEP activates only 25% of expert layers, reducing inference energy consumption by 68%. For medium-complexity analysis and summarization tasks, it activates 50%. Only for multi-step reasoning, mathematical proofs, and code architecture design does it activate the full compute pathway.

On benchmarks including MMLU-Pro and HumanEval, MoEPs full-inference performance matches that of equivalently sized dense models, but average inference energy consumption is only 41% of the latter. In enterprise customer service and document processing scenarios, energy savings are even more pronounced since simple queries account for 80% of traffic.

Google DeepMind and Anthropic research teams commented on the MoEP paper, acknowledging dynamic routing as an important direction for LLM efficiency optimization while emphasizing the need for robustness validation across a wider range of task types.