ProbMoE: Differentiable Probabilistic Routing for Mixture-of-Experts
arXiv:2606. 01509v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) models scale by activating only a small subset of experts per token.
The paper presents a statistical framework for Mixture-of-Experts (MoE) models, treating them as localized aggregation systems. It derives oracle risk bounds that separate approximation, expert‑learning, and router‑estimation errors for both dense and sparse routing with evolving experts. The authors also analyze how sparse Top‑K routing balances computational cost with performance, interpret gating geometrically, and explain how shared experts can capture common predictive structure while allowing routed experts to focus on local residuals.
arXiv:2606. 01509v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) models scale by activating only a small subset of experts per token.
GaP-MoE introduces a method for explicit expert specialization in Sparse Mixture-of-Experts models by manipulating expert output gradients instead of adding auxiliary loss terms. The technique selects a leader expert from the top‑K routed experts and projects follower experts’ gradients onto the orthogonal complement of the leader’s recent gradients, reducing directional overlap. Experiments on fine‑tuning BERT for GLUE and pretraining GPT‑2 on various corpora show that GaP‑MoE matches or surpasses standard SMoE performance while remaining a drop‑in module.
The paper introduces Multi-View Latent MoE, a sparse mixture‑of‑experts architecture that uses multiple independent routers to generate richer selection signals and to modulate expert computation. By aggregating sigmoid scores from these routers, the model selects experts that are strongly favored by one or several routers and mixes their relative scores into each expert’s input, thereby reusing routing preferences for both selection and computation. Experiments on a 24‑layer 10B‑A1B model trained on 708 B tokens show a 1.18‑point improvement in nine‑task downstream mean over the baseline, with the Multi‑Router submodule alone providing a 0.6‑point gain and significantly better alignment between router scores and expert utility. whyItMatters":"The approach demonstrates that leveraging multiple routers can both improve expert selection quality and maintain computational efficiency, leading to measurable downstream performance gains on large‑scale language models."
MASKerade is a new dense‑to‑MoE training method that learns experts as sparse subnetworks of a frozen pretrained feed‑forward network (FFN) using binary masks. A token‑level router selects which masked FFNs to execute, and both the router and mask scores are jointly optimized while the underlying FFN weights remain unchanged. In experiments on five vision‑language benchmarks with Qwen and Gemma backbones, a configuration of four 2:4 experts with top‑2 routing achieved the best performance among compared baselines, demonstrating that mask learning over frozen weights is a practical alternative for constructing token‑routed MoE experts.
arXiv:2608. 10392v1 Announce Type: new Abstract: Mixture-of-experts (MoE) models have recently moved beyond routing a fixed number of complete experts.
The paper introduces Elastic Expert Routing, a method that replaces the fixed top‑k expert selection in Sparse Mixture‑of‑Experts models with a stochastic sampling of the expert budget around k. This approach smooths the hard boundary between fully supervised and zero‑feedback experts, maintaining the same expected compute cost while improving training stability. Experiments show that elastic routing boosts downstream performance on supervised fine‑tuning and from‑scratch pretraining, achieving higher macro‑averages on OLMoE‑1B‑7B, Qwen3‑30B‑A3B, and outperforming static top‑k baselines by an average of 1.6 points.
MoRE: Mixture of Reused Experts is a hybrid architecture that combines Mixture-of-Experts (MoE) with weight‑sharing techniques. It shares expert pools across adjacent layers while each layer keeps its own router, and introduces lightweight depth embeddings to help shared experts differentiate layer contexts. Experiments on models ranging from 114 M to 1.15 B parameters show MoRE achieves lower perplexity and better downstream performance than standard MoEs and other weight‑sharing models, with only minimal changes to existing MoE implementations.
arXiv:2512. 13996v3 Announce Type: replace Abstract: Sparse Mixture-of-Experts architectures are essential for scaling model capacity efficiently, yet the standard Top-$k$ routing imposes a rigid sparsity pattern that ignores the intrinsic variance in token difficulty and layer-specific computational needs.
The paper investigates how routing decisions in sparse mixture-of-experts (MoE) models evolve across layers. By aligning router control subspaces with generalized orthogonal Procrustes analysis, the authors find that a single linear transition can predict routing states across depth with substantial accuracy, revealing a shared geometric structure. They further demonstrate that these canonical states preserve expert selection better than generic hidden representations and improve next‑step routing predictions, reducing negative log‑likelihood by up to 15.7% on OLMoE and 6.2% on Phi.
arXiv:2606. 17952v1 Announce Type: cross Abstract: Sparse Mixture-of-Experts (MoE) architectures enable scaling LLM parameters under a fixed inference budget by activating only a small subset of experts via top-$k$ routing.
The paper discusses how modern mixture-of-experts (MoE) models combine fine‑grained expert segmentation with shared experts, noting that shared experts can reduce routing exploration and limit the benefits of fine‑grained segmentation. This creates an exploration dilemma that hampers the full expressiveness of the model. The authors propose MirrorMoE, which adaptively suppresses unnecessary experts to shrink the exploration space, and report consistent performance gains across different model scales.
arXiv:2602. 06154v2 Announce Type: replace Abstract: Mixture-of-Experts (MoE) models scale large language models efficiently by sparsely activating experts, but once an expert is selected, it is executed fully.