6arXiv cs.CL (Computation and Language)·
MAESTRO: Markov-chain-based structured pruning framework for Mixture-of-Experts language models
Researchers introduce MAESTRO, a structured pruning method for sparse Mixture-of-Experts (MoE) language models that models expert activation trajectories as Ergodic Markov chains to capture cross-layer routing dependencies. Unlike existing pruning methods adapted from dense transformers, MAESTRO uses stationary distributions of these chains as globally-aware importance heuristics. Under a 50% compression regime evaluated across five domains including safety and ethics, MAESTRO outperforms state-of-the-art baselines by up to 10.61% in average performance retention with lower cross-task variance.
Related guides (2)
Related events (8)
Mixture of Experts Explained
This Hugging Face blog post provides a technical overview of the Mixture of Experts (MoE) architecture, explaining how sparse gating mechanisms route tokens to subsets of expert feed-forward layers to achieve computational efficiency. The post covers training dynamics, inference considerations, and the tradeoffs between dense and sparse models. It serves as a reference document contextualizing MoE's growing relevance following high-profile model releases using the architecture.
EMO: Pretraining Mixture of Experts for Emergent Modularity
AllenAI introduces EMO, a pretraining approach for Mixture of Experts (MoE) models that aims to produce emergent modularity during training. The work explores how MoE architectures can develop specialized expert routing without explicit supervision. Published on the Hugging Face blog, this represents research-level work on improving MoE training dynamics and efficiency.
Causal audit finds routing statistics do not predict expert importance in MoE pruning
A new arXiv paper conducts a token-level interventional audit of Mixture-of-Experts (MoE) pruning heuristics across three architectures (OLMoE-1B-7B, Qwen1.5-MoE, DeepSeek-V2-Lite), finding that no standard observational metric — utilization rates, activation norms, routing weight distributions — reliably predicts which experts can be removed without functional cost. Effect sizes fall below Cohen's d = 0.17 across all 60 metric-layer combinations after multiple-comparison correction, with only a single significant signal at OLMoE's final layer. The authors argue that existing pruning methods succeed not because they identify dispensable experts but because early-layer redundancy makes most selection criteria interchangeable. The work frames this as a concrete counterexample to the broader interpretability practice of treating associational (rung-1) evidence as interventional (rung-2) conclusions.
Expert Tying reduces MoE LLM memory footprint by ~2x with minimal quality loss
Researchers introduce Expert Tying, an architectural modification for Mixture-of-Experts LLMs that shares expert parameters across consecutive transformer layers while keeping routing and attention layer-independent. Evaluated on OLMoE, Qwen3, and DeepSeek-style MoE architectures, the method achieves nearly 2x memory reduction with negligible perplexity or downstream quality degradation. The approach exploits parameter redundancy in MoE pathways to improve the compute-to-memory trade-off for training and inference.
ZEDA: Post-Trained MoE Models Can Skip Half Their Experts via Self-Distillation
This paper introduces Zero-Expert Self-Distillation Adaptation (ZEDA), a framework that converts static post-trained Mixture-of-Experts (MoE) language models into dynamic ones without pre-training from scratch. ZEDA injects parameter-free zero-output experts into each MoE layer and uses two-stage self-distillation with the original model as a frozen teacher. Applied to Qwen3-30B-A3B and GLM-4.7-Flash across 11 benchmarks, ZEDA eliminates over 50% of expert FLOPs with marginal accuracy loss and achieves approximately 1.20× end-to-end inference speedup, outperforming the strongest dynamic MoE baseline by 4–6 points.
Complete-muE: Optimal Hyperparameter Transfer and Scaling for MoE Models
Complete-muE is a framework for transferring hyperparameters across dense FFN and Mixture-of-Experts (MoE) transformer architectures, addressing limitations of existing tools like μP and SDE that cannot handle simultaneous architecture and token-per-expert changes. It uses a two-bridge system: Bridge I maps dense FFN to Dense MoE via active-width μP with normalized router scale, and Bridge II maps Dense MoE to sparse MoE via activated-expert scaling with a first-order SDE correction. The practical outcome is a 'tune dense once, transfer to all' recipe that enables near-optimal hyperparameter reuse across MoE configurations without costly re-tuning. Experiments on language model and diffusion model pretraining confirm stable hyperparameter optima across architectures and parameter counts.
Expert-aware causal tracing of factual recall in sparse MoE language models
A new arXiv preprint extends causal tracing methodology to sparse mixture-of-experts (MoE) language models, asking which routed experts mediate factual recall rather than just which layers or feed-forward modules. Using CounterFact facts, the authors apply noise-corruption and clean-patch interventions to Qwen3-30B-A3B-Base and Mixtral-8x7B-v0.1, finding that expert-level localization is possible in the former (a single expert at layer 44) but requires multi-expert coalition recovery in the latter. The results indicate that factual localization in MoE models is model- and protocol-dependent rather than universal.
CHERRY: Compressed Hierarchical Experts with Recurrent Representational Yield — three compute-efficient LM training techniques
A preprint from arXiv introduces CHERRY, a suite of three complementary techniques for compute-efficient language model training: Selective Ground Truth Token Training (SGT) that concentrates supervision on ~15% of semantically loaded tokens while recovering ~67% of full-sequence loss reduction; depth compression that shrinks a 48-layer 1B-parameter model to 6 layers (227M) via layer averaging and recurrent unrolling, matching a 566M dense model's loss; and a Mixture of Efficient Experts (MoEE) assembly that outperforms individual compressed models at comparable active parameters. The techniques are validated on CHERRY-1.8B, a Korean-language foundation model trained entirely from scratch using these methods. Authors are transparent about scope limitations: one model family, Korean data, and loss-based metrics only.