The paper investigates how different expert‑deactivation strategies at inference time—removal, freezing, remaining‑expert renormalization, and unnormalized re‑routing—affect token routing and expert weights. Experiments across five instruction‑tuned checkpoints from four model families, plus a base OLMoE checkpoint, show that the choice of strategy interacts with the training‑time weight rule: renormalization performs best when models were trained with raw router probabilities, while unnormalized re‑routing is superior when trained with normalized weights. A weight‑sum model links loss of expert contributions and branch‑scale changes to the training rule, revealing that mismatches can inflate perplexity by up to 28,740× and that additional renormalization penalties grow as fewer experts are activated.
arXiv:2609. 04575v1 Announce Type: cross Abstract: Modern fine-grained Mixture-of-Experts (MoE) models route each token to a small number of experts and renormalize their router probabilities.
The paper investigates training‑free pruning of Mixture‑of‑Experts (MoE) models, focusing on how to handle the computation when a router selects a pruned expert. It compares replacement operators from three families on several models (OLMoE, DeepSeek‑V2‑Lite, Moonlight, Granite) while keeping the pruning mask fixed, and introduces a new operator called the tilted novice. The study finds that no single operator consistently dominates, with performance varying by model, pruning level, and evaluation metric, and it proposes a five‑item checklist for reporting results.
UpgradeBench is a decision‑centric longitudinal benchmark that evaluates how fine‑tuned language‑model specialists should be handled when new base‑model releases occur. It covers four consecutive Qwen releases, a continuation checkpoint, six tasks, two model sizes, and OLMo checkpoints with known training lineage, and examines whether retraining, adapter transfer, or other recovery strategies improve specialist performance. The benchmark reveals that upgrade gains vary by task and release interval, that direct adapter copying is sensitive to pretraining distance, and that teacher relabeling can recover specialists without new annotations. "whyItMatters":"The study provides actionable insights into the cost‑effective management of specialist models across model releases, showing how to balance retraining effort with performance gains."
The paper examines how expert pruning—removing low‑importance experts in Mixture‑of‑Experts models—fails when the router is over‑dispersed, a condition caused by aggressive load‑balancing that spreads tokens nearly uniformly across experts. In this regime, traditional importance signals from router probabilities collapse, making perplexity an unreliable predictor of downstream accuracy; for example, the lowest‑perplexity pruning on gpt‑oss‑20B harms mathematical reasoning while the highest‑perplexity pruning preserves it. To address this, the authors introduce Minimax Expert Score Allocation (MESA), a domain‑aware method that iteratively boosts scores for the most affected domain, achieving minimal worst‑case degradation across domains and outperforming baseline pruning strategies on multiple benchmarks while reducing memory usage.
ACE introduces a training‑free, calibration‑free framework for adaptive expert skipping in Mixture‑of‑Experts LLMs. It combines a Global Spectral Proxy that estimates global transformation capacity with a Router‑Conditioned Refinement that builds expert‑specific direction prototypes, enabling the model to skip low‑contribution experts while always keeping the top‑1 expert. Offline computation of expert statistics leaves only lightweight table lookups during inference, and experiments on three MoE‑based LLMs show ACE outperforms static and dynamic baselines, especially at high skipping ratios.
arXiv:2606. 10338v1 Announce Type: cross Abstract: Machine unlearning is increasingly important for large language models, yet unlearning in Mixture-of-Experts (MoE) architectures remains underexplored.
arXiv:2609.36222v1 Announce Type: new Abstract: Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferrin...
The paper introduces SafeStart, a technique for Mixture-of-Experts (MoE) language models that mitigates the safety degradation caused by active expert reduction during serving. By running all experts for the prompt and the first eight generated tokens before switching to a reduced set, SafeStart largely restores the model’s refusal to harmful requests while preserving most of the speedup achieved by expert reduction. Experiments on several MoE models show that SafeStart maintains high offline decode speedups (≈85% of the original) and limits the increase in harmful responses to within a few percentage points, even under jailbreak prompts.
The paper studies how the design of Mixture-of-Experts (MoE) routers affects inference speed when combined with Speculative Decoding (SD). It shows that routers promoting high expert coactivation reduce memory transfer costs and improve runtime. By integrating a global load‑balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanism, the authors achieve a 21% throughput gain over baseline MoEs while preserving accuracy.
The paper introduces a two‑level readout for mixture‑of‑experts reasoning models. First, it compresses the model’s internal reasoning states into a 64‑dimensional semantic frame (J64) that reveals process dynamics beyond the emitted trace. Second, it reconstructs this frame from native expert‑routing statistics (R64), achieving high correlation and preserving most predictive gains while enabling low‑overhead, test‑time decision making.
arXiv:2506. 14126v2 Announce Type: replace-cross Abstract: Modern deep learning is increasingly characterized by the use of open-weight foundation models that can be fine-tuned on specialized datasets.
The paper introduces SAME (Stabilized Mixture-of-Experts) to address challenges in Multimodal Continual Instruction Tuning (MCIT) for large language models. SAME mitigates router drift by decomposing routing dynamics into orthogonal subspaces and updating only task-relevant directions, while preventing expert drift through curvature‑aware scaling that uses historical input covariance without rehearsal. The method also employs adaptive expert activation to freeze selected experts during training, reducing redundant computation and cross‑task interference, and demonstrates state‑of‑the‑art performance on a new long‑task‑sequence benchmark.