Softmax Routing Dynamics in Mixture-of-Experts Models: Collapse and Recovery
Read the original on Hugging Face Trending Papers →The Flow has not summarised this story yet — read it at Hugging Face Trending Papers.
The Flow has not summarised this story yet — read it at Hugging Face Trending Papers.
arXiv:2610.07207v1 Announce Type: cross Abstract: Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability probl...
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 investigates a dense analogue of Sparse Mixture-of-Experts (MoE) models by using $K$ SwiGLU experts that are all active for every token and combined via a softmax gate, keeping the total feed‑forward network (FFN) width fixed. Validation loss shows a non‑monotonic relationship with $K$: $K=2$ slightly improves performance over the single‑expert baseline, while $K=4$ and $K=6$ degrade it. The study also finds that the gating mechanism remains largely soft and balanced, except for a near one‑hot routing in the first layer of the $K=4$ model, which is functionally important as forcing uniform routing increases loss significantly.
The paper investigates how Mixture-of-Experts (MoE) models encode moral content compared to dense models. While linear probes can recover moral valence from almost all expert-layer combinations with high accuracy, these representations are far more fragile to activation noise, showing a 4.2‑fold drop in robustness. The authors attribute this fragility to output dilution: the MoE block averages across active experts, reducing the feedforward signal by nearly two orders of magnitude, which makes moral information vulnerable to perturbation even though routing remains stable.
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.
arXiv:2609.16179v1 Announce Type: new Abstract: Z-loss has been widely applied to the logits of language-model output heads and sparse mixture-of-experts routers. Z-loss constrains the softmax log-no...