Saved in:
Bibliographic Details
Main Authors: Yoon, Youngsik, Wang, Siwei, Chen, Wei, Ok, Jungseul
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.07260
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909025754415104
author Yoon, Youngsik
Wang, Siwei
Chen, Wei
Ok, Jungseul
author_facet Yoon, Youngsik
Wang, Siwei
Chen, Wei
Ok, Jungseul
contents Mixture-of-Experts (MoE) language models route each token to a small subset of experts, but whether the routes selected by a trained top-$k$ router are good ones is rarely evaluated directly. Holding the model fixed, we compare each standard route against sampled equal-compute alternatives for the same token and score each by the next-token probability it assigns to the realized token in a verified reasoning trajectory. The result is sharply token-conditional: the standard router is well-aligned with route utility on confident tokens but uninformative on the fragile tokens that drive hard reasoning, where lower-loss equal-compute routes consistently exist inside the frozen model but are not selected. The same pattern holds across Qwen3-30B-A3B, GPT-OSS-20B, DeepSeek-V2-Lite, and OLMoE-1B-7B, and follows structurally from how standard top-$k$ training evaluates routing decisions: the language modeling loss scores only the executed route, and load balancing depends only on aggregate routing statistics. A minimal router-only update to the final-layer router, leaving every expert and every other router frozen, is sufficient to shift pass@K on AIME 2024+2025 and HMMT 2025 for both Qwen3-30B-A3B and GPT-OSS-20B, suggesting that at least part of the failure reflects router-reachable misallocation rather than expert capacity alone.
format Preprint
id arxiv_https___arxiv_org_abs_2605_07260
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle When Are Experts Misrouted? Counterfactual Routing Analysis in Mixture-of-Experts Language Models
Yoon, Youngsik
Wang, Siwei
Chen, Wei
Ok, Jungseul
Machine Learning
Computation and Language
Mixture-of-Experts (MoE) language models route each token to a small subset of experts, but whether the routes selected by a trained top-$k$ router are good ones is rarely evaluated directly. Holding the model fixed, we compare each standard route against sampled equal-compute alternatives for the same token and score each by the next-token probability it assigns to the realized token in a verified reasoning trajectory. The result is sharply token-conditional: the standard router is well-aligned with route utility on confident tokens but uninformative on the fragile tokens that drive hard reasoning, where lower-loss equal-compute routes consistently exist inside the frozen model but are not selected. The same pattern holds across Qwen3-30B-A3B, GPT-OSS-20B, DeepSeek-V2-Lite, and OLMoE-1B-7B, and follows structurally from how standard top-$k$ training evaluates routing decisions: the language modeling loss scores only the executed route, and load balancing depends only on aggregate routing statistics. A minimal router-only update to the final-layer router, leaving every expert and every other router frozen, is sufficient to shift pass@K on AIME 2024+2025 and HMMT 2025 for both Qwen3-30B-A3B and GPT-OSS-20B, suggesting that at least part of the failure reflects router-reachable misallocation rather than expert capacity alone.
title When Are Experts Misrouted? Counterfactual Routing Analysis in Mixture-of-Experts Language Models
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2605.07260