---
id: 20261110-T0-10
title: "AlterNet：按输入难度动态分配算力的MoE架构，可弹性推理"
title_en: "AlterNet Combines Mixture-of-Experts with Learned Halting for Elastic Inference"
url: https://ai.daily.yangsir.net/daily/20261110-T0-10
issue_date: 2026-11-10
publish_date: 2026-11-09T05:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2611.08741
---

# AlterNet：按输入难度动态分配算力的MoE架构，可弹性推理

多数LLM架构的计算量在训练时就固定了，推理时无法按需伸缩。AlterNet把混合专家（MoE）与浅层专家、动态深度机制结合，并用学到的停止策略决定前向传播走多深，从而根据输入难度弹性分配计算。简单说：简单问题少算，复杂问题多算。对推理成本敏感的部署场景，这类架构的价值在于不牺牲固定算力预算就能应对难度波动。

## English Version

**AlterNet Combines Mixture-of-Experts with Learned Halting for Elastic Inference**

Most LLM architectures have fixed compute budgets set at training time, leaving no room to adapt at inference. AlterNet combines mixture-of-experts with shallow experts and a dynamic-depth system, using a learned halting policy to decide how deep a forward pass should go—allocating more compute to hard inputs and less to easy ones. For cost-sensitive deployments, the appeal is handling difficulty variance without raising the fixed compute budget.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2611.08741)

**详情页**：https://ai.daily.yangsir.net/daily/20261110-T0-10

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*