Research Papers 论文研究 1d ago Updated 15h ago 更新于 15小时前 46

Hierarchical Copula-Gumbel-Top-K Routing: Two-Sided Dependence Control for Frozen Mixture-of-Experts at Fixed Per-Token Routing Laws 分层Copula-Gumbel-Top-K路由:固定每令牌路由定律下冻结混合专家的双侧依赖控制

Introduces Hierarchical Copula-Gumbel-Top-K (CGA) routing, a two-sided dependence control mechanism for frozen Mixture-of-Experts (MoE) models that preserves individual token routing laws while enabling joint distribution control across tokens Uses an exchangeable Gaussian copula to positively correlate Gumbel perturbations within token groups, increasing within-group expert-set coherence Employs a tunable antithetic construction across disjoint group pairs to introduce negative dependence, redu 提出分层Copula-Gumbel-Top-K(CGA)路由框架,在固定每个token边际路由律的前提下,控制不同token间路由选择的联合分布 组内通过可交换高斯Copula引入正相关,提升专家集一致性;组间通过可调对偶构造引入负相关,降低负载方差 严格证明两种操作均保持每个token的Top-K样本、混合权重和包含概率分布不变,条件期望专家流量守恒 正耦合仅膨胀负载方差,负对偶仅压缩负载方差,两个独立旋钮在不变约束面上实现一致性-分散度权衡 基础模型完全冻结,仅训练一个轻量控制器(score-function估计器),前向评估冻结网络,梯度仅流经控制器

62
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Introduces Hierarchical Copula-Gumbel-Top-K (CGA) routing, a two-sided dependence control mechanism for frozen Mixture-of-Experts (MoE) models that preserves individual token routing laws while enabling joint distribution control across tokens
  • Uses an exchangeable Gaussian copula to positively correlate Gumbel perturbations within token groups, increasing within-group expert-set coherence
  • Employs a tunable antithetic construction across disjoint group pairs to introduce negative dependence, reducing expert load variance
  • Proves that both operations preserve each token's ordered Top-K sample, mixture weights, and inclusion probabilities identically to independent routing conditioned on pre-routing logits
  • Validates the mechanism via a small-scale pilot; the frozen base model is evaluated only in the forward direction with gradients confined to a small trainable controller using a score-function estimator

Why It Matters

This work addresses a critical practical challenge in MoE models: balancing expert load dispersion against routing coherence without retraining the entire base model. For AI practitioners deploying large MoE systems, the ability to tune dependence structures via a lightweight controller offers a parameter-efficient path to managing expert utilization and reducing routing instability.

Technical Details

  • Core Problem: Given fixed per-token routing laws in a stochastic Gumbel-Top-K router, characterizes the space of reachable joint distributions over routing choices across tokens
  • Within-Group Positive Dependence: An exchangeable Gaussian copula positively correlates Gumbel perturbations at each expert coordinate for related tokens, increasing coherence in expert-set selection within groups
  • Cross-Group Negative Dependence: A tunable antithetic construction introduces selectable negative dependence across disjoint token groups, counterbalancing load concentration
  • Invariance Guarantee: Both operations provably preserve individual token marginal distributions (Top-K samples, mixture weights, inclusion probabilities) conditioned on pre-routing logits, ensuring conditional expected expert traffic remains unchanged
  • Training Architecture: A small controller driven by frozen features adjusts the two dependence dials; the frozen network runs only in the forward pass, with gradients confined to the controller via score-function estimation

Industry Insight

  • The two-dial dependence control framework offers a principled, training-efficient alternative to full fine-tuning for stabilizing MoE routing, particularly valuable for large-scale models where retraining is prohibitively expensive
  • The theoretical trade-off characterization (positive coupling inflates load variance, cross-group opposition reduces it) provides actionable guidance for practitioners tuning expert load balancing in production MoE deployments
  • While the pilot validates the mechanism, the lack of demonstrated task-level gains suggests further research is needed to connect routing coherence improvements to downstream performance; this remains an open opportunity for the community

TL;DR

  • 提出分层Copula-Gumbel-Top-K(CGA)路由框架,在固定每个token边际路由律的前提下,控制不同token间路由选择的联合分布
  • 组内通过可交换高斯Copula引入正相关,提升专家集一致性;组间通过可调对偶构造引入负相关,降低负载方差
  • 严格证明两种操作均保持每个token的Top-K样本、混合权重和包含概率分布不变,条件期望专家流量守恒
  • 正耦合仅膨胀负载方差,负对偶仅压缩负载方差,两个独立旋钮在不变约束面上实现一致性-分散度权衡
  • 基础模型完全冻结,仅训练一个轻量控制器(score-function估计器),前向评估冻结网络,梯度仅流经控制器

为什么值得看

MoE系统的核心矛盾在于路由一致性与负载均衡之间的张力——本文提供了第一个在保持边际分布严格不变约束下的双端依赖控制理论框架,为冻结大模型的高效微调开辟了新路径。

技术解析

  • 组内正耦合:在相关token组内,对每个专家坐标的Gumbel扰动施加可交换高斯Copula正相关,使组内token倾向于路由到相同专家集,提升语义一致性
  • 组间负对偶:在不相交组对之间,引入可调强度的对偶(antithetic)构造,产生可控负依赖,用于压缩跨组负载方差
  • 不变性定理:两种操作均保持每个token的条件Top-K样本分布、混合权重分布和包含概率分布与独立路由完全相同;条件期望专家流量随之守恒
  • 方差权衡刻画:正组内耦合仅使 realized expert load 方差相对独立路由膨胀;非负组间对偶仅使方差相对平耦合压缩——两个互补旋钮在不变约束面上独立调节
  • 训练机制:基础模型完全冻结,仅一个轻量控制器作用于冻结特征,通过score-function估计器训练;前向传播评估冻结网络,梯度仅流经控制器,计算开销极低

行业启示

  • 为MoE系统的负载管理提供了理论完备的"旋钮式"控制工具,无需重新训练基础模型即可调节一致性与负载均衡的权衡
  • 冻结模型+轻量控制器的范式适合大规模部署场景,可快速迭代路由策略而无需承担全量微调的成本
  • 当前仅为小规模验证,尚未建立任务级微调收益;后续需在真实MoE基准上验证CGA对吞吐、负载均衡和下游性能的改善效果

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Research 科学研究 Training 训练 Inference 推理