Layer-wise MoE Routing Locality under Shared-Prefix Code Generation: Token-Identity Decomposition and Compile-Equivalent Fork Redundancy

2026-04-18

AI Summary

This study addresses the limited understanding of inter-layer routing dynamics in Mixture-of-Experts (MoE) large language models during multi-candidate code generation under shared prefixes. The authors employ tree-based search to generate diverse code completions sharing a common prefix and leverage GCC assembly alignment to disambiguate token identities. By applying Jaccard similarity and layer-wise decomposition, they analyze expert routing behavior across layers. Their analysis reveals, for the first time, a pronounced cross-layer similarity pattern in MoE routing—reaching 0.649 at identical positions, forty times higher than random baselines—thereby challenging the assumption of context-independent routing. Additionally, they introduce a redundancy metric based on compilation equivalence, showing that 67% of compilable outputs cluster into three assembly-equivalent groups, with 99.6% of intra-group differences attributable solely to comments and blank lines.

Technology Category

• Machine Learning: Mixture of Experts (MoE)
• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

In LLM-based code generation, multiple code candidates are often generated in parallel from the same prompt -- for example, in best-of-N sampling or multi-candidate code completion. These requests can share KV caches through a common prefix, yet the extent to which their Mixture-of-Experts (MoE) expert routing overlaps, and how this overlap varies across layers, remains insufficiently understood. We study Qwen3.5-35B-A3B-FP8 (256 routed experts, top-8) by performing tree-search-based branching generation from a shared prefix (851 completed codes, temperature 0.7) and analyzing the results with a compiler-output-based alignment (gcc -S -O0 assembly) that controls for token-identity confounds. Our findings are threefold: (1) At positions where both sequences generated the same token, Jaccard similarity reaches 0.649 (40x random), while even at positions with different tokens it remains 0.175 (11x random). (2) A layer-wise decomposition reveals a crossing pattern: same-token routing similarity exceeds different-token similarity across all layers, but dips in the middle layers (L14-20), while different-token similarity peaks in the middle layers at 14x random. (3) In tree-search code generation, 67% of successfully compiled codes concentrate in the top three assembly-equivalent groups, and 99.6% of within-group differences consist of comments and blank lines. We show that diversity in top-P search, including beam search, poses a significant challenge. These results refine the "context-independent routing" claim of prior work through layer-wise decomposition and suggest opportunities for improving search efficiency in LLM code generation.

Problem

• Mixture-of-Experts
• code generation
• routing locality
• shared-prefix
• layer-wise analysis

Innovation

• Mixture-of-Experts
• routing locality
• shared-prefix generation
• compiler-equivalent alignment
• layer-wise decomposition