Feilong Liu

Feilong Liu

Mixture-of-Experts (MoE) architectures achieve scalable capacity through sparse routing, yet the geometric structure of expert specialization remains poorly understood. We introduce a unified Jacobian-PCA-Grassmann framework for analyzing MoE layers in both function space and representation space. Across pretrained MoE Transformers (Mistral, Qwen), we find a consistent structural asymmetry: experts exhibit strong functional decorrelation (consistently low, near-zero cross-expert Jacobian alignment) while their routed representations occupy distinct but partially overlapping subspaces. This indicates that functional decorrelation and representation overlap coexist rather than coincide in MoE specialization. Controlled routing experiments further indicate that routing sparsity appears to be a key factor shaping this geometry: top-k routing induces sharper functional separation and larger subspace divergence, whereas fully soft routing yields more entangled expert structure. Together, these results suggest a geometric interpretation in which MoE layers may be viewed as implementing locally decorrelated operators over overlapping submanifolds on a shared representation manifold, and provide a general diagnostic framework for studying conditional computation in modern Transformer architectures.

Feilong Liu

Mixture-of-Experts (MoE) architectures are widely used for efficiency and conditional computation, but their effect on the geometry of learned functions and representations remains poorly understood. We study MoEs through a geometric lens, interpreting routing as soft partitioning into overlapping expert-local charts. We introduce a Dual Jacobian-PCA spectral probe that analyzes local function geometry via Jacobian singular value spectra and representation geometry via weighted PCA of routed hidden states. Using a controlled MLP-MoE setting with exact Jacobian computation, we compare dense, Top-k, and fully soft routing under matched capacity. Across random seeds, MoE routing consistently reduces local sensitivity: expert-local Jacobians show smaller leading singular values and faster spectral decay than dense baselines. Weighted PCA reveals that expert-local representations distribute variance across more principal directions, indicating higher effective rank. We further observe low alignment among expert Jacobians, suggesting decomposition into low-overlap expert-specific transformations. Routing sharpness modulates these effects: Top-k routing yields more concentrated, lower-rank expert structure, while fully soft routing produces broader, higher-rank representations. Experiments on a 3-layer transformer with WikiText confirm curvature reduction on natural language and show lower cross-expert alignment for Top-k routing. These findings support interpreting MoEs as soft partitionings of function space that flatten local curvature while redistributing representation variance, yielding testable predictions for expert scaling, hallucination reduction, and ensemble diversity.

Feilong Liu

Rotary positional embeddings (RoPE) are widely used in large language models to encode token positions through multiplicative rotations, yet their behavior at long context lengths remains poorly characterized. In this work, we reinterpret RoPE as phase modulation applied to a bank of complex oscillators, enabling analysis through classical signal processing theory. Under this formulation, we derive principled lower bounds on the RoPE base parameter that are necessary to preserve positional coherence over a target context length. These include a fundamental aliasing bound, analogous to a Nyquist limit, and a DC-component stability bound that constrains phase drift in low-frequency positional modes. We further extend this analysis to deep transformers, showing that repeated rotary modulation across layers compounds angular misalignment, tightening the base requirement as depth increases. Complementing these results, we derive a precision-dependent upper bound on the RoPE base arising from finite floating-point resolution. Beyond this limit, incremental phase updates become numerically indistinguishable, leading to positional erasure even in the absence of aliasing. Together, the lower and upper bounds define a precision- and depth-dependent feasibility region a Goldilocks zone for long-context transformers. We validate the framework through a comprehensive case study of state-of-the-art models, including LLaMA, Mistral, and DeepSeek variants, showing that observed successes, failures, and community retrofits align closely with the predicted bounds. Notably, models that violate the stability bound exhibit attention collapse and long-range degradation, while attempts to scale beyond one million tokens encounter a hard precision wall independent of architecture or training.