Xueqing Zhou

Hongxuan Wu, Yukun Zhang, Xueqing Zhou

When a multimodal Transformer answers a visual question, is the prediction driven by visual evidence, linguistic reasoning, or genuinely fused cross-modal computation -- and how does this structure evolve across layers? We address this question with a layer-wise framework based on Partial Information Decomposition (PID) that decomposes the predictive information at each Transformer layer into redundant, vision-unique, language-unique, and synergistic components. To make PID tractable for high-dimensional neural representations, we introduce \emph{PID Flow}, a pipeline combining dimensionality reduction, normalizing-flow Gaussianization, and closed-form Gaussian PID estimation. Applying this framework to LLaVA-1.5-7B and LLaVA-1.6-7B across six GQA reasoning tasks, we uncover a consistent \emph{modal transduction} pattern: visual-unique information peaks early and decays with depth, language-unique information surges in late layers to account for roughly 82\% of the final prediction, and cross-modal synergy remains below 2\%. This trajectory is highly stable across model variants (layer-wise correlations $>$0.96) yet strongly task-dependent, with semantic redundancy governing the detailed information fingerprint. To establish causality, we perform targeted Image$\rightarrow$Question attention knockouts and show that disrupting the primary transduction pathway induces predictable increases in trapped visual-unique information, compensatory synergy, and total information cost -- effects that are strongest in vision-dependent tasks and weakest in high-redundancy tasks. Together, these results provide an information-theoretic, causal account of how vision becomes language in multimodal Transformers, and offer quantitative guidance for identifying architectural bottlenecks where modality-specific information is lost.

Yukun Zhang, Xueqing Zhou

The Transformer architecture has revolutionized artificial intelligence, yet a principled theoretical understanding of its internal mechanisms remains elusive. This paper introduces a novel analytical framework that reconceptualizes the Transformer's discrete, layered structure as a continuous spatiotemporal dynamical system governed by a master Partial Differential Equation (PDE). Within this paradigm, we map core architectural components to distinct mathematical operators: self-attention as a non-local interaction, the feed-forward network as a local reaction, and, critically, residual connections and layer normalization as indispensable stabilization mechanisms. We do not propose a new model, but rather employ the PDE system as a theoretical probe to analyze the mathematical necessity of these components. By comparing a standard Transformer with a PDE simulator that lacks explicit stabilizers, our experiments provide compelling empirical evidence for our central thesis. We demonstrate that without residual connections, the system suffers from catastrophic representational drift, while the absence of layer normalization leads to unstable, explosive training dynamics. Our findings reveal that these seemingly heuristic "tricks" are, in fact, fundamental mathematical stabilizers required to tame an otherwise powerful but inherently unstable continuous system. This work offers a first-principles explanation for the Transformer's design and establishes a new paradigm for analyzing deep neural networks through the lens of continuous dynamics.

Yukun Zhang, Xueqing Zhou

We propose a novel framework, Continuous_Time Attention, which infuses partial differential equations (PDEs) into the Transformer's attention mechanism to address the challenges of extremely long input sequences. Instead of relying solely on a static attention matrix, we allow attention weights to evolve over a pseudo_time dimension via diffusion, wave, or reaction_diffusion dynamics. This mechanism systematically smooths local noise, enhances long_range dependencies, and stabilizes gradient flow. Theoretically, our analysis shows that PDE_based attention leads to better optimization landscapes and polynomial rather than exponential decay of distant interactions. Empirically, we benchmark our method on diverse experiments_demonstrating consistent gains over both standard and specialized long sequence Transformer variants. Our findings highlight the potential of PDE_based formulations to enrich attention mechanisms with continuous_time dynamics and global coherence.

Yukun Zhang, Xueqing Zhou

We propose PDE-Transformer, a novel sequence modeling paradigm that casts the forward pass of a Transformer as the numerical discretization of a continuous reaction-diffusion system derived from a variational energy functional. In our framework, token embeddings evolve under a partial differential equation whose nonlocal integral term models self-attention, local reaction term models feed-forward layers, diffusion term encodes positional smoothing, and a stability control term corresponds to layer normalization. From this unifying perspective, we design an Adaptive PDE Diffusion Layer-an efficient, learnable finite-difference stencil that enforces local smoothness in feature space with linear time complexity and complements self-attention's global routing. Through a systematic theoretical analysis based on four pillars:stability, diffusion geometry, multi-scale dynamics, and component coupling, we derive principled guidelines for integrating the PDE layer at seven candidate points in the Transformer. Empirically, on the Long Range Arena benchmark, placing the layer immediately after embedding yields a 4.1 pp average accuracy gain over a strong baseline, and an adaptive multi-scale variant delivers further improvements. Our work thus offers a principled, lightweight mechanism to bolster long-range dependency modeling by harmonizing continuous PDE smoothing with discrete self-attention.