arXiv subjects

Yukun Zhang

Publications and source records attributed to Yukun Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2026

Rethinking Structure Preservation in Text-Guided Image Editing with Visual Autoregressive Models

Visual autoregressive (VAR) models have recently emerged as a promising family of generative models, enabling a wide range of downstream vision tasks such as text-guided image editing. By shifting the editing paradigm from noise manipulation in diffusion-based methods to token-level operations, VAR-based approaches achieve better background preservation and significantly faster inference. However, existing VAR-based editing methods still face two key challenges: accurately localizing editable tokens and maintaining structural consistency in the edited results. In this work, we propose a novel text-guided image editing framework rooted in an analysis of intermediate feature distributions within VAR models. First, we introduce a coarse-to-fine token localization strategy that can refine editable regions, balancing editing fidelity and background preservation. Second, we analyze the intermediate representations of VAR models and identify structure-related features, by which we design a simple yet effective feature injection mechanism to enhance structural consistency between the edited and source images. Third, we develop a reinforcement learning-based adaptive feature injection scheme that automatically learns scale- and layer-specific injection ratios to jointly optimize editing fidelity and structure preservation. Extensive experiments demonstrate that our method achieves superior structural consistency and editing quality compared with state-of-the-art approaches, across both local and global editing scenarios.

cs.CV↗

arXiv2026

How Vision Becomes Language: A Layer-wise Information-Theoretic Analysis of Multimodal Reasoning

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.

cs.AI↗

arXiv2026

Where to Add PDE Diffusion in Transformers

Transformers enable powerful content-based global routing via self-attention, but they lack an explicit local geometric prior along the sequence axis. As a result, the placement of locality-inducing modules in hybrid architectures has largely been empirical. We study a simple deterministic PDE diffusion layer implemented as one explicit Euler step of one-dimensional heat smoothing using a discrete Neumann Laplacian under a spectral stability constraint, and ask a structural question: where should diffusion be inserted relative to attention? Our central claim is that diffusion and attention generally do not commute, so inserting the same local operator before versus after attention leads to qualitatively different behaviors. We develop a three-layer operator-theoretic framework that (1) establishes unconditional guarantees for the diffusion subsystem, including spectral non-expansiveness and monotone Dirichlet-energy dissipation when the diffusion step size is smaller than one half, (2) derives compositional perturbation bounds linking insertion effects to representation roughness and downstream amplification, and (3) uses diffusion-attention non-commutativity as a diagnostic for structural double-mixing conflicts. Guided by theory, we evaluate seven insertion positions on the Long Range Arena benchmark. Early diffusion acts as effective pre-regularization, improving average accuracy by 4.1 percentage points when applied after embedding, while post-attention diffusion degrades performance by 2.5 percentage points, consistent with the predicted conflict. A multi-scale diffusion variant yields consistent gains under the same global stability constraint. Our analysis provides a general template for reasoning about local-global compositions in sequence models by separating provable guarantees, compositional bounds, and mechanistic diagnostics.

cs.LG↗

arXiv2026

SMES: Towards Scalable Multi-Task Recommendation via Expert Sparsity

Industrial recommender systems typically rely on multi-task learning to estimate diverse user feedback signals and aggregate them for ranking. Recent advances in model scaling have shown promising gains in recommendation. However, naively increasing model capacity imposes prohibitive online inference costs and often yields diminishing returns for sparse tasks with skewed label distributions. This mismatch between uniform parameter scaling and heterogeneous task capacity demands poses a fundamental challenge for scalable multi-task recommendation. In this work, we investigate parameter sparsification as a principled scaling paradigm and identify two critical obstacles when applying sparse Mixture-of-Experts (MoE) to multi-task recommendation: exploded expert activation that undermines instance-level sparsity and expert load skew caused by independent task-wise routing. To address these challenges, we propose SMES, a scalable sparse MoE framework with progressive expert routing. SMES decomposes expert activation into a task-shared expert subset jointly selected across tasks and task-adaptive private experts, explicitly bounding per-instance expert execution while preserving task-specific capacity. In addition, SMES introduces a global multi-gate load-balancing regularizer that stabilizes training by regulating aggregated expert utilization across all tasks. SMES has been deployed in Kuaishou large-scale short-video services, supporting over 400 million daily active users. Extensive online experiments demonstrate stable improvements, with GAUC gain of 0.29% and a 0.31% uplift in user watch time.

cs.IR↗

arXiv2026

The Economics of Digital Intelligence Capital: Endogenous Depreciation and the Structural Jevons Paradox

This paper develops a micro-founded economic theory of the AI industry by modeling large language models as a distinct asset class-Digital Intelligence Capital-characterized by data-compute complementarities, increasing returns to scale, and relative (rather than absolute) valuation. We show that these features fundamentally reshape industry dynamics along three dimensions. First, because downstream demand depends on relative capability, innovation by one firm endogenously depreciates the economic value of rivals' existing capital, generating a persistent innovation pressure we term the Red Queen Effect. Second, falling inference prices induce downstream firms to adopt more compute-intensive agent architectures, rendering aggregate demand for compute super-elastic and producing a structural Jevons paradox. Third, learning from user feedback creates a data flywheel that can destabilize symmetric competition: when data accumulation outpaces data decay, the market bifurcates endogenously toward a winner-takes-all equilibrium. We further characterize conditions under which expanding upstream capabilities erode downstream application value (the Wrapper Trap). A calibrated agent-based model confirms these mechanisms and their quantitative implications. Together, the results provide a unified framework linking intelligence production upstream with agentic demand downstream, offering new insights into competition, scalability, and regulation in the AI economy.

econ.GN↗

arXiv2026

Generative AI as a Non-Convex Supply Shock: Market Bifurcation and Welfare Analysis

The diffusion of Generative AI (GenAI) constitutes a supply shock of a fundamentally different nature: while marginal production costs approach zero, content generation creates congestion externalities through information pollution. We develop a three-layer general equilibrium framework to study how this non-convex technology reshapes market structure, transition dynamics, and social welfare. In a static vertical differentiation model, we show that the GenAI cost shock induces a kinked production frontier that bifurcates the market into exit, AI, and human segments, generating a ``middle-class hollow'' in the quality distribution. To analyze adjustment paths, we embed this structure in a mean-field evolutionary system and a calibrated agent-based model with bounded rationality. The transition to the AI-integrated equilibrium is non-monotonic: rather than smooth diffusion, the economy experiences a temporary ecological collapse driven by search frictions and delayed skill adaptation, followed by selective recovery. Survival depends on asymmetric skill reconfiguration, whereby humans retreat from technical execution toward semantic creativity. Finally, we show that the welfare impact of AI adoption is highly sensitive to pollution intensity: low congestion yields monotonic welfare gains, whereas high pollution produces an inverted-U relationship in which further AI expansion reduces total welfare. These results imply that laissez-faire adoption can be inefficient and that optimal governance must shift from input regulation toward output-side congestion management.

econ.GN↗

arXiv2026

The Impact of Generative AI on Content Platforms: A Two-Sided Market Analysis with Multi-Dimensional Quality Heterogeneity

This paper presents a unified computational framework to examine how generative AI (GenAI) reshapes welfare, inequality, and diversity in content platform economies. By integrating welfare economics with agent-based simulations, we model the co-evolutionary dynamics among AI generators, human creators, and consumers within a two-sided market characterized by multi-dimensional quality heterogeneity. Unlike static models, our framework endogenizes AI learning as a function of human data synthesis and models human adaptation as a strategic reallocation of skills toward creative niches. The results reveal that while GenAI significantly enhances consumer surplus through technical quality gains and price depression, it triggers a skill-biased displacement of human incumbents and intensifies market concentration. Through the evaluation of six governance regimes, we identify a fundamental ``Policy Trilemma'' where platforms must navigate non-trivial trade-offs between allocative efficiency, distributional equity, and ecosystem sustainability. Our findings highlight that algorithmic diversity and pro-creative commission structures function as essential economic mechanisms for sustaining long-tail participation and inclusive social welfare in the generative AI era.

cs.CY↗

arXiv2026

Low-Complexity Monitoring and Compensation of Transceiver IQ Imbalance by Multi-dimensional Architecture for Dual-Polarization 16 Quadrature Amplitude Modulation

In this paper, a low-complexity multi-dimensional architecture for IQ imbalance compensation is proposed, which reduces the effects of in-phase (I) and quadrature (Q) imbalance. The architecture use a transceiver IQ skew estimation structure to compensate for IQ skew, and then use a low-complexity MIMO equalizer to compensate for IQ amplitude/phase imbalance. In the transceiver IQ skew estimation structure, the receiver(RX) IQ skew is estimated by Gardner's phase detector, and the transmitter TX skew is estimated by finding the value that yields the lowest equalizer error. The low-complexity MIMO equalizer consists of a complex-valued MIMO (CV-MIMO) and a two-layer multimodulus algorithm real-valued MIMO (TMMA-RV-MIMO), which employ a butterfly and a non-butterfly structure, respectively. The CV-MIMO is used to perform polarization demultiplexing and the TMMA-RV-MIMO equalizes each of the two polarizations. In addition, the TMMA-RV-MIMO can recovery the carrier phase. A 100 km transmission simulation and experiment with 36 Gbaud dual-polarization 16 quadrature amplitude modulation (DP-16QAM) signals showed that, with the TX/RX IQ skew estimation, the estimation error is less than 0.9/0.25 ps. The low-complexity MIMO equalizer can tolerate 0.1 TX IQ amplitude imbalance and 5 degrees at a 0.3 dB Q-factor penalty. The number of real multiplications is reduced by 55% compared with conventional cases in total.

cs.NI↗

arXiv2026

The Economics of Information Pollution in the Age of AI: General Equilibrium, Welfare, and Policy Design

The advent of Large Language Models (LLMs) represents a fundamental shock to the economics of information production. By asymmetrically collapsing the marginal cost of generating low-quality, synthetic content while leaving high-quality production costly, AI systematically incentivizes information pollution. This paper develops a general equilibrium framework to analyze this challenge. We model the strategic interactions among a monopolistic platform, profit-maximizing producers, and utility-maximizing consumers in a three-stage game. The core of our model is a production technology with differential elasticities of substitution (σ_L > 1 > σ_H), which formalizes the insight that AI is a substitute for labor in low-quality production but a complement in high-quality creation. We prove the existence of a unique "Polluted Information Equilibrium" and demonstrate its inefficiency, which is driven by a threefold market failure: a production externality, a platform governance failure, and an information commons externality. Methodologically, we derive a theoretically-grounded Information Pollution Index (IPI) with endogenous welfare weights to measure ecosystem health. From a policy perspective, we show that a first-best outcome requires a portfolio of instruments targeting each failure. Finally, considering the challenges of deep uncertainty, we advocate for an adaptive governance framework where policy instruments are dynamically adjusted based on real-time IPI readings, offering a robust blueprint for regulating information markets in the age of AI.

cs.CY↗

arXiv2025

Integrating Domain Knowledge for Financial QA: A Multi-Retriever RAG Approach with LLMs

This research project addresses the errors of financial numerical reasoning Question Answering (QA) tasks due to the lack of domain knowledge in finance. Despite recent advances in Large Language Models (LLMs), financial numerical questions remain challenging because they require specific domain knowledge in finance and complex multi-step numeric reasoning. We implement a multi-retriever Retrieval Augmented Generators (RAG) system to retrieve both external domain knowledge and internal question contexts, and utilize the latest LLM to tackle these tasks. Through comprehensive ablation experiments and error analysis, we find that domain-specific training with the SecBERT encoder significantly contributes to our best neural symbolic model surpassing the FinQA paper's top model, which serves as our baseline. This suggests the potential superior performance of domain-specific training. Furthermore, our best prompt-based LLM generator achieves the state-of-the-art (SOTA) performance with significant improvement (>7%), yet it is still below the human expert performance. This study highlights the trade-off between hallucinations loss and external knowledge gains in smaller models and few-shot examples. For larger models, the gains from external facts typically outweigh the hallucination loss. Finally, our findings confirm the enhanced numerical reasoning capabilities of the latest LLM, optimized for few-shot learning.

cs.CL↗

arXiv2025

Quantum computing quantum Monte Carlo algorithm

Quantum computing and quantum Monte Carlo (QMC) are respectively the state-of-the-art quantum and classical computing methods for understanding many-body quantum systems. Here, we propose a hybrid quantum-classical algorithm that integrates these two methods, inheriting their distinct features in efficient representation and manipulation of quantum states and overcoming their limitations. We first introduce non-stoquasticity indicators (NSIs) and their upper bounds, which measure the sign problem, the most notable limitation of QMC. We show that our algorithm could greatly mitigate the sign problem, which decreases NSIs with the assistance of quantum computing. Meanwhile, the use of quantum Monte Carlo also increases the expressivity of shallow quantum circuits, allowing more accurate computation that is conventionally achievable only with much deeper circuits. We numerically test and verify the method for the N$_2$ molecule (12 qubits) and the Hubbard model (16 qubits). Our work paves the way to solving practical problems with intermediate-scale and early-fault tolerant quantum computers, with potential applications in chemistry, condensed matter physics, materials, high energy physics, etc.

quant-ph↗

arXiv2025

Obtaining Accurate Ground-State Properties on Near-term Quantum Devices

Accurate ground-state calculations on noisy quantum computers are fundamentally limited by restricted ansatz expressivity and unavoidable hardware errors. We introduce a hybrid-quantum classical framework that simultaneously addresses these challenges. Our method systematically purifies noisy two electron reduced density matrices from quantum devices by enforcing N-representability conditions through efficient semidefinite programming, guided by a norm-based distance constraint to the experimental data. To implement this constraint, we develop a hardware efficient calibration protocol based on Clifford circuits. We demonstrate near full configuration interaction accuracy for ground-state energies of H2, LiH, and H4, and compute precise scattering intensities for C6H8 on noisy hardware. This approach surpasses conventional methods by simultaneously overcoming both ansatz limitations and hardware noise, establishing a scalable route to quantum advantage and marking a critical step toward reliable simulations of complex molecularnsystems on noisy devices.

quant-ph↗

arXiv2025

Heisenberg-Limited Quantum Eigenvalue Estimation for Non-normal Matrices

Estimating the eigenvalues of non-normal matrices is a foundational problem with far-reaching implications, from modeling non-Hermitian quantum systems to analyzing complex fluid dynamics. Yet, this task remains beyond the reach of standard quantum algorithms, which are predominantly tailored for Hermitian matrices. Here we introduce a new class of quantum algorithms that directly address this challenge. The central idea is to construct eigenvalue signals through customized quantum simulation protocols and extract them using advanced classical signal-processing techniques, thereby enabling accurate and efficient eigenvalue estimation for general non-normal matrices. Crucially, when supplied with purified quantum state inputs, our algorithms attain Heisenberg-limited precision--achieving optimal performance. These results extend the powerful guided local Hamiltonian framework into the non-Hermitian regime, significantly broadening the frontier of quantum computational advantage. Our work lays the foundation for a rigorous and scalable quantum computing approach to one of the most demanding problems in linear algebra.

quant-ph↗

arXiv2025

Hierarchical Alignment: Surgical Fine-Tuning via Functional Layer Specialization in Large Language Models

Existing alignment techniques for Large Language Models (LLMs), such as Direct Preference Optimization (DPO), typically treat the model as a monolithic entity, applying uniform optimization pressure across all layers. This approach overlooks the functional specialization within the Transformer architecture, where different layers are known to handle distinct tasks from syntax to abstract reasoning. In this paper, we challenge this one-size-fits-all paradigm by introducing Hierarchical Alignment, a novel method that applies targeted DPO to distinct functional blocks of a model's layers: local (syntax), intermediate (logic), and global (factuality). Through a series of controlled experiments on state-of-the-art models like Llama-3.1-8B and Qwen1.5-7B using LoRA for surgical fine-tuning, our results, evaluated by a powerful LLM-as-Judge, demonstrate significant and predictable improvements. Specifically, aligning the local layers (Local-Align) enhances grammatical fluency. More importantly, aligning the global layers (Global-Align) not only improves factual consistency as hypothesized but also proves to be the most effective strategy for enhancing logical coherence, outperforming all baselines. Critically, all hierarchical strategies successfully avoid the "alignment tax" observed in standard DPO, where gains in fluency come at the cost of degraded logical reasoning. These findings establish a more resource-efficient, controllable, and interpretable path for model alignment, highlighting the immense potential of shifting from monolithic optimization to structure-aware surgical fine-tuning to build more advanced and reliable LLMs.

cs.CL↗

arXiv2025

Multi-Scale Manifold Alignment for Interpreting Large Language Models: A Unified Information-Geometric Framework

We present Multi-Scale Manifold Alignment(MSMA), an information-geometric framework that decomposes LLM representations into local, intermediate, and global manifolds and learns cross-scale mappings that preserve geometry and information. Across GPT-2, BERT, RoBERTa, and T5, we observe consistent hierarchical patterns and find that MSMA improves alignment metrics under multiple estimators (e.g., relative KL reduction and MI gains with statistical significance across seeds). Controlled interventions at different scales yield distinct and architecture-dependent effects on lexical diversity, sentence structure, and discourse coherence. While our theoretical analysis relies on idealized assumptions, the empirical results suggest that multi-objective alignment offers a practical lens for analyzing cross-scale information flow and guiding representation-level control.

cs.CL↗

arXiv2025

Exponential quantum advantages for practical non-Hermitian eigenproblems

Non-Hermitian physics has emerged as a rich field of study, with applications ranging from $PT$-symmetry breaking and skin effects to non-Hermitian topological phase transitions. Yet most studies remain restricted to small-scale or classically tractable systems. While quantum computing has shown strong performance in Hermitian eigenproblems, its extension to the non-Hermitian regime remains largely unexplored. Here, we develop a quantum algorithm to address general non-Hermitian eigenvalue problems, specifically targeting eigenvalues near a given line in the complex plane -- thereby generalizing previous results on ground state energy and spectral gap estimation for Hermitian matrices. Our method combines a fuzzy quantum eigenvalue detector with a divide-and-conquer strategy to efficiently isolate relevant eigenvalues. This yields a provable exponential quantum speedup for non-Hermitian eigenproblems. Furthermore, we discuss the broad applications in detecting spontaneous $PT$-symmetry breaking, estimating Liouvillian gaps, and analyzing classical Markov processes. These results highlight the potential of quantum algorithms in tackling challenging problems across quantum physics and beyond.

quant-ph↗

arXiv2025

Understanding Transformer Architecture through Continuous Dynamics: A Partial Differential Equation Perspective

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.

cs.LG↗

arXiv2025

Classical Algorithms for Hamiltonian Dynamics Mean Value and Guided Local Hamiltonian Problem

The efficient simulation of quantum dynamics and ground states is a central challenge in physics and a key frontier for quantum advantage. While short-time evolution in one-dimensional systems can often be simulated classically, extending this to higher dimensions remains difficult. Here, we introduce an efficient classical algorithm for simulating the short-time dynamics of arbitrary local quantum systems. For any local Hamiltonian $H$ and constant evolution time $t$, our method estimates expectation values of the form $\langleψ|e^{iHt}Oe^{-iHt}|ψ\rangle$ for global Pauli observables $O$ and stabilizer states $|ψ\rangle$, with high precision and exponentially small failure probability. Furthermore, we present a classical dequantization of a tailored quantum algorithm that efficiently solves the guided local Hamiltonian (GLH) problem to constant additive error - previously considered classically hard and hence a promising candidate for quantum computational advantage. These results reveal unexpected classical tractability in constant-time quantum dynamics and fundamental connections between Hamiltonian dynamics mean value and the GLH problem. Our work refines the boundary between classical and quantum computational power, identifying sharper criteria for regimes where quantum advantage may genuinely emerge.

quant-ph↗