arXiv subjects

Zhenyu Zhao

Publications and source records attributed to Zhenyu Zhao.

At least 37 records · Page 2Linked to original sources

arXiv2026

HERO: Improving the Reliability and Sensitivity of Generative Model Evaluation Using Historical Data

Reliable generative AI models critically rely on expert human annotations to evaluate output quality, yet these "gold" labels are expensive to collect and limited in quantity. Organizations thus often turn to collecting vast but noisy "silver" labels from crowdsourced workers or vendor annotators as proxies for gold labels. Because gold remains the evaluation target, naively aggregating noisy silver labels may introduce bias, and estimators built on sparsely observed gold labels may have high variance to resolve the model performance gaps that guide practical decisions. Model evaluation has become an ongoing operational practice rather than a one-time exercise, with evaluation rounds repeating across model versions, releases, and content domains. A natural question is whether the previous historical evaluation data can be used to improve each new round of evaluation. We introduce HERO (History Enhanced RObust model evaluation), a novel framework that uses historical data to suppress bias (improve reliability) and reduce variance (improve sensitivity) in model performance evaluation. HERO calibrates silver labelers' performance learned from historical gold annotations, and stabilizes the resulting estimator by anchoring it to covariate information measured with high precision in the historical data. HERO can be broadly applied across multiple common evaluation tasks, and remains valid when only a subset of historical labelers appears in the current round. We establish conditions under which the bias and variance reductions hold, showcase HERO's performance in simulation studies, and demonstrate its effectiveness on real-world model evaluation benchmarking datasets.

stat.ME↗

arXiv2026

FrameOracle: Learning What to See and How Much to See in Videos

Vision-language models (VLMs) advance video understanding but operate under tight computational budgets, making performance dependent on selecting a small, high-quality subset of frames. Existing frame sampling strategies, such as uniform or fixed-budget selection, fail to adapt to variations in content density or task complexity. To address this, we present FrameOracle, a lightweight, plug-and-play module that predicts both (1) which frames are most relevant to a given query and (2) how many frames are needed. FrameOracle is trained via a curriculum that progresses from weak proxy signals, such as cross-modal similarity, to stronger supervision with FrameOracle-41K, the first large-scale VideoQA dataset with validated keyframe annotations specifying minimal sufficient frames per question. Extensive experiments across five VLMs and six benchmarks show that FrameOracle reduces 16-frame inputs to an average of 10.4 frames without accuracy loss. When starting from 64-frame candidates, it reduces inputs to 13.9 frames on average while improving accuracy by 1.5%, achieving state-of-the-art efficiency-accuracy trade-offs for scalable video understanding.

cs.CV↗

arXiv2026

The Price of Agreement: Measuring LLM Sycophancy in Agentic Financial Applications

Given the increased use of LLMs in financial systems today, it becomes important to evaluate the safety and robustness of such systems. One failure mode that LLMs frequently display in general domain settings is that of sycophancy. That is, models prioritize agreement with expressed user beliefs over correctness, leading to decreased accuracy and trust. In this work, we focus on evaluating sycophancy that LLMs display in agentic financial tasks. Our findings are three-fold: first, we find the models show only low to modest drops in performance in the face of user rebuttals or contradictions to the reference answer, which distinguishes sycophancy that models display in financial agentic settings from findings in prior work. Second, we introduce a suite of tasks to test for sycophancy by user preference information that contradicts the reference answer and find that most models fail in the presence of such inputs. Lastly, we benchmark different modes of recovery such as input filtering with a pretrained LLM.

cs.AI↗

arXiv2026

Approaching physical limits of latent dimensionality in optical computing

The physical implementation of artificial intelligence requires mapping computational processes onto the dynamic physical processes of the underlying computing platform. The photonic processors offer an intrinsically parallel and low energy framework for this mapping, however, a mismatch between the potential computing capability of a bounded optical domain and the human accessible manipulation range sets a hard integration density ceiling on existing architectures. Here, we address this challenge by investigating the integration density limits in photonic processors through exploring the fundamental physical limits on the latent dimensionality for maximum expressivity of a bounded optical domain. These physical limits potentially serve as universal metrics for evaluating optical computing capacity. To validate these, we design and realize ultracompact multimode photonic processors approaching these limits: a 2.2 um by 8 um processor achieves 86.7 % accuracy in experiment for iris flower classification, and a 20.6 um by 44.8 um processor reaches 92.9% accuracy in handwritten digit recognition. Finally, we scale this architecture to highly complex tasks by implementing a generative diffusion model for image synthesis. By grounding photonic processor design in the wave physics origin of latent dimensionality, our results supply the missing theoretical reference point for optical computing architecture.

physics.optics↗

arXiv2026

FASE: FPGA-Assisted Syscall Emulation for Rapid End-to-End Processor Performance Validation

The rapid advancement of AI workloads and domain-specific architectures has led to increasingly diverse processor microarchitectures, whose design exploration requires fast and accurate performance validation. However, traditional workflows defer validation process until RTL design and SoC integration are complete, significantly prolonging development and iteration cycle. In this work, we present FASE framework, FPGA-Assisted Syscall Emulation, the first work for adapt syscall emulation on FPGA platforms, enabling complex multi-thread benchmarks to directly run on the processor design without integrating SoC or target OS for early-stage performance validation. FASE introduces three key innovations to address three critical challenges for adapting FPGA-based syscall emulation: (1) only a minimal CPU interface is exposed, with other hardware components untouched, addressing the lack of a unified hardware interface in FPGA systems; (2) a Host-Target Protocol (HTP) is proposed to minimize cross-device data traffic, mitigating the low-bandwidth and high-latency communication between FPGA and host; and (3) a host-side runtime is proposed to remotely handle Linux-style system calls, addressing the challenge of cross-device syscall delegation. Experiments ware conducted on Xilinx FPGA with open-sourced RISC-V SMP processor Rocket. With single-thread CoreMark, FASE introduces less than 1% performance error and achieves over 2000x higher efficiency compared to Proxy Kernel due to FPGA acceleration. With complex OpenMP benchmarks, FASE demonstrates over 96% performance validation accuracy for most single-thread workloads and over 91.5% for most multi-thread workloads compared to full SoC validation, significantly reducing development complexity and time-to-feedback. All components of FASE framework are released as open-source.

cs.AR↗

arXiv2026

Diffusion Attention Expert Model for Predicting and Semi-automatic Localizing STAS in Lung Cancer Histopathological Images

Accurate intraoperative and postoperative diagnosis of spread through air spaces (STAS) is essential for guiding surgical decisions and postoperative management in lung cancer. However, histopathological assessment is labor-intensive and is prone to missed or incorrect diagnoses. We propose a Diffusion Attention Expert Model (DAEM) to detect STAS in frozen sections (FSs) and paraffin sections (PSs). Its diffusion attention expert module leverages full attention aggregation to learn multi-scale features from histopathological images, while a dual-branch architecture strengthens multi-scale feature representation. On an internal dataset, DAEM achieves AUCs of 0.8946 for FSs and 0.9112 for PSs. Validation on external multi-center datasets from eight institutions demonstrates strong generalizability and interpretability. Using tumor microenvironment (TME) features in PSs, we further enable semi-automatic measurement of STAS location and its distance from the primary tumor. Several quantitative TME metrics are identified as potential biomarkers for STAS, including micropapillary-type STAS. Overall, DAEM offers a clinically actionable framework for STAS assessment by enabling accurate and interpretable detection on FSs and PSs, supporting postoperative risk stratification through quantitative TME-based analysis.

cs.CV↗

arXiv2026

A Patchwise Local Fourier Extension Method for Function Approximation on General Two-Dimensional Domains

We propose a patchwise local Fourier extension method for approximating smooth functions on general two dimensional domains with curved boundaries. The domain is embedded into a Cartesian background grid and decomposed into rectangular interior patches and one-side curved trapezoidal boundary patches. After local data transfer, all patches are converted into fixed-size tensor-product arrays and approximated by a truncated-SVD stabilized local Fourier extension procedure. Unlike global Fourier frame approximations, the proposed method localizes both the geometry and the ill-conditioned extension process. For fixed local parameters, the local algebraic operations are performed on fixed-size systems, and the reference Fourier extension matrices and their singular value decompositions are reused across patches. Boundary patches require additional one-dimensional transfer or completion steps, but their costs remain uniformly bounded by the local resolution. Consequently, the online complexity is $O(N)$, where $N$ denotes the total number of retained output points for fixed local resolution. Numerical experiments on smooth curved domains and on a mildly rough boundary domain demonstrate that the method achieves high accuracy with a fixed set of local parameters. The smooth-cover correction reduces the boundary-induced error by several orders of magnitude in the full-domain rough-boundary test, without changing the underlying scan-based partition.

math.NA↗

arXiv2026

Local Legendre Frame Approximation from Equispaced Data

We propose a local Legendre frame (LLF) method for function approximation from equispaced data on a finite interval. Motivated by the difficulty of stable high-order polynomial approximation at equispaced points, especially in the presence of the Runge phenomenon, the method partitions the interval into subintervals, maps each subinterval to a common reference interval, and computes local coefficients by a truncated singular value decomposition (TSVD) regularization. Since all subintervals share the same local sampling matrix, the method admits a natural offline--online implementation. We establish a quasi-optimal estimate for the regularized reconstruction and discuss practical parameter selection. Numerical results show that LLF attains high accuracy for relatively smooth and moderately oscillatory functions, while it remains applicable to highly oscillatory functions, although comparable accuracy generally requires more sampling points. For continuous piecewise smooth functions with derivative singularities, the method also provides an effective detect--localize--correct strategy based on one-sided coefficient-energy indicators. These results indicate that LLF provides a stable and flexible local approximation framework for equispaced data.

math.NA↗

arXiv2026

NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results

This paper presents a comprehensive review of the NTIRE 2026 3D Restoration and Reconstruction (3DRR) Challenge, detailing the proposed methods and results. The challenge seeks to identify robust reconstruction pipelines that are robust under real-world adverse conditions, specifically extreme low-light and smoke-degraded environments, as captured by our RealX3D benchmark. A total of 279 participants registered for the competition, of whom 33 teams submitted valid results. We thoroughly evaluate the submitted approaches against state-of-the-art baselines, revealing significant progress in 3D reconstruction under adverse conditions. Our analysis highlights shared design principles among top-performing methods and provides insights into effective strategies for handling 3D scene degradation.

cs.CV↗

arXiv2026

High-precision quadrature via local Fourier extension: analytic integration, uniform sampling, and correction for piecewise smooth integrands

We propose a high-precision numerical quadrature framework based on local Fourier extension (LFE) approximations. The method constructs, on each subinterval, a truncated-SVD stabilized local Fourier continuation of the integrand on an extended periodic domain, and then evaluates the integral \emph{analytically} from the resulting Fourier coefficients. Under uniform sampling, the discrete LFE matrix and its TSVD factors are precomputed once and reused across all windows, yielding an efficient offline/online implementation that remains compatible with classical composite rules. We provide an error bound that reduces the quadrature error to the LFE approximation error and derive algebraic convergence rates for Sobolev-regular integrands. Numerical experiments demonstrate that, on smooth functions, the proposed quadrature reaches near machine precision with substantially fewer nodes than the composite Simpson rule. The advantage persists for oscillatory and variable-frequency integrands and becomes more pronounced for nonuniform phase structures. For continuous piecewise smooth integrands, we develop a correction strategy driven by coefficient-energy outliers to identify singularity-containing windows, followed by a localized procedure that brackets the singular point within one grid cell and corrects only the affected window contribution. The corrected quadrature restores near-spectral accuracy in the reported tests, including cases where the singularity is not aligned with the window endpoints.

math.NA↗

arXiv2026

$Ψ_0$: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation

We introduce $Ψ_0$ (Psi-Zero), an open foundation model to address challenging humanoid loco-manipulation tasks. While existing approaches often attempt to address this fundamental problem by co-training on large and diverse human and humanoid data, we argue that this strategy is suboptimal due to the fundamental kinematic and motion disparities between humans and humanoid robots. Therefore, data efficiency and model performance remain unsatisfactory despite the considerable data volume. To address this challenge, \ours\;decouples the learning process to maximize the utility of heterogeneous data sources. Specifically, we propose a staged training paradigm with different learning objectives: First, we autoregressively pre-train a VLM backbone on large-scale egocentric human videos to acquire generalizable visual-action representations. Then, we post-train a flow-based action expert on high-quality humanoid robot data to learn precise robot joint control. Our research further identifies a critical yet often overlooked data recipe: in contrast to approaches that scale with noisy Internet clips or heterogeneous cross-embodiment robot datasets, we demonstrate that pre-training on high-quality egocentric human manipulation data followed by post-training on domain-specific real-world humanoid trajectories yields superior performance. Extensive real-world experiments demonstrate that \ours\ achieves the best performance using only about 800 hours of human video data and 30 hours of real-world robot data, outperforming baselines pre-trained on more than 10$\times$ as much data by over 40\% in overall success rate across multiple tasks. We will open-source the entire ecosystem to the community, including a data processing and training pipeline, a humanoid foundation model, and a real-time action inference engine.

cs.RO↗

arXiv2026

Dynamic Prior Thompson Sampling for Cold-Start Exploration in Recommender Systems

Cold-start exploration is a core challenge in large-scale recommender systems: new or data-sparse items must receive traffic to estimate value, but over-exploration harms users and wastes impressions. In practice, Thompson Sampling (TS) is often initialized with a uniform Beta(1,1) prior, implicitly assuming a 50% success rate for unseen items. When true base rates are far lower, this optimistic prior systematically over-allocates to weak items. The impact is amplified by batched policy updates and pipeline latency: for hours, newly launched items can remain effectively "no data," so the prior dominates allocation before feedback is incorporated. We propose Dynamic Prior Thompson Sampling, a prior design that directly controls the probability that a new arm outcompetes the incumbent winner. Our key contribution is a closed-form quadratic solution for the prior mean that enforces P(X_j > Y_k) = epsilon at introduction time, making exploration intensity predictable and tunable while preserving TS Bayesian updates. Across Monte Carlo validation, offline batched simulations, and a large-scale online experiment on a thumbnail personalization system serving millions of users, dynamic priors deliver precise exploration control and improved efficiency versus a uniform-prior baseline.

cs.LG↗

arXiv2026

Multipath Routing for Multi-Hop UAV Networks

Multi-hop uncrewed aerial vehicle (UAV) networks are promising to extend the terrestrial network coverage. Existing multi-hop UAV networks employ a single routing path by selecting the next-hop forwarding node in a hop-by-hop manner, which leads to local congestion and increases traffic delays. In this paper, a novel traffic-adaptive multipath routing method is proposed for multi-hop UAV networks, which enables each UAV to dynamically split and forward traffic flows across multiple next-hop neighbors, thus meeting latency requirements of diverse traffic flows in dynamic mobile environments. An on-time packet delivery ratio maximization problem is formulated to determine the traffic splitting ratios at each hop. This sequential decision-making problem is modeled as a decentralized partially observable Markov decision process (Dec-POMDP). To solve this Dec-POMDP, a novel multi-agent deep reinforcement leaning (MADRL) algorithm, termed Independent Proximal Policy Optimization with Dirichlet Modeling (IPPO-DM), is developed. Specifically, the IPPO serves as the core optimization framework, where the Dirichlet distribution is leveraged to parameterize a continuous stochastic policy network on the probability simplex, inherently ensuring feasible traffic splitting ratios. Simulation results demonstrate that IPPO-DM outperforms benchmark schemes in terms of both delivery latency guarantee and packet loss performance.

cs.MA↗

arXiv2026

Fourier Extension Based on Weighted Generalized Inverse

This paper introduces a weighted generalized inverse framework for Fourier extensions, designed to suppress spurious oscillations in the extended region while maintaining high approximation accuracy on the original interval. By formulating the Fourier extension problem as a compact operator equation, we propose a weighted best-approximation solution that incorporates a priori smoothness information through suitable weight operators on the Fourier coefficients. This leads to a regularization scheme based on the generalized truncated singular value decomposition (GTSVD). Under algebraic and exponential smoothness assumptions, convergence analysis demonstrates optimal $L^2$ accuracy and improved stability for derivatives. Compared with classical Fourier extension using standard TSVD, the proposed method effectively controls high-frequency components and yields smoother extensions. A practical discretization using uniform sampling is developed, along with an adaptive design of weight functions. Numerical experiments confirm that the method significantly improves derivative approximations and reduces oscillations in the extended domain without compromising accuracy on the original interval.

math.NA↗

arXiv2025

CoPHo: Classifier-guided Conditional Topology Generation with Persistent Homology

The structure of topology underpins much of the research on performance and robustness, yet available topology data are typically scarce, necessitating the generation of synthetic graphs with desired properties for testing or release. Prior diffusion-based approaches either embed conditions into the diffusion model, requiring retraining for each attribute and hindering real-time applicability, or use classifier-based guidance post-training, which does not account for topology scale and practical constraints. In this paper, we show from a discrete perspective that gradients from a pre-trained graph-level classifier can be incorporated into the discrete reverse diffusion posterior to steer generation toward specified structural properties. Based on this insight, we propose Classifier-guided Conditional Topology Generation with Persistent Homology (CoPHo), which builds a persistent homology filtration over intermediate graphs and interprets features as guidance signals that steer generation toward the desired properties at each denoising step. Experiments on four generic/network datasets demonstrate that CoPHo outperforms existing methods at matching target metrics, and we further validate its transferability on the QM9 molecular dataset.

cs.LG↗

arXiv2025

Deep Photonic Reservoir Computing with On-chip Nonlinearity

Reservoir computing, renowned for its low training cost, has emerged as a promising lightweight paradigm for efficient spatiotemporal processing,it remains challenging to realize deep photonic reservoir computing (DPRC) systems, due to the lack of scalable on-chip nonlinearity. Here, we introduce a versatile time delayed DPRC framework that natively supports deep and concurrent spatiotemporal processing entirely in the optical domain. At its core, the system leverages free carrier dynamics in silicon microring resonators to provide the fundamental nonlinearity and short term memory, and these nonlinear nodes are interconnected through true time delay lines that establish shared long-term memory. Benefiting from intrinsic physical nonlinearity and multi-timescale fading memory, this simple yet effective architecture demonstrates remarkable high dimensional representation capabilities. On the NTU RGB D benchmark, the parameter efficient DPRC system achieves superior action recognition accuracies compared to mainstream deep learning models, while requiring only a single shot regression training procedure. We further verify a prototype DPRC chip that excels across diverse dataset classification and time series prediction tasks. It enables a streamlined all optical pipeline between hierarchical layers, delivering a consistent computational density of 334.25 TOPs/mm2, independent of the reservoir depth and three orders of magnitude higher than conventional approaches. Moreover, its performance scales with near-zero hardware overhead by utilizing additional wavelength channels. This DPRC network is highly scalable on a silicon photonic platform, with flexible extension to hundreds of deep reservoir layers and parallel channels, paving the way toward intelligent optoelectronic systems for advanced real time processing and parallel decision making.

physics.optics↗

arXiv2025

Causal Feature Selection Method for Contextual Multi-Armed Bandits in Recommender System

Effective feature selection is essential for optimizing contextual multi-armed bandits (CMABs) in large-scale online systems, where suboptimal features can degrade rewards, interpretability, and efficiency. Traditional feature selection often prioritizes outcome correlation, neglecting the crucial role of heterogeneous treatment effects (HTE) across arms in CMAB decision-making. This paper introduces two novel, model-free filter methods, Heterogeneous Incremental Effect (HIE) and Heterogeneous Distribution Divergence (HDD), specifically designed to identify features driving HTE. HIE quantifies a feature's value based on its ability to induce changes in the optimal arm, while HDD measures its impact on reward distribution divergence across arms. These methods are computationally efficient, robust to model mis-specification, and adaptable to various feature types, making them suitable for rapid screening in dynamic environments where retraining complex models is infeasible. We validate HIE and HDD on synthetic data with known ground truth and in a large-scale commercial recommender system, demonstrating their consistent ability to identify influential HTE features and thereby enhance CMAB performance.

cs.LG↗

arXiv2025

Dependency-Aware Task Offloading in Multi-UAV Assisted Collaborative Mobile Edge Computing

This paper proposes a novel multi-unmanned aerial vehicle (UAV) assisted collaborative mobile edge computing (MEC) framework, where the computing tasks of terminal devices (TDs) can be decomposed into serial or parallel sub-tasks and offloaded to collaborative UAVs. We first model the dependencies among all sub-tasks as a directed acyclic graph (DAG) and design a two-timescale frame structure to decouple the sub-task interdependencies for sub-task scheduling. Then, a joint sub-task offloading, computational resource allocation, and UAV trajectories optimization problem is formulated, which aims to minimize the system cost, i.e., the weighted sum of the task completion delay and the system energy consumption. To solve this non-convex mixed-integer nonlinear programming (MINLP) problem, a penalty dual decomposition and successive convex approximation (PDD-SCA) algorithm is developed. Particularly, the original MINLP problem is equivalently transferred into a continuous form relying on PDD theory. By decoupling the resulting problem into three nested subproblems, the SCA method is further combined to recast the non-convex components and obtain desirable solutions. Numerical results demonstrate that: 1) Compared to the benchmark algorithms, the proposed scheme can significantly reduce the system cost, and thus realize an improved trade-off between task latency and energy consumption; 2) The proposed algorithm can achieve an efficient workload balancing for distributed computation across multiple UAVs.

cs.CY↗