arXiv subjects

Zhaohui Xu

Publications and source records attributed to Zhaohui Xu.

9 recordsLinked to original sources

arXiv2026

Continual Learning in Transition

Classical continual learning (CL) has primarily focused on enabling models to update and retain knowledge through parameter-centric mechanisms, e.g., training strategies, architectural designs, and weight adaptation. However, emerging paradigms are reshaping the scope of CL beyond this traditional model adaptation view. For instance, on-policy learning broadens the space of update mechanisms; test-time training extends CL from the training phase to inference; and external harness components such as memory, skill libraries, and interaction protocols extend the evolutionary boundaries of model capabilities far beyond the static parameter space. Collectively, these developments indicate a transition from parameter-centric learning toward system-level adaptation. To characterize this transition, we examine the evolution of continual learning through three dimensions: When, How, and Where learning occurs. The How dimension encompasses off-policy, on-policy, and beyond-gradient optimization mechanics. The When dimension captures evolution across pre-training, post-training, and inference-time stages. The Where dimension delineates updates occurring within internal parameters versus external structural constraints. Anchored by this tri-axial framework, we systematically survey representative methods, trace the ongoing transition of continual learning, and discuss the key challenges, broader implications, and future directions arising from this paradigm shift.

cs.LG↗

arXiv2026

Extra-dimensional Origins of Chemical Potentials at the Cosmological Collider

We study the realization of the chemical potential mechanism in cosmological collider physics in a robust class of inflationary models arising from multiple higher-dimensional gauge fields. The rolling inflaton background corresponds to an electric field in the extra dimension in which charged particles are produced analogously to the Schwinger mechanism, while neutral particles can be produced through non-minimal interactions. We show that particles heavier than the inflationary Hubble scale can be created without Boltzmann suppression in both cases. In particular, charged Kaluza-Klein excitations can be created through minimal gauge interactions. We construct realistic models along these lines consistent with both theoretical and experimental constraints.

hep-ph↗

arXiv2026

Robust Subgroup Analysis for Heterogeneous Censored Data

Subgroup analysis is important in practice because real-world data typically come from heterogeneous populations, where meaningful patterns can differ substantially across subpopulations. Correctly identifying these subgroups can improve prediction accuracy, prevent biased or misleading conclusions, and support more effective, targeted decision-making. While most existing subgroup analysis methods are developed for complete data, in this paper we propose a novel and robust approach for censored data under heterogeneous accelerated failure time (AFT) models. Specifically, we combine inverse probability weighting, M-estimation, and concave pairwise fusion penalization to simultaneously identify subgroups and estimate covariate effects for heterogeneous censored data, without requiring prior knowledge of individual subgroup memberships. We further develop an efficient RISA-ADMM algorithm to implement the method and establish its convergence. Furthermore, we derive the theoretical properties of the proposed estimators under mild regularity conditions. Extensive simulations and an application to the German credit dataset demonstrate the robustness and effectiveness of our approach.

stat.ME↗

arXiv2026

YouZhi: Towards High-Concurrency Financial LLMs via Adaptive GQA-to-MLA Transition

Large language models (LLMs) drive significant financial innovations, yet their high-concurrency deployment is severely bottlenecked by KV cache memory overhead, which inflates infrastructure costs and throttles scalability. To address this, we propose YouZhi-LLM, a highly efficient financial LLM empowered by a comprehensive structural transition and training pipeline natively built on the Huawei Ascend ecosystem. At its algorithmic core, YouZhi-LLM features a layer-adaptive GQA-to-MLA transition framework that dynamically assigns per-layer FreqFold sizes, maximizing KV-cache compression while minimizing perplexity degradation. To recover representation capacity and inject domain expertise, the Ascend-based training pipeline seamlessly integrates generalized knowledge distillation with financial-specific supervised fine-tuning. Evaluations demonstrate the superiority of this systematic approach, with the adaptive transition reducing perplexity degradation by up to 35% over uniform baselines. Crucially, when evaluated on Ascend NPUs via vLLM-Ascend, the massive KV-cache reduction translates directly into deployment efficiency. Compared to their respective base models, YouZhi-7B yields a 12.3% improvement in average financial benchmark score alongside a 2.69$\times$ increase in maximum concurrency; similarly, YouZhi-14B achieves a 7.0% accuracy gain and a 2.43$\times$ concurrency boost, establishing a new paradigm for cost-effective, high-throughput financial inference.

cs.CL↗

arXiv2026

Charged Loops at the Cosmological Collider with Chemical Potential

Cosmological collider physics allows the detection of heavy particles at inflationary scales through their imprints on primordial non-Gaussianities. We study the chemical potential mechanism applied to a pair of charged scalars. We analytically evaluate the resulting one-loop contribution to the bispectrum, using the spectral decomposition. In this way we are able to determine the parametric dependences for both the signal and the background. We show that a signal strength $f_{\mathrm{NL}}\sim O(0.01)$ can be obtained within theoretical control, potentially reachable by 21cm tomography. As an application we consider the colored Higgs bosons in $\mathrm{SU}(5)$ supersymmetric orbifold grand unification with masses $M\lesssim10^{15}\:\mathrm{GeV}$.

hep-ph↗

arXiv2025

Make LLM Inference Affordable to Everyone: Augmenting GPU Memory with NDP-DIMM

The billion-scale Large Language Models (LLMs) need deployment on expensive server-grade GPUs with large-storage HBMs and abundant computation capability. As LLM-assisted services become popular, achieving cost-effective LLM inference on budget-friendly hardware becomes the trend. Extensive researches relocate LLM parameters from expensive GPUs to host memory. However, the restricted bandwidth between the host and GPU memory limits the inference performance. This work introduces Hermes, a budget-friendly system that leverages the near-data processing (NDP) within commodity DRAM DIMMs to enhance the performance of a single consumer-grade GPU, achieving efficient LLM inference. The inherent activation sparsity in LLMs naturally divides weight parameters into two categories, termed ``hot" and ``cold" neurons, respectively. Hot neurons, which consist of only approximately 20\% of all weight parameters, account for 80\% of the total computational load, while cold neurons make up the other 80\% of parameters but are responsible for just 20\% of the computational load. Therefore, we propose a heterogeneous computing strategy: mapping hot neurons to a single computation-efficient GPU, while offloading cold neurons to NDP-DIMMs, which offer large memory size but limited computation capabilities. Meanwhile, the dynamic nature of activation sparsity needs a real-time partition of hot/cold neurons and adaptive remapping of cold neurons across multiple NDP-DIMM modules. Therefore, we introduce a lightweight predictor optimizing real-time neuron partition and adjustment between GPU and NDP-DIMMs. We also utilize a window-based online scheduling mechanism to maintain load balance among NDP-DIMM modules. Hermes facilitates the deployment of LLaMA2-70B on consumer-grade hardware at 13.75 tokens/s and realizes an average 75.24$\times$ speedup over the state-of-the-art offloading-based inference system.

cs.AR↗

arXiv2024

COMET: Towards Partical W4A4KV4 LLMs Serving

Quantization is a widely-used compression technology to reduce the overhead of serving large language models (LLMs) on terminal devices and in cloud data centers. However, prevalent quantization methods, such as 8-bit weight-activation or 4-bit weight-only quantization, achieve limited performance improvements due to poor support for low-precision (e.g., 4-bit) activation. This work, for the first time, realizes practical W4A4KV4 serving for LLMs, fully utilizing the INT4 tensor cores on modern GPUs and reducing the memory bottleneck caused by the KV cache. Specifically, we propose a novel fine-grained mixed-precision quantization algorithm (FMPQ) that compresses most activations into 4-bit with negligible accuracy loss. To support mixed-precision matrix multiplication for W4A4 and W4A8, we develop a highly optimized W4Ax kernel. Our approach introduces a novel mixed-precision data layout to facilitate access and fast dequantization for activation and weight tensors, utilizing the GPU's software pipeline to hide the overhead of data loading and conversion. Additionally, we propose fine-grained streaming multiprocessor (SM) scheduling to achieve load balance across different SMs. We integrate the optimized W4Ax kernel into our inference framework, COMET, and provide efficient management to support popular LLMs such as LLaMA-3-70B. Extensive evaluations demonstrate that, when running LLaMA family models on a single A100-80G-SMX4, COMET achieves a kernel-level speedup of \textbf{$2.88\times$} over cuBLAS and a \textbf{$2.02 \times$} throughput improvement compared to TensorRT-LLM from an end-to-end framework perspective.

cs.AR↗

arXiv2012

The fundamental Diagram of Pedestrian Model with Slow Reaction

The slow-to-start models are a classical cellular automata model in simulating vehicle traffic. However, to our knowledge, the slow-to-start effect has not considered in modeling pedestrian dynamic. We verify the similar behavior between pedestrian and vehicle, and propose an new lattice gas (LG) model called the slow reaction (SR) model to describe the pedestrian's delayed reaction in single-file movement. We simulate and reproduce the Seyfried's field experiments at the research centre Julich, and use its empirical data to validate our SR model. We compare the SR model with the standard LG model. We test different probability of slow reaction ps in SR model and found the simulation data of ps=0.3 fit the empirical data best. The RMS error of mean velocity of SR model is smaller than that of standard LG model. In the range of ps=0.1~0.3, our fundamental diagram between velocity and density by simulation coincides with field experiments. The distribution of individual velocity in fundamental diagram in SR model agrees with the empirical data better than that of standard LG model. In addition, we observe the stop-and-go waves and phase separation in pedestrian flow by simulation. We reproduced the phenomena of uneven distribution of interspaces by SR model while the standard LG model did not implement. The SR model can reproduce the evolution of spatio-temporal structures of pedestrian flow with higher fidelity to Seyfried's experiments than the standard LG model.

physics.soc-ph↗

arXiv2012

Jamming Transition of Point-to-Point Traffic Through Cooperative Mechanisms

We study the jamming transition of two-dimensional point-to-point traffic through cooperative mechanisms using computer simulation. We propose two decentralized cooperative mechanisms which are incorporated into the point-to-point traffic models: stepping aside (CM-SA) and choosing alternative routes (CM-CAR). Incorporating CM-SA is to prevent a type of ping-pong jumps from happening when two objects standing face-to-face want to move in opposite directions. Incorporating CM-CAR is to handle the conflict when more than one object competes for the same point in parallel update. We investigate and compare four models mainly from fundamental diagrams, jam patterns and the distribution of cooperation probability. It is found that although it decreases the average velocity a little, the CM-SA increases the critical density and the average flow. Despite increasing the average velocity, the CM-CAR decreases the average flow by creating substantially vacant areas inside jam clusters. We investigate the jam patterns of four models carefully and explain this result qualitatively. In addition, we discuss the advantage and applicability of decentralized cooperation modeling.

physics.soc-ph↗