arXiv subjects

Shiyong Li

Publications and source records attributed to Shiyong Li.

At least 19 recordsLinked to original sources

arXiv2026

CVPO: Enhancing LLM Reinforcement Learning Reasoning via Value-Variance Adaptation and Dynamic Curriculum Learning

Reinforcement learning (RL) has emerged as an effective method for enhancing the reasoning capabilities of large language models (LLMs). However, existing methods suffer from insufficient precision in feedback on generated answer trajectories and exhibit the phenomenon of problem difficulty drift. To address these challenges, we propose CVPO - Curriculum-guided Value-Variance Policy Optimization. At the response trajectory level, we find that token-level value-variance correlates with exploration intensity. Our theoretical analysis shows this variance bounds policy update magnitude. We then use the estimated trajectory value-variance to quantify the intrinsic randomness in generation. Based on this, we design a variance-aware advantage adjustment mechanism for different reward types. At the question level, we introduce a dynamic curriculum weighting method that adapts to question difficulty. This helps the model focus on tasks matched to its current ability during each training stage. Experimental results show our method outperforms strong value-based baselines like VAPO. It achieves better performance and stronger exploration, enabling more accurate and robust reasoning in language models across various math tasks.

cs.CL↗

arXiv2026

TAOT: Topology-Aware Optimal Transport for Dynamic Expert Replica Placement in MoE Training

Mixture-of-Experts (MoE) has become a key architecture for scaling large language models (LLMs), yet its dynamic routing causes severe load imbalance in expert-parallel training. Existing dynamic-replica methods copy hot experts onto idle ranks to share computation, but they optimize load balance alone and ignore the cost of moving expert weights across a multi-node topology, so the resulting cross-node communication can outweigh the balancing gain and inflate training cost. We present TAOT, a topology-aware optimal transport method for dynamic expert-replica placement. TAOT models the overload on hot ranks and the spare capacity on lightly loaded ranks as a balanced entropy-regularized optimal transport problem with a communication-cost matrix, solves it with Sinkhorn-Knopp iterations to produce rank-level flow hints, and combines integer replica matching with token assignment into an executable schedule. At the system level, it overlaps guest-weight transfer with home-expert computation to hide the communication overhead. Experiments show TAOT achieves a 1.43x end-to-end MoE training speedup, reaches balance quality competitive with or better than existing state-of-the-art methods, and attains the lowest weighted expert-communication cost across all configurations, with up to a 74% reduction.

cs.DC↗

arXiv2026

Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction

Given the quadratic complexity of attention, KV cache eviction is vital to accelerate model inference. Current KV cache eviction methods typically rely on instantaneous heuristic metrics, implicitly assuming that score magnitudes are consistent proxies for importance across all heads. However, this overlooks the heterogeneity in predictive fidelity across attention heads. While certain heads prioritize the instantaneous contribution of tokens, others are dedicated to capturing long-horizon utility. In this paper, we propose that optimal budget allocation should be governed by the marginal utility in preserving long-term semantic information. Building on this insight, we propose LU-KV, a novel framework that formulates head-level budget allocation as a global combinatorial optimization problem to maximize the long-horizon marginal contribution of reserved tokens. To solve this non-convex problem, we employ a convex-hull relaxation and a marginal-utility-based greedy solver, achieving near-optimal solutions. Furthermore, we implement a data-driven offline profiling protocol to facilitate the practical deployment of LU-KV. Evaluations on LongBench and RULER benchmarks demonstrate that LU-KV reduces KV cache size by 80% with minimal performance degradation, while also decreasing inference latency and GPU memory footprint.

cs.LG↗

arXiv2025

ESS: An Offload-Centric Latent-Cache Management Architecture for DeepSeek-V3.2-Exp

DeepSeek-V3.2-Exp introduces a sparse attention mechanism that significantly reduces inference latency in long-context scenarios. Although the overall throughput has improved greatly, the Decode-stage of PD disaggregation remains to be a major bottleneck. This bottleneck primarily stems from the conflict between linear growth of Latent-Cache with sequence length and the limited GPU memory capacity, which constrains the feasible batch-size and thereby suppresses Decode-stage throughput. To address this challenge, we propose ESS (Extended Sparse Server), an offload-centric system design tailored for DeepSeek-V3.2-Exp. ESS selectively offloads Latent-Cache to CPU memory while preserving latency-critical components on GPU. By freeing up GPU memory, ESS effectively decoupling batch-size scaling from GPU memory constraints. This design significantly improves Decode-stage throughput, thereby reducing deployment costs in real-world settings. Our high-fidelity simulations show that ESS delivers 69.4\% throughput improvement at 32K context length and up to 123\% throughput improvement at 128K, demonstrating its effectiveness for large-context inference workloads. These results highlight ESS as a practical and scalable solution for long-context LLM serving.

cs.DC↗

arXiv2025

FEVO: Financial Knowledge Expansion and Reasoning Evolution for Large Language Models

Advancements in reasoning for large language models (LLMs) have lead to significant performance improvements for LLMs in various fields such as mathematics and programming. However, research applying these advances to the financial domain, where considerable domain-specific knowledge is necessary to complete tasks, remains limited. To address this gap, we introduce FEVO (Financial Evolution), a multi-stage enhancement framework developed to enhance LLM performance in the financial domain. FEVO systemically enhances LLM performance by using continued pre-training (CPT) to expand financial domain knowledge, supervised fine-tuning (SFT) to instill structured, elaborate reasoning patterns, and reinforcement learning (RL) to further integrate the expanded financial domain knowledge with the learned structured reasoning. To ensure effective and efficient training, we leverage frontier reasoning models and rule-based filtering to curate FEVO-Train, high-quality datasets specifically designed for the different post-training phases. Using our framework, we train the FEVO series of models - C32B, S32B, R32B - from Qwen2.5-32B and evaluate them on seven benchmarks to assess financial and general capabilities, with results showing that FEVO-R32B achieves state-of-the-art performance on five financial benchmarks against much larger models as well as specialist models. More significantly, FEVO-R32B demonstrates markedly better performance than FEVO-R32B-0 (trained from Qwen2.5-32B-Instruct using only RL), thus validating the effectiveness of financial domain knowledge expansion and structured, logical reasoning distillation

cs.AI↗

arXiv2025

Astra: Efficient and Money-saving Automatic Parallel Strategies Search on Heterogeneous GPUs

In this paper, we introduce an efficient and money-saving automatic parallel strategies search framework on heterogeneous GPUs: Astra. First, Astra searches for the efficiency-optimal parallel strategy in both GPU configurations search space (GPU types and GPU numbers) and parallel parameters search space. Then, Astra also provides the solution on heterogeneous GPUs by mathematically modeling the time consumption of heterogeneous training. At last, Astra is the first to propose the automatic parallel strategy search on money-saving. The experiment results demonstrate that Astra can achieve better throughput than expert-designed strategies. The search time cost for Astra can also be limited to 1.27 seconds in a single-GPU setting and less than 1.35 minutes in a heterogeneous-GPU setting on average with an accuracy of over 95%.

cs.DC↗

arXiv2024

Yi-Lightning Technical Report

This technical report presents Yi-Lightning, our latest flagship large language model (LLM). It achieves exceptional performance, ranking 6th overall on Chatbot Arena, with particularly strong results (2nd to 4th place) in specialized categories including Chinese, Math, Coding, and Hard Prompts. Yi-Lightning leverages an enhanced Mixture-of-Experts (MoE) architecture, featuring advanced expert segmentation and routing mechanisms coupled with optimized KV-caching techniques. Our development process encompasses comprehensive pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF), where we devise deliberate strategies for multi-stage training, synthetic data construction, and reward modeling. Furthermore, we implement RAISE (Responsible AI Safety Engine), a four-component framework to address safety issues across pre-training, post-training, and serving phases. Empowered by our scalable super-computing infrastructure, all these innovations substantially reduce training, deployment and inference costs while maintaining high-performance standards. With further evaluations on public academic benchmarks, Yi-Lightning demonstrates competitive performance against top-tier LLMs, while we observe a notable disparity between traditional, static benchmark results and real-world, dynamic human preferences. This observation prompts a critical reassessment of conventional benchmarks' utility in guiding the development of more intelligent and powerful AI systems for practical applications. Yi-Lightning is now available through our developer platform at https://platform.lingyiwanwu.com.

cs.CL↗

arXiv2024

Mao: Machine learning approach for NUMA optimization in Warehouse Scale Computers

Non-Uniform Memory Access (NUMA) architecture imposes numerous performance challenges to today's cloud workloads. Due to the complexity and the massive scale of modern warehouse-scale computers (WSCs), a lot of efforts need to be done to improve the memory access locality on the NUMA architecture. In Baidu, we have found that NUMA optimization has significant performance benefit to the major workloads like Search and Feed (Baidu's recommendation system). But how to conduct NUMA optimization within the large scale cluster brings a lot of subtle complexities and workload-specific scenario optimizations. In this paper, we will present a production environment deployed solution in Baidu called MAP (Memory Access Optimizer) that helps improve the memory access locality for Baidu's various workloads. MAO includes an online module and an offline module. The online module is responsible for the online monitoring, dynamic NUMA node binding and runtime optimization. Meanwhile the offline workload characterization module will proceed with the data analysis and resource-sensitivity module training. We also proposed a new performance model called "NUMA Sensitivity model" to address the impact of remote memory access to workload performance and projection of the potential performance improvements via NUMA optimization for a specific workload. Based on continuous data collected from online monitoring, this model is proved to be working properly in MAO. As of today, we have successfully deployed MAO to more than one hundred thousand servers. In our Feed product, we have achieved 12.1% average latency improvements and 9.8% CPU resource saving.

cs.DC↗

arXiv2023

Towards Large-scale Single-shot Millimeter-wave Imaging for Low-cost Security Inspection

Millimeter-wave (MMW) imaging is emerging as a promising technique for safe security inspection. It achieves a delicate balance between imaging resolution, penetrability and human safety, resulting in higher resolution compared to low-frequency microwave, stronger penetrability compared to visible light, and stronger safety compared to X ray. Despite of recent advance in the last decades, the high cost of requisite large-scale antenna array hinders widespread adoption of MMW imaging in practice. To tackle this challenge, we report a large-scale single-shot MMW imaging framework using sparse antenna array, achieving low-cost but high-fidelity security inspection under an interpretable learning scheme. We first collected extensive full-sampled MMW echoes to study the statistical ranking of each element in the large-scale array. These elements are then sampled based on the ranking, building the experimentally optimal sparse sampling strategy that reduces the cost of antenna array by up to one order of magnitude. Additionally, we derived an untrained interpretable learning scheme, which realizes robust and accurate image reconstruction from sparsely sampled echoes. Last, we developed a neural network for automatic object detection, and experimentally demonstrated successful detection of concealed centimeter-sized targets using 10% sparse array, whereas all the other contemporary approaches failed at the same sample sampling ratio. The performance of the reported technique presents higher than 50% superiority over the existing MMW imaging schemes on various metrics including precision, recall, and mAP50. With such strong detection ability and order-of-magnitude cost reduction, we anticipate that this technique provides a practical way for large-scale single-shot MMW imaging, and could advocate its further practical applications.

eess.IV↗

arXiv2023

Population-Based Evolutionary Gaming for Unsupervised Person Re-identification

Unsupervised person re-identification has achieved great success through the self-improvement of individual neural networks. However, limited by the lack of diversity of discriminant information, a single network has difficulty learning sufficient discrimination ability by itself under unsupervised conditions. To address this limit, we develop a population-based evolutionary gaming (PEG) framework in which a population of diverse neural networks is trained concurrently through selection, reproduction, mutation, and population mutual learning iteratively. Specifically, the selection of networks to preserve is modeled as a cooperative game and solved by the best-response dynamics, then the reproduction and mutation are implemented by cloning and fluctuating hyper-parameters of networks to learn more diversity, and population mutual learning improves the discrimination of networks by knowledge distillation from each other within the population. In addition, we propose a cross-reference scatter (CRS) to approximately evaluate re-ID models without labeled samples and adopt it as the criterion of network selection in PEG. CRS measures a model's performance by indirectly estimating the accuracy of its predicted pseudo-labels according to the cohesion and separation of the feature space. Extensive experiments demonstrate that (1) CRS approximately measures the performance of models without labeled samples; (2) and PEG produces new state-of-the-art accuracy for person re-identification, indicating the great potential of population-based network cooperative training for unsupervised learning.

cs.CV↗

arXiv2022

AcroFOD: An Adaptive Method for Cross-domain Few-shot Object Detection

Under the domain shift, cross-domain few-shot object detection aims to adapt object detectors in the target domain with a few annotated target data. There exists two significant challenges: (1) Highly insufficient target domain data; (2) Potential over-adaptation and misleading caused by inappropriately amplified target samples without any restriction. To address these challenges, we propose an adaptive method consisting of two parts. First, we propose an adaptive optimization strategy to select augmented data similar to target samples rather than blindly increasing the amount. Specifically, we filter the augmented candidates which significantly deviate from the target feature distribution in the very beginning. Second, to further relieve the data limitation, we propose the multi-level domain-aware data augmentation to increase the diversity and rationality of augmented data, which exploits the cross-image foreground-background mixture. Experiments show that the proposed method achieves state-of-the-art performance on multiple benchmarks.

cs.CV↗

arXiv2022

Compressive Sensing Based Sparse MIMO Array Optimization for Wideband Near-Field Imaging

In the area of near-field millimeter-wave imaging, the generalized sparse array synthesis (SAS) method is in great demand. The traditional methods usually employ the greedy algorithms, which may have the convergence problem. This paper proposes a convex optimization model for the multiple-input multiple-output (MIMO) array design based on the compressive sensing (CS) approach. We generate a block shaped reference pattern, to be used as an optimizing target. The pattern occupies the entire imaging area of interest in order to involve the effect of each pixel into the optimization model. In MIMO scenarios, we can fix the transmit subarray and synthesize the receive subarray, and vice versa, or doing the synthesis sequentially. The problems associated with focusing, sidelobes suppression, and grating lobes suppression of the synthesized array are examined in details. Numerical and experimental results demonstrate that the synthesized sparse array can offer better image qualities than the sparse arrays with equally spaced or randomly spaced antennas with the same number of antenna elements.

eess.SP↗

arXiv2021

ERNIE 3.0 Titan: Exploring Larger-scale Knowledge Enhanced Pre-training for Language Understanding and Generation

Pre-trained language models have achieved state-of-the-art results in various Natural Language Processing (NLP) tasks. GPT-3 has shown that scaling up pre-trained language models can further exploit their enormous potential. A unified framework named ERNIE 3.0 was recently proposed for pre-training large-scale knowledge enhanced models and trained a model with 10 billion parameters. ERNIE 3.0 outperformed the state-of-the-art models on various NLP tasks. In order to explore the performance of scaling up ERNIE 3.0, we train a hundred-billion-parameter model called ERNIE 3.0 Titan with up to 260 billion parameters on the PaddlePaddle platform. Furthermore, we design a self-supervised adversarial loss and a controllable language modeling loss to make ERNIE 3.0 Titan generate credible and controllable texts. To reduce the computation overhead and carbon emission, we propose an online distillation framework for ERNIE 3.0 Titan, where the teacher model will teach students and train itself simultaneously. ERNIE 3.0 Titan is the largest Chinese dense pre-trained model so far. Empirical results show that the ERNIE 3.0 Titan outperforms the state-of-the-art models on 68 NLP datasets.

cs.CL↗

arXiv2021

The BEST framework for the search for the QCD critical point and the chiral magnetic effect

The Beam Energy Scan Theory (BEST) Collaboration was formed with the goal of providing a theoretical framework for analyzing data from the Beam Energy Scan (BES) program at the relativistic heavy ion collider (RHIC) at Brookhaven National Laboratory. The physics goal of the BES program is the search for a conjectured QCD critical point as well as for manifestations of the chiral magnetic effect. We describe progress that has been made over the previous five years. This includes studies of the equation of state and equilibrium susceptibilities, the development of suitable initial state models, progress in constructing a hydrodynamic framework that includes fluctuations and anomalous transport effects, as well as the development of freezeout prescriptions and hadronic transport models. Finally, we address the challenge of integrating these components into a complete analysis framework. This document describes the collective effort of the BEST Collaboration and its collaborators around the world.

nucl-th↗

arXiv2021

Near-Field Millimeter-Wave Imaging via Arrays in the Shape of Polyline

This paper proposes a polyline shaped array based system scheme, associated with mechanical scanning along the perpendicular direction of the array, for near-field millimeter-wave (MMW) imaging. Each section of the polyline is a chord of a circle with equal length. The polyline array, which can be realized as a monostatic array or a multistatic one, is capable of providing more observation angles than the linear or planar arrays. Further, we present the related three-dimensional (3-D) imaging algorithms based on a hybrid processing in the time domain and the spatial frequency domain. The nonuniform fast Fourier transform (NUFFT) is utilized to improve the computational efficiency. Simulations and experimental results are provided to demonstrate the efficacy of the proposed method in comparison with the back-projection (BP) algorithm.

eess.SP↗

arXiv2021

Non-dissipative second-order transport, spin, and pseudo-gauge transformations in hydrodynamics

We derive a set of nontrivial relations between second-order transport coefficients which follow from the second law of thermodynamics upon considering a regime close to uniform rotation of the fluid. We demonstrate that extension of hydrodynamics by spin variable is equivalent to modifying conventional hydrodynamics by a set of second-order terms satisfying the relations we derived. We point out that a novel contribution to the heat current orthogonal to vorticity and temperature gradient reminiscent of the thermal Hall effect is constrained by the second law.

hep-th↗

arXiv2021

Near-Field Millimeter-Wave Imaging via Circular-Arc MIMO Arrays

Millimeter-wave (MMW) imaging has a wide prospect in application of concealed weapons detection. We propose a circular-arc multiple-input multiple-output (MIMO) array scheme with uniformly spaced transmit and receive antennas along the horizontal-arc direction, while scanning along the vertical direction. The antenna beams of the circular-arc MIMO array can provide more uniform coverage of the imaging scene than those of the linear or planar MIMO arrays. Further, a near-field three-dimensional (3-D) imaging algorithm, based on the spatial frequency domain processing, is presented with analysis of sampling criteria and resolutions. Numerical simulations, as well as comparisons with the back-projection (BP) algorithm, are provided to show the efficacy of the proposed approach.

eess.SP↗

arXiv2021

Efficient Near-Field Imaging Using Cylindrical MIMO Arrays

Multiple-input multiple-output (MIMO) array based millimeter-wave (MMW) imaging has a tangible prospect in applications of concealed weapons detection. A near-field imaging algorithm based on wavenumber domain processing is proposed for a cylindrical MIMO array scheme with uniformly spaced transmit and receive antennas over both the vertical and horizontal-arc directions. The spectrum aliasing associated with the proposed MIMO array is analyzed through a zero-filling discrete-time Fourier transform. The analysis shows that an undersampled array can be used in recovering the MMW image by a wavenumber domain algorithm. The requirements for the antenna inter-element spacing of the MIMO array are delineated. Numerical simulations as well as comparisons with the backprojection (BP) algorithm are provided to demonstrate the effectiveness of the proposed method.

eess.SP↗