arXiv subjects

Yanpeng Wang

Publications and source records attributed to Yanpeng Wang.

18 recordsLinked to original sources

arXiv2026

CogEvol: Towards Efficient and Reliable Learning Environment Generation

We present CogEvol, a family of models trained specifically for Learning Environment Generation: turning a course brief into a finished learning artifact (structured-JSON slides or self-contained interactive HTML pages) in a single pass. Across 220k production requests, CogEvol completes a slide in a median of 17 seconds and an interactive page in 59, replacing minutes-long multi-turn agent scaffolding. Reliability is enforced rather than hoped for: a production-grounded data pipeline turns real failures into 53,687 verified SFT samples, and a hybrid rule-plus-VLM reward drives GRPO-based RL, hardened after we caught and fixed a reward-hacking episode that produced visually convincing but unplayable games. CogEvol-27B scores 83.7 on slide quality and 63.7 on a 500-case interactive-HTML benchmark with 26.9x fewer parameters than flagship coding models, and, in collaboration with the OpenMAIC team, serves their live production traffic. CogEvol-4B is released openly under the Apache 2.0 license at https://github.com/CogEvol/CogEvol-4B; external flagships are measured on the same suites under the identical harness. Scaffold editing cuts interactive-page generation cost by a further ~76%, and the full stack runs on domestic Ascend accelerators at application-level parity with A800 GPUs, lowering the unit cost of AI-native education at scale.

cs.CL↗

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

Harness Engineering for LLM-Driven GPU Kernel Generation

Large language models (LLMs) can assist GPU kernel generation, but their practical effectiveness depends on whether generated code can be reliably constrained, validated, profiled, and selected. This paper presents a harness-centered system for LLM-driven GPU kernel optimization in the MLSys 2026 FlashInfer AI Kernel Generation Contest on NVIDIA Blackwell B200 GPUs. The system separates an evaluation harness from a profile-backed optimization controller: the harness enforces compilation, correctness, official-aligned timing, and artifact archival, while the controller turns profiler and workload evidence into bounded candidate-generation decisions. Human-authored skills capture operator constraints, references, profiling procedures, and promotion rules, while Codex and Claude Code agents generate candidate kernels inside those constraints. Across five operator definitions, the retained official-aligned artifacts achieved mean-latency speedups over supplied FlashInfer baselines of 1.62x, 18.05x, 29.68x, 1.12x, and 13.70x. The Agent-Assisted kernels outperform the Full-Agent artifacts across the evaluated definitions, indicating that expert-provided optimization directions, high-quality references, and workload context remain critical for reliable AI-driven kernel optimization.

cs.LG↗

arXiv2026

King Function for Shifted Gaussian: Laguerre Structure, Spectral Theory and Density

We study King function arising as radial kernels in the laboratory-frame spherical harmonic expansion of shifted Gaussian distributions. We first clarify their relation with the co-moving Laguerre hierarchy by means of a King--Laguerre expansion. We then derive the King differential equation and show that the associated self-adjoint operator in a Gaussian-weighted Hilbert space is unitarily equivalent to the free radial Schr\"odinger operator on the half-line. This yields the spectral representation and generalized eigenfunction. Finally, we prove that real-parameter King function, lies in the resolvent set, form a dense non-orthogonal system in a natural radial velocity space, providing an approximation-theoretic basis for King mixture representations. Weighted $L^1$-integrability criteria and closed-form moment formulas are also derived, justifying the normalization of King function.

math-ph↗

arXiv2026

Revealing the Challenges of Attention-FFN Disaggregation for Modern MoE Models and Hardware Systems

Deploying large-scale MoE models presents challenges in memory capacity and bandwidth for expert activation. While Attention-FFN Disaggregation (AFD) has emerged as a potential architecture to decouple compute and memory resources, its performance boundaries compared to standard large-scale Expert Parallelism (EP) remain underexplored. In this paper, we conduct a systematic analysis of AFD by extending the roofline model to the communication level, correlating interconnect bandwidth, arithmetic intensity, and Hardware FLOPS Utilization (HFU). Our analysis reveals a dead zone on standard clusters: increasing FFN instance count fails to improve HFU as computational workload is capped by scale-out bandwidth, causing operator active time to shrink relative to the fixed latency budget. We further show that AFD's discrete node-level scaling incurs higher imbalance penalties than EP's continuous batch adjustment. Nevertheless, these limitations diminish under specific conditions: Superpod-class hardware with abundant interconnect bandwidth and models with coarse-grained experts and lower sparsity are more likely to benefit from AFD. These findings position AFD as a promising approach for specific hardware-model combinations rather than a universal solution.

cs.DC↗

arXiv2025

LoongFlow: Directed Evolutionary Search via a Cognitive Plan-Execute-Summarize Paradigm

The transition from static Large Language Models (LLMs) to self-improving agents is hindered by the lack of structured reasoning in traditional evolutionary approaches. Existing methods often struggle with premature convergence and inefficient exploration in high-dimensional code spaces. To address these challenges, we introduce LoongFlow, a self-evolving agent framework that achieves state-of-the-art solution quality with significantly reduced computational costs. Unlike "blind" mutation operators, LoongFlow integrates LLMs into a cognitive "Plan-Execute-Summarize" (PES) paradigm, effectively mapping the evolutionary search to a reasoning-heavy process. To sustain long-term architectural coherence, we incorporate a hybrid evolutionary memory system. By synergizing Multi-Island models with MAP-Elites and adaptive Boltzmann selection, this system theoretically balances the exploration-exploitation trade-off, maintaining diverse behavioral niches to prevent optimization stagnation. We instantiate LoongFlow with a General Agent for algorithmic discovery and an ML Agent for pipeline optimization. Extensive evaluations on the AlphaEvolve benchmark and Kaggle competitions demonstrate that LoongFlow outperforms leading baselines (e.g., OpenEvolve, ShinkaEvolve) by up to 60% in evolutionary efficiency while discovering superior solutions. LoongFlow marks a substantial step forward in autonomous scientific discovery, enabling the generation of expert-level solutions with reduced computational overhead.

cs.AI↗

arXiv2025

Staggered Batch Scheduling: Co-optimizing Time-to-First-Token and Throughput for High-Efficiency LLM Inference

The evolution of Large Language Model (LLM) serving towards complex, distributed architectures--specifically the P/D-separated, large-scale DP+EP paradigm--introduces distinct scheduling challenges. Unlike traditional deployments where schedulers can treat instances as black boxes, DP+EP architectures exhibit high internal synchronization costs. We identify that immediate request dispatching in such systems leads to severe in-engine queuing and parallelization bubbles, degrading Time-to-First-Token (TTFT). To address this, we propose Staggered Batch Scheduling (SBS), a mechanism that deliberately buffers requests to form optimal execution batches. This temporal decoupling eliminates internal queuing bubbles without compromising throughput. Furthermore, leveraging the scheduling window created by buffering, we introduce a Load-Aware Global Allocation strategy that balances computational load across DP units for both Prefill and Decode phases. Deployed on a production H800 cluster serving Deepseek-V3, our system reduces TTFT by 30%-40% and improves throughput by 15%-20% compared to state-of-the-art immediate scheduling baselines.

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↗

arXiv2025

Transport theory in moderately anisotropic plasmas: I, Collisionless aspects of axisymmetric velocity space

A novel transport theory, based on the finitely distinguishable independent features (FDIF) hypothesis, is presented for scenarios when velocity space exhibits axisymmetry. In this theory, the transport equations are derived from the 1D-2V Vlasov equation, employing the spherical harmonics expansions (SHE) together with the King function expansion (KFE) in velocity space. The characteristic parameter equations (CPEs) are provided based on the general King mixture model (GKMM), serving as the constraint equations of the transport equations. It is a nature process to present the closure relations of transport equations based on SHE and KFE, successfully providing a kinetic moment-closed model (KMCM). This model is typically a nonlinear system, effective for moderately anisotropic non-equilibrium plasmas.

physics.plasm-ph↗

arXiv2024

Higher-order moment convergent method in weakly anisotropic plasma and the NLVFP code for solution of the 0D-2V Vlasov-Fokker-Planck equation

Fusion plasma and space plasma are typical non-equilibrium and nonlinear systems, with the interactions between different species well described by the Vlasov-Fokker-Planck (VFP) equations. The transport of mass, momentum, energy, and temperature relaxation are important issues, which are affected by the collision term of VFP even in so-called collisionless plasma domain. Hence, nonlinearity and collisions are important features in large regime. A successful numerical simulation for non-equilibrium plasma has to be able to conserve mass, momentum and energy, while satisfying Boltzmann's H-theorem and higher-order moment convergence. An expansion of the distribution function in spherical harmonics (Legendre basis when the velocity space exhibits axisymmetry) in angle coordinate and in King basis in speed coordinate of velocity space is well suited to address these requirements. This paper reviews the formulation of the 0D-2V VFP equation in terms of spherical harmonics coupled with King function and its solution in our NLVFP code. In this topic review, we will introduce the background physics related to the nonlinear VFP simulation, then describe NLVFP for 0D-2V homogeneous, weakly anisotropic plasma with utilization of the Shkarofsky's form of Fokker-Planck-Rosenbluth (FPRS) collision operator.

physics.plasm-ph↗

arXiv2024

General relaxation model for a homogeneous plasma with spherically symmetric velocity space

A kinetic moment-closed model (KMCM), derived from the Vlasov-Fokker-Planck (VFP) equation with spherically symmetric velocity space, is introduced as a general relaxation model for homogeneous plasmas. The closed form of this model is presented by introducing a set new functions called $R$ function and $R$ integration. This nonlinear model, based on the finitely distinguishable independent features (FDIF) hypothesis, enables the capture of the nature of the equilibrium state. From this relaxation model, a general temperature relaxation model is derived when velocity space exhibits spherical symmetry, and the general characteristic frequency of temperature relaxation is presented.

physics.plasm-ph↗

arXiv2024

A conservative, implicit solver for 0D-2V multi-species nonlinear Fokker-Planck collision equations

In this study, we present an optimal implicit algorithm specifically designed to accurately solve the multi-species nonlinear 0D-2V axisymmetric Fokker-Planck-Rosenbluth (FPR) collision equation while preserving mass, momentum, and energy. Our approach relies on the utilization of nonlinear Shkarofsky's formula of FPR (FPRS) collision operator in the spherical-polar coordinate. The key innovation lies in the introduction of a new function named King, with the adoption of the Legendre polynomial expansion for the angular coordinate and King function expansion for the speed coordinate. The Legendre polynomial expansion will converge exponentially and the King method, a moment convergence algorithm, could ensure the conservation with high precision in discrete form. Additionally, post-step projection onto manifolds is employed to exactly enforce symmetries of the collision operators. Through solving several typical problems across various nonequilibrium configurations, we demonstrate the high accuracy and superior performance of the presented algorithm for weakly anisotropic plasmas.

math.NA↗

arXiv2024

Relaxation model for a homogeneous plasma with spherically symmetric velocity space

We derive the transport equations from the Vlasov-Fokker-Planck equation when the velocity space is spherically symmetric. The Shkarofsky's form of Fokker-Planck-Rosenbluth collision operator is employed in the Vlasov-Fokker-Planck equation. A closed-form relaxation model for homogeneous plasmas could be presented in terms of Gauss hypergeometric2F1 functions. This has been accomplished based on the Maxwellian mixture model. Furthermore, we demonstrate that classic models such as two-temperature thermal equilibrium model and thermodynamic equilibrium model are special cases of our relaxation model and the zeroth-order Braginskii heat transfer model can also be derived. The present relaxation model is a nonequilibrium model based on the hypothesis that the plasmas system possesses finitely distinguishable independent features, without relying on the conventional near-equilibrium assumption.

physics.plasm-ph↗

arXiv2022

Regular sets in Cayley graphs

In a graph $Γ$ with vertex set $V$, a subset $C$ of $V$ is called an $(a,b)$-perfect set if every vertex in $C$ has exactly $a$ neighbors in $C$ and every vertex in $V\setminus C$ has exactly $b$ neighbors in $C$, where $a$ and $b$ are nonnegative integers. In the literature $(0,1)$-perfect sets are known as perfect codes and $(1,1)$-perfect sets are known as total perfect codes. In this paper we prove that, for any finite group $G$, if a non-trivial normal subgroup $H$ of $G$ is a perfect code in some Cayley graph of $G$, then it is also an $(a,b)$-perfect set in some Cayley graph of $G$ for any pair of integers $a$ and $b$ with $0\leqslant a\leqslant|H|-1$ and $0\leqslant b\leqslant |H|$ such that $\gcd(2,|H|-1)$ divides $a$. A similar result involving total perfect codes is also proved in the paper.

math.CO↗

arXiv2022

Subgroup regular sets in Cayley graphs

Let $Γ$ be a graph with vertex set $V$, and let $a$ and $b$ be nonnegative integers. A subset $C$ of $V$ is called an $(a,b)$-regular set in $Γ$ if every vertex in $C$ has exactly $a$ neighbors in $C$ and every vertex in $V\setminus C$ has exactly $b$ neighbors in $C$. In particular, $(0, 1)$-regular sets and $(1, 1)$-regular sets in $\Ga$ are called perfect codes and total perfect codes in $\Ga$, respectively. A subset $C$ of a group $G$ is said to be an $(a,b)$-regular set of $G$ if there exists a Cayley graph of $G$ which admits $C$ as an $(a,b)$-regular set. In this paper we prove that, for any generalized dihedral group $G$ or any group $G$ of order $4p$ or $pq$ for some primes $p$ and $q$, if a nontrivial subgroup $H$ of $G$ is a $(0, 1)$-regular set of $G$, then it must also be an $(a,b)$-regular set of $G$ for any $0\leqslant a\leqslant|H|-1$ and $0\leqslant b\leqslant |H|$ such that $a$ is even when $|H|$ is odd. A similar result involving $(1, 1)$-regular sets of such groups is also obtained in the paper.

math.CO↗

arXiv2019

Characterization of subgroup perfect codes in Cayley graphs

A subset $C$ of the vertex set of a graph $Γ$ is called a perfect code in $Γ$ if every vertex of $Γ$ is at distance no more than $1$ to exactly one vertex of $C$. A subset $C$ of a group $G$ is called a perfect code of $G$ if $C$ is a perfect code in some Cayley graph of $G$. In this paper we give sufficient and necessary conditions for a subgroup $H$ of a finite group $G$ to be a perfect code of $G$. Based on this, we determine the finite groups that have no nontrivial subgroup as a perfect code, which answers a question by Ma, Walls, Wang and Zhou.

math.CO↗