arXiv subjects

Yipeng Shi

Publications and source records attributed to Yipeng Shi.

12 recordsLinked to original sources

arXiv2026

SiriusDeliver: Automating Data Warehouse Delivery at Tencent

Enterprise data warehouses (DWs) support business-critical analytics, but warehouse task delivery remains a complicated production process involving context retrieval, workflow configuration, code generation, platform submission, and failure diagnosis. Although large language models (LLMs) and coding agents have improved software development, they are insufficient for production DW delivery, which requires dependency-aware orchestration, lifecycle-aware artifact control, and continuous adaptation to evolving platform practices. We present SiriusDeliver, an end-to-end delivery automation agent for production warehouse task submission. SiriusDeliver integrates three components: a hierarchical delivery agent that orchestrates warehouse skills, an artifact lifecycle control module that verifies and revises artifacts before and after platform execution, and a trace-driven skill evolution mechanism that maintains reusable skills from delivery trajectories. We evaluate SiriusDeliver through offline datasets and large-scale production deployment on Tencent Cloud WeData. Offline experiments on real-world warehouse delivery cases show that SiriusDeliver improves delivery success and automation efficiency over representative baselines. During a two-month deployment across 6 business teams and 4 warehouse task types, SiriusDeliver served 3,600 monthly active users and supported 18,240 delivery sessions, achieving an 87.2% end-to-end success rate and a 73.5% autonomous submission rate. A one-month A/B test shows that SiriusDeliver reduces median delivery time from 228 to 23 minutes and engineer effort from 95 to 11 minutes, while maintaining comparable final delivery success.

cs.DB↗

arXiv2026

PATS: Policy-Aware Training Scaffolding for Agentic Reinforcement Learning

In long-horizon LLM agent reinforcement learning, weak policies often repeat similar failures, producing uninformative rollout trajectories and limiting effective policy optimization. Existing skill-centric methods improve exploration by optimizing, filtering, or internalizing reusable skills. However, they remain centered on the skills themselves rather than being designed as adaptive training-time support for the evolving policy. To address this, we propose a policy-centric training paradigm that reframes skills as a dynamic training scaffold. Our framework, PATS, converts rollout groups from the latest policy into evidence cards and uses task-specific evaluation to adjust the context used in subsequent rollouts. Concrete guidance helps weak policies to complete challenging tasks. As policy improves, redundant context is revised or removed to reduce reliance on explicit guidance while preserving useful rollout variation. The policy is optimized with environmental rewards using standard RLVR, and the training scaffold is discarded at deployment. Across ALFWorld, WebShop, and seven search-augmented QA benchmarks, PATS achieves performance competitive with SOTA baselines while using 25%-50% fewer tokens.

cs.AI↗

arXiv2026

ATOD: Annealed Turn-Aware On-Policy Distillation for Multi-Turn Agentic Tasks

Training small language-model agents for long-horizon interactive tasks requires both fast imitation and reward-driven improvement. On-policy distillation (OPD) provides dense teacher guidance and typically improves rapidly in the early stage, but its gains saturate once the student approaches the teacher, limiting the final performance ceiling. Reinforcement learning (RL) directly optimizes environment rewards and encourages exploratory improvement toward a higher reward-defined ceiling, but sparse and delayed feedback makes early-stage learning much less efficient than OPD. In this paper, we propose ATOD (Annealed Turn-aware On-policy Distillation), a hybrid online distillation algorithm that explicitly exploits this complementarity. (1) ATOD uses an annealed OPD-RL schedule: OPD dominates early training to approach teacher-level behavior, while RL is gradually strengthened to drive reward-based exploration. (2) ATOD introduces Turn-level Disagreement-Uncertainty Reweighting (T-DUR), which softly gates the distillation sig- nal to prioritize turns with high disagreement or uncertainty in long trajectories. Experiments on ALFWorld, WebShop, and Search-QA show that ATOD consistently outperforms competing post-training baselines: across the three student sizes, ATOD improves average success rate by 4.16 points over OPD and 23.62 points over GRPO, while surpassing the corresponding teacher models by 2.16 points.

cs.AI↗

arXiv2025

Large Language Model Driven Development of Turbulence Models

Artificial intelligence (AI) has achieved human-level performance in specialized tasks such as Go, image recognition, and protein folding, raising the prospect of an AI singularity-where machines not only match but surpass human reasoning. Here, we demonstrate a step toward this vision in the context of turbulence modeling. By treating a large language model (LLM), DeepSeek-R1, as an equal partner, we establish a closed-loop, iterative workflow in which the LLM proposes, refines, and reasons about near-wall turbulence models under adverse pressure gradients (APGs), system rotation, and surface roughness. Through multiple rounds of interaction involving long-chain reasoning and a priori and a posteriori evaluations, the LLM generates models that not only rediscover established strategies but also synthesize new ones that outperform baseline wall models. Specifically, it recommends incorporating a material derivative to capture history effects in APG flows, modifying the law of the wall to account for system rotation, and developing rough-wall models informed by surface statistics. In contrast to conventional data-driven turbulence modeling-often characterized by human-designed, black-box architectures-the models developed here are physically interpretable and grounded in clear reasoning.

physics.flu-dyn↗

arXiv2024

RISecure-PUF: Multipurpose PUF-Driven Security Extensions with Lookaside Buffer in RISC-V

RISC-V's limited security features hinder its use in confidential computing and heterogeneous platforms. This paper introduces RISecure-PUF, a security extension utilizing existing Physical Unclonable Functions for key generation and secure protocol purposes. A one-way hash function is integrated to ensure provable security against modeling attacks, while a lookaside buffer accelerates batch sampling and minimizes reliance on error correction codes. Implemented on the Genesys 2 FPGA, RISecure-PUF improves at least $2.72\times$ in batch scenarios with negligible hardware overhead and a maximum performance reduction of $10.7\%$, enabled by reusing the hash function module in integrated environments such as cryptographic engines.

cs.CR↗

arXiv2023

A priori screening of data-enabled turbulence models

Assessing the compliance of a white-box turbulence model with known turbulent knowledge is straightforward. It enables users to screen conventional turbulence models and identify apparent inadequacies, thereby allowing for a more focused and fruitful validation and verification. However, comparing a black-box machine-learning model to known empirical scalings is not straightforward. Unless one implements and tests the model, it would not be clear if a machine-learning model, trained at finite Reynolds numbers preserves the known high Reynolds number limit. This is inconvenient, particularly because model implementation involves retraining and re-interfacing. This work attempts to address this issue, allowing fast a priori screening of machine-learning models that are based on feed-forward neural networks (FNN). The method leverages the mathematical theorems we present in the paper. These theorems offer estimates of a network's limits even when the exact weights and biases are unknown. For demonstration purposes, we screen existing machine-learning wall models and RANS models for their compliance with the log layer physics and the viscous layer physics in a priori manner. In addition, the theorems serve as essential guidelines for future machine-learning models.

physics.flu-dyn↗

arXiv2022

Modeling and simulation in supersonic three-temperature carbon dioxide turbulent channel flow

This paper pioneers the direct numerical simulation (DNS) and physical analysis in supersonic three-temperature carbon dioxide (CO2) turbulent channel flow. CO2 is a linear and symmetric triatomic molecular, with the thermal non-equilibrium three-temperature effects arising from the interactions among translational, rotational and vibrational modes under room temperature. Thus, the rotational and vibrational modes of CO2 are addressed. Thermal non-equilibrium effect of CO2 has been modeled in an extended three-temperature BGK-type model, with the calibrated translational, rotational and vibrational relaxation time. To solve the extended BGK-type equation accurately and robustly, non-equilibrium high-accuracy gas-kinetic scheme is proposed within the well-established two-stage fourth-order framework. Compared with the one-temperature supersonic turbulent channel flow, supersonic three-temperature CO2 turbulence enlarges the ensemble heat transfer of the wall by approximate 20%, and slightly decreases the ensemble frictional force. The ensemble density and temperature fields are greatly affected, and there is little change in Van Driest transformation of streamwise velocity. The thermal non-equilibrium three-temperature effects of CO2 also suppress the peak of normalized root-mean-square of density and temperature, normalized turbulent intensities and Reynolds stress. The vibrational modes of CO2 behave quite differently with rotational and translational modes. Compared with the vibrational temperature fields, the rotational temperature fields have the higher similarity with translational temperature fields, especially in temperature amplitude. Current thermal non-equilibrium models, high-accuracy DNS and physical analysis in supersonic CO2 turbulent flow can act as the benchmark for the long-term applicability of compressible CO2 turbulence.

physics.flu-dyn↗

arXiv2021

LES wall modeling for heat transfer at high speeds

A practical application of universal wall scalings is near-wall turbulence modeling. In this paper, we exploit temperature's semi-local scaling [Patel, Boersma, and Pecnik, {Scalar statistics in variable property turbulent channel flows}, Phys. Rev. Fluids, 2017, 2(8), 084604] and derive an eddy conductivity closure for wall-modeled large-eddy simulation of high-speed flows. We show that while the semi-local scaling does not collapse high-speed direct numerical simulation (DNS) data, the resulting eddy conductivity and the wall model work fairly well. The paper attempts to answer the following outstanding question: why the semi-local scaling fails but the resulting eddy conductivity works well. We conduct DNSs of Couette flows at Mach numbers from $M=1.4$ to 6. We add a source term in the energy equation to get a cold, a close-to-adiabatic wall, and a hot wall. Detailed analysis of the flows' energy budgets shows that aerodynamic heating is the answer to our question: aerodynamic heating is not accounted for in Patel et al.'s semi-local scaling but is modeled in the equilibrium wall model. We incorporate aerodynamic heating in semi-local scaling and show that the new scaling successfully collapses the high-speed DNS data. We also show that incorporating aerodynamic heating or not, the semi-local scaling gives rise to the exact same eddy conductivity, thereby answering the outstanding question.

physics.flu-dyn↗

arXiv2019

Role of magnetic field curvature in magnetohydrodynamic turbulence

Magnetic field are transported and tangled by turbulence, even as they lose identity due to nonideal or resistive effects. On balance field lines undergo stretch-twist-fold processes. The curvature field, a scalar that measures the tangling of the magnetic field lines, is studied in detail here, in the context of magnetohydrodynamic turbulence. A central finding is that the magnitudes of the curvature and the magnetic field are anti-correlated. High curvature co-locates with low magnetic field, which gives rise to power-law tails of the probability density function of the curvature field. The curvature drift term that converts magnetic energy into flow and thermal energy, largely depends on the curvature field behavior, a relationship that helps to explain particle acceleration due to curvature drift. This adds as well to evidence that turbulent effects most likely play an essential role in particle energization since turbulence drives stronger tangled field configurations, and therefore curvature.

physics.space-ph↗

arXiv2018

Dual Objective Approach Using A Convolutional Neural Network for Magnetic Resonance Elastography

Traditionally, nonlinear inversion, direct inversion, or wave estimation methods have been used for reconstructing images from MRE displacement data. In this work, we propose a convolutional neural network architecture that can map MRE displacement data directly into elastograms, circumventing the costly and computationally intensive classical approaches. In addition to the mean squared error reconstruction objective, we also introduce a secondary loss inspired by the MRE mechanical models for training the neural network. Our network is demonstrated to be effective for generating MRE images that compare well with equivalents from the nonlinear inversion method.

physics.med-ph↗

arXiv2018

Scale dependence of energy transfer in turbulent plasma

In the context of space and astrophysical plasma turbulence and particle heating, several vocabularies emerge for estimating turbulent energy dissipation rate, including Kolmogorov-Yaglom third-order law and, in its various forms, $\boldsymbol{j}\cdot\boldsymbol{E}$ (work done by the electromagnetic field on particles), and $-\left( \boldsymbol{P} \cdot \nabla \right) \cdot \boldsymbol{u}$ (pressure-strain interaction), to name a couple. It is now understood that these energy transfer channels, to some extent, are correlated with coherent structures. In particular, we find that different energy dissipation proxies, although not point-wise correlated, are concentrated in proximity to each other, for which they decorrelate in a few $d_i$(s). However, the energy dissipation proxies dominate at different scales. For example, there is an inertial range over which the third-order law is meaningful. Contributions from scale bands stemming from scale-dependent spatial filtering show that, the energy exchange through $\boldsymbol{j}\cdot\boldsymbol{E}$ mainly results from large scales, while the energy conversion from fluid flow to internal through $-\left( \boldsymbol{P} \cdot \nabla \right) \cdot \boldsymbol{u}$ dominates at small scales.

physics.space-ph↗

arXiv2015

A numerical investigation on active and passive scalars in isotropic compressible turbulence

In this paper, we investigated the statistical differences between active and passive scalars in isotropic compressible turbulence. In the inertial range, the kinetic energy and scalar have the Kolmogorov spectra, and the Kolmogorov and Obukhov-Corrsin constants are 2.06 and 0.87, respectively. The local scaling exponents computed from the second-order structure functions exist plateaus for velocity and active scalar, while that for passive scalar takes first a minimum of 0.61 then a maximum of 0.73. The mixed third-order structure function of velocity and passive scalar satisfies the 4/3-law. For the scaling exponent, the one of velocity and passive scalar mixing is between those of velocity and passive scalar, while the one of velocity and active scalar mixing is below those of velocity and active scalar. At large amplitudes, the p.d.f. of active scalar fluctuations is super-Gaussian, whereas that of passive scalar fluctuations is sub-Gaussian. Moreover, the p.d.f.s of the two scalar increments are concave and convex shapes, respectively, which exhibit strong intermittency at small scales, and approach Gaussian as scale increases. The active scalar has "ramp-cliff" structures, while the passive scalar seems to be dominated by rarefaction and compression. By employing a "coarse-graining" approach, we study scalar cascades. Unlike passive scalar, the cascade of active scalar is mainly determined by the viscous dissipation at small scales and the pressure-dilatation at large scales, where the latter is substantial in the vicinity of small-scale shocklets but is negligible after space averages, because of the cancelations between rarefaction and compression regions. Finally, in the inertial range, the p.d.f.s for the subgrid-scale fluxes of scalars can collapse to the same distribution, revealing the scale-invariant feature for the statistics of active and passive scalars.

physics.flu-dyn↗