arXiv subjects

Zhe Fu

Publications and source records attributed to Zhe Fu.

27 records · Page 2Linked to original sources

arXiv2024

DeepSeek LLM: Scaling Open-Source Language Models with Longtermism

The rapid development of open-source large language models (LLMs) has been truly remarkable. However, the scaling law described in previous literature presents varying conclusions, which casts a dark cloud over scaling LLMs. We delve into the study of scaling laws and present our distinctive findings that facilitate scaling of large scale models in two commonly used open-source configurations, 7B and 67B. Guided by the scaling laws, we introduce DeepSeek LLM, a project dedicated to advancing open-source language models with a long-term perspective. To support the pre-training phase, we have developed a dataset that currently consists of 2 trillion tokens and is continuously expanding. We further conduct supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) on DeepSeek LLM Base models, resulting in the creation of DeepSeek Chat models. Our evaluation results demonstrate that DeepSeek LLM 67B surpasses LLaMA-2 70B on various benchmarks, particularly in the domains of code, mathematics, and reasoning. Furthermore, open-ended evaluations reveal that DeepSeek LLM 67B Chat exhibits superior performance compared to GPT-3.5.

cs.CL↗

arXiv2023

Cooperative Driving for Speed Harmonization in Mixed-Traffic Environments

Autonomous driving systems present promising methods for congestion mitigation in mixed autonomy traffic control settings. In particular, when coupled with even modest traffic state estimates, such systems can plan and coordinate the behaviors of automated vehicles (AVs) in response to observed downstream events, thereby inhibiting the continued propagation of congestion. In this paper, we present a two-layer control strategy in which the upper layer proposes the desired speeds that predictively react to the downstream state of traffic, and the lower layer maintains safe and reasonable headways with leading vehicles. This method is demonstrated to achieve an average of over 15% energy savings within simulations of congested events observed in Interstate 24 with only 4% AV penetration, while restricting negative externalities imposed on traveling times and mobility. The proposed strategy that served as part of the "speed planner" was deployed on 100 AVs in a massive traffic experiment conducted on Nashville's I-24 in November 2022.

eess.SY↗

arXiv2022

Learning energy-efficient driving behaviors by imitating experts

The rise of vehicle automation has generated significant interest in the potential role of future automated vehicles (AVs). In particular, in highly dense traffic settings, AVs are expected to serve as congestion-dampeners, mitigating the presence of instabilities that arise from various sources. However, in many applications, such maneuvers rely heavily on non-local sensing or coordination by interacting AVs, thereby rendering their adaptation to real-world settings a particularly difficult challenge. To address this challenge, this paper examines the role of imitation learning in bridging the gap between such control strategies and realistic limitations in communication and sensing. Treating one such controller as an "expert", we demonstrate that imitation learning can succeed in deriving policies that, if adopted by 5% of vehicles, may boost the energy-efficiency of networks with varying traffic conditions by 15% using only local observations. Results and code are available online at https://sites.google.com/view/il-traffic/home.

cs.RO↗

arXiv2021

From Cloud to Edge: A First Look at Public Edge Platforms

Public edge platforms have drawn increasing attention from both academia and industry. In this study, we perform a first-of-its-kind measurement study on a leading public edge platform that has been densely deployed in China. Based on this measurement, we quantitatively answer two critical yet unexplored questions. First, from end users' perspective, what is the performance of commodity edge platforms compared to cloud, in terms of the end-to-end network delay, throughput, and the application QoE. Second, from the edge service provider's perspective, how are the edge workloads different from cloud, in terms of their VM subscription, monetary cost, and resource usage. Our study quantitatively reveals the status quo of today's public edge platforms, and provides crucial insights towards developing and operating future edge services.

cs.NI↗

arXiv2019

The three-state Potts model on the centered triangular lattice

We study phase transitions of the Potts model on the centered-triangular lattice with two types of couplings, namely $K$ between neighboring triangular sites, and $J$ between the centered and the triangular sites. Results are obtained by means of a finite-size analysis based on numerical transfer matrix calculations and Monte Carlo simulations. Our investigation covers the whole $(K, J)$ phase diagram, but we find that most of the interesting physics applies to the antiferromagnetic case $K<0$, where the model is geometrically frustrated. In particular, we find that there are, for all finite $J$, two transitions when K is varied. Their critical properties are explored. In the limits $J\to \pm \infty$ we find algebraic phases with infinite-order transitions to the ferromagnetic phase.

cond-mat.stat-mech↗

arXiv2016

Special transitions in an O($n$) loop model with an Ising-like constraint

We investigate the O($n$) nonintersecting loop model on the square lattice under the constraint that the loops consist of ninety-degree bends only. The model is governed by the loop weight $n$, a weight $x$ for each vertex of the lattice visited once by a loop, and a weight $z$ for each vertex visited twice by a loop. We explore the $(x,z)$ phase diagram for some values of $n$. For $0 1$, the O($n$)-like transition line appears to be absent. Thus, for $z=0$, the $(n,x)$ phase diagram displays a line of phase transitions for $n\le 1$. The line ends at $n=1$ in an infinite-order transition. We determine the conformal anomaly and the critical exponents along this line. These results agree accurately with a recent proposal for the universal classification of this type of model, at least in most of the range $-1 \leq n \leq 1$. We also determine the exponent describing crossover to the generic O($n$) universality class, by introducing topological defects associated with the introduction of `straight' vertices violating the ninety-degree-bend rule. These results are obtained by means of transfer-matrix calculations and finite-size scaling.

cond-mat.stat-mech↗

arXiv2013

Ising-like transitions in the O($n$) loop model on the square lattice

We explore the phase diagram of the O($n$) loop model on the square lattice in the $(x,n)$ plane, where $x$ is the weight of a lattice edge covered by a loop. These results are based on transfer-matrix calculations and finite-size scaling. We express the correlation length associated with the staggered loop density in the transfer-matrix eigenvalues. The finite-size data for this correlation length, combined with the scaling formula, reveal the location of critical lines in the diagram. For $n>>2$ we find Ising-like phase transitions associated with the onset of a checkerboard-like ordering of the elementary loops, i.e., the smallest possible loops, with the size of an elementary face, which cover precisely one half of the faces of the square lattice at the maximum loop density. In this respect, the ordered state resembles that of the hard-square lattice gas with nearest-neighbor exclusion, and the finiteness of $n$ represents a softening of its particle-particle potentials. We also determine critical points in the range $-2\leq n\leq 2$. It is found that the topology of the phase diagram depends on the set of allowed vertices of the loop model. Depending on the choice of this set, the $n>2$ transition may continue into the dense phase of the $n \leq 2$ loop model, or continue as a line of $n \leq 2$ O($n$) multicritical points.

cond-mat.stat-mech↗

arXiv2012

Exact critical points of the O($n$) loop model on the martini and the 3-12 lattices

We derive the exact critical line of the O($n$) loop model on the martini lattice as a function of the loop weight $n$.A finite-size scaling analysis based on transfer matrix calculations is also performed.The numerical results coincide with the theoretical predictions with an accuracy up to 9 decimal places. In the limit $n\to 0$, this gives the exact connective constant $μ=1.7505645579...$ of self-avoiding walks on the martini lattice. Using similar numerical methods, we also study the O($n$) loop model on the 3-12 lattice. We obtain similarly precise agreement with the exact critical points given by Batchelor [J. Stat. Phys. 92, 1203 (1998)].

cond-mat.stat-mech↗

arXiv2010

Critical frontier for the Potts and percolation models on triangular-type and kagome-type lattices II: Numerical analysis

In a recent paper (arXiv:0911.2514), one of us (FYW) considered the Potts model and bond and site percolation on two general classes of two-dimensional lattices, the triangular-type and kagome-type lattices, and obtained closed-form expressions for the critical frontier with applications to various lattice models. For the triangular-type lattices Wu's result is exact, and for the kagome-type lattices Wu's expression is under a homogeneity assumption. The purpose of the present paper is two-fold: First, an essential step in Wu's analysis is the derivation of lattice-dependent constants $A, B, C$ for various lattice models, a process which can be tedious. We present here a derivation of these constants for subnet networks using a computer algorithm. Secondly, by means of a finite-size scaling analysis based on numerical transfer matrix calculations, we deduce critical properties and critical thresholds of various models and assess the accuracy of the homogeneity assumption. Specifically, we analyze the $q$-state Potts model and the bond percolation on the 3-12 and kagome-type subnet lattices $(n\times n):(n\times n)$, $n\leq 4$, for which the exact solution is not known. To calibrate the accuracy of the finite-size procedure, we apply the same numerical analysis to models for which the exact critical frontiers are known. The comparison of numerical and exact results shows that our numerical determination of critical thresholds is accurate to 7 or 8 significant digits. This in turn infers that the homogeneity assumption determines critical frontiers with an accuracy of 5 decimal places or higher. Finally, we also obtained the exact percolation thresholds for site percolation on kagome-type subnet lattices $(1\times 1):(n\times n)$ for $1\leq n \leq 6$.

cond-mat.stat-mech↗