arXiv subjects

Yangyang Zhou

Publications and source records attributed to Yangyang Zhou.

10 recordsLinked to original sources

arXiv2026

An MLIR-Based Compilation Method for Large Language Models

Large Language Models (LLMs) have become the dominant workload on modern AI accelerators, yet deploying them on specialized hardware still faces two core challenges: how to import a trained model into a compiler-friendly intermediate representation, and how to efficiently schedule the autoregressive inference loop under limited on-chip memory. This paper presents an MLIR (Multi-Level Intermediate Representation) based compilation method for large language models, illustrated using two dialects of operators, TopOp and TpuOp. TopOp serves as a high-level graph dialect that is independent of both the source framework and the target chip, and is responsible for expressing model semantics; TpuOp serves as the target hardware dialect, carrying chip-related decisions such as quantization, layer groups, and memory layout. A model is first represented as TopOp, then lowered layer by layer to TpuOp, and finally a deployable binary is generated. In addition, each Transformer layer is split into three stages for static compilation: prefill, prefill_kv (prefill with historical key-value cache), and decode, so as to accommodate the different computational characteristics of prompt-parallel processing and per-token generation. The method has been implemented in the TPU-MLIR compiler {https://github.com/sophgo/tpu-mlir} and the LLM-TPU deployment project {https://github.com/sophgo/LLM-TPU}, supporting a variety of generative models including the Qwen, Llama, InternVL, and MiniCPM-V series, as well as multiple quantization and deployment forms such as GPTQ, AWQ, and AutoRound.

cs.CL↗

arXiv2026

RMGAP: Benchmarking the Generalization of Reward Models across Diverse Preferences

Reinforcement Learning from Human Feedback has become the standard paradigm for language model alignment, where reward models directly determine alignment effectiveness. In this work, we focus on how to evaluate the generalizability of reward models. By "generalizability", we mean the ability of RMs to correctly rank responses to align with diverse user preferences. However, existing reward model benchmarks are typically designed around a universal preference, failing to assess this generalization. To address this critical gap, we introduce RMGAP, a benchmark comprising 1,097 instances across Chat, Writing, Reasoning, and Safety domains. Since different users exhibit diverse preferences for the same task, we first generate four distinct responses with different linguistic profiles for each collected prompt. However, the original prompt set lacks the specificity to convey different preferences. We therefore construct tailored prompts by contrasting these candidates and designing scenarios in which one response becomes the uniquely appropriate choice. Moreover, we observe that users often express the same preference using different phrasings, and thus extend each prompt with two paraphrased variants. Our evaluation of 24 state-of-the-art RMs reveals their substantial limitations: even the best RM achieves only 49.27% Best-of-N accuracy, highlighting considerable room for improvement in reward model generalization. Related data and code are available at https://github.com/nanzhi84/RMGAP.

cs.CL↗

arXiv2022

Conformal optical black hole for cavity

Whispering gallery mode (WGM) cavity is important for exploring physics of strong light-matter interaction. Yet it suffers from the notorious radiation loss universally due to the light tunneling effect through the curved boundary. In this work, we propose and demonstrate an optical black hole (OBH) cavity based on transformation optics. The radiation loss of all WGMs in OBH cavity is completely inhibited by an infinite wide potential barrier. Besides, the WGM field outside the cavity is revealed to follow $1/r^α$ decay rule based on conformal mapping, which is fundamentally different from the conventional Hankel-function distributions in a homogeneous cavity. Experimentally, a truncated OBH cavity is achieved based on the effective medium theory, and both the Q-factor enhancement and tightly confined WGM field are measured in the microwave spectra which agree well with the theoretical results. The circular OBH cavity is further applied to the arbitrary-shaped cavities including single-core and multi-core structures with high-Q factor via the conformal mapping. The OBH cavity design strategy can be generalized to resonant modes of various wave systems, such as acoustic and elastic waves, and finds applications in energy harvesting and optoelectronics.

physics.class-ph↗

arXiv2021

Square Maxwell's fish-eye lens for broadband achromatic super-resolution imaging

Broadband super-resolution imaging is important in optical field. To achieve super-resolution imaging, various lenses from superlens to solid immersion lens have been designed and fabricated in recent years. However, the imaging is unsatisfactory due to low work efficiency and narrow band. In this work, we propose a solid immersion square Maxwell's fish-eye lens which realizes broadband (7-16GHz) achromatic super-resolution imaging with full width at half maximum around 0.2 wavelength based on transformation optics at microwave frequencies. In addition, a super-resolution information transmission channel is also designed to realize long-distance multi-source super-resolution information transmission based on the super-resolution lens.

physics.app-ph↗

arXiv2021

Multiple super-resolution imaging from Mikaelian lens to generalized Maxwell's fish-eye lens

Super-resolution imaging is vital for optical applications, such as high capacity information transmission, real-time bio-molecular imaging and nanolithography. Technology and method of super-resolution imaging have attracted much attention. In recent years, different kinds of novel lens, from the superlens to the super-oscillatory lens, has been designed and fabricated to break out the diffraction limit. However, the effect of the super-resolution imaging in these lenses is not satisfactory due to intrinsic loss, aberration, bigger sidebands and single super-resolution imaging. In this research, we introduce the solid immersion mechanism to Mikaelian lens (ML) for multiple super-resolution imaging.

physics.optics↗

arXiv2021

Super-resolution imaging in absolute instruments

It has been shown that negative refraction makes a perfect lens. However, with little loss, the imaging functionality will be strongly compromised. Later on, it was proved that positive refraction from Maxwell's fish-eye lens can also makes a perfect lens. However, strong debating happens on the introduced drain problem at the imaging position. In this work, we for the first time find that a solid immersion Maxwell's fish-eye lens could be used for super-resolution imaging. We find that it is due to the perfect focusing and total reflection at the outer interface, such that a super-resolution image is formed at the required position in the air background. This simple mechanism will also be valid for other absolute instruments and more versatile super-imaging systems will be anticipated.

physics.optics↗

arXiv2020

Enumeration of row-increasing tableaux of two-row skew shapes

In this paper, we firstly extend a result of Bonin, Shapiro and Simion by giving the distribution of the major index over generalized Schröder paths. Then by providing a bijection between generalized Schröder paths and row-increasing tableaux of skew shapes with two rows, we obtain the distribution of the major index and the amajor index over these tableaux, which extends a result of Du, Fan and Zhao. We also generalize a result of Pechenik and give the distribution of the major index over increasing tableaux of skew shapes with two rows. Especially, a bijection from row-increasing tableaux with shape $(n,m)$ and maximal value $n+m-k$ to standard Young tableaux with shape $((n-k+1,m-k+1,1^k)/(1^2))$ is obtained.

math.CO↗

arXiv2019

Effects of Some Operations on Domination Chromatic Number in Graphs

For a simple graph $G$, a domination coloring of $G$ is a proper vertex coloring such that every vertex of $G$ dominates at least one color class, and every color class is dominated by at least one vertex. The domination chromatic number, denoted by $χ_{dd}(G)$, is minimum number of colors among all domination colorings of $G$. In this paper, we discuss the effects of some typical operations on $χ_{dd}(G)$, such as vertex (edge) removal, vertex (edge) contraction, edge subdivision, and cycle extending.

cs.DM↗

arXiv2019

On Domination Coloring in Graphs

A domination coloring of a graph $G$ is a proper vertex coloring of $G$ such that each vertex of $G$ dominates at least one color class, and each color class is dominated by at least one vertex. The minimum number of colors among all domination colorings is called the domination chromatic number, denoted by $χ_{dd}(G)$. In this paper, we study the complexity of this problem by proving its NP-Completeness for arbitrary graphs, and give general bounds and characterizations on several classes of graphs. We also show the relation between dominator chromatic number $χ_{d}(G)$, dominated chromatic number $χ_{dom}(G)$, chromatic number $χ(G)$, and domination number $γ(G)$. We present several results on graphs with $χ_{dd}(G)=χ(G)$.

cs.DM↗

arXiv2018

Transformation devices with optical nihility media and reduced realizations

Starting from optical nihility media (ONM), we design several intriguing devices with transformation optics method in two dimensions, such as a wave splitter, a concave lens, a field rotator, a concentrator and an invisibility cloak. The extreme anisotropic property of ONM hinders the fabrication of these devices, which could be effectively realized by simplified materials with Fabry-Pérot resonances (FPs) at discrete frequencies. Moreover, we propose a reduced version of simplified materials with FPs to construct a concentrator and a rotator, which is feasible in experimental fabrications. The simulations of total scattering cross sections confirm their functionalities.

physics.optics↗