arXiv subjects

Chenyang Zhou

Publications and source records attributed to Chenyang Zhou.

13 recordsLinked to original sources

arXiv2026

RhinoVLA Technical Report

Vision-Language-Action (VLA) models have shown strong potential for robotic manipulation, but real-time deployment on edge hardware remains challenging. In this work, we identify VLM visual and context tokens as a major source of deployment latency: for GEMM-dominated projection operators, computation grows linearly with the number of input tokens when model dimensions are fixed. Motivated by this observation, we propose RhinoVLA, a deployment-oriented VLA model co-designed with the Huixi R1 edge SoC. RhinoVLA adopts a token-efficient Qwen3-VL backbone and a continuous Action Expert, reducing the VLM-side token and computation burden while preserving pretrained multimodal capability. To support cross-robot learning, RhinoVLA further introduces a unified interface that combines View Registry, 72D physical state-action slot space, and robotinstance LoRA, allowing heterogeneous robot observations and action schemas to be aligned under a shared policy. On the deployment side, RhinoVLA is optimized through hardware-aware compilation, mixed-precision execution, and parallel visual encoding. Experiments show that RhinoVLA achieves downstream performance comparable to {\pi}0.5 at a similar parameter scale, while reaching 11.69 Hz end-to-end inference on Huixi R1, meeting the 10 Hz real-time closedloop control target. The project will be open-sourced at https://github.com/HuixiAI/RhinoVLA.

cs.RO↗

arXiv2026

ChipMATE: Multi-Agent Training via Reinforcement Learning for Enhanced RTL Generation

Existing API-based agentic systems for RTL code generation are fundamentally misaligned with industrial practice: they assume a golden testbench is available at generation time, rely on closed-source APIs incompatible with chip vendors' air-gapped security requirements, and cannot be trained on vendors' proprietary RTL codebases, leaving valuable internal data unused. Recent self-trained models address the deployment constraint but remain single-turn generators that overlook the critical role of verification in real industrial flows. To bridge these gaps, we present ChipMATE, the first self-trained multi-agent framework for RTL generation. Inspired by industrial practice where correctness emerges from cross-comparison between independently written RTL modules and reference models, ChipMATE pairs a Verilog agent with a Python reference-model agent that mutually verify each other's outputs without any golden oracle. We design a backtrack-based inference workflow to prevent error propagation across turns, and a two-stage training pipeline that first trains each agent individually to saturate its code-generation capability, then trains the team jointly to collaborate effectively. To support the training, we further build a hybrid data-generation framework that produces 64.4K high-quality reference model training samples. ChipMATE achieves 75.0\% and 80.1\% pass@1 on VerilogEval V2 with 4B and 9B base models, outperforming all existing self-trained models and even DeepSeek V4 with 1600B parameters. Our code and model weights are publicly available in https://github.com/zhongkaiyu/ChipMATE.

cs.MA↗

arXiv2026

AMMA: A Multi-Chiplet Memory-Centric Architecture for Low-Latency 1M Context Attention Serving

All current LLM serving systems place the GPU at the center, from production-level attention-FFN disaggregation to NVIDIA's Rubin GPU-LPU heterogeneous platform. Even academic PIM/PNM proposals still treat the GPU as the central hub for cross-device communication. Yet the GPU's compute-rich architecture is fundamentally mismatched with the memory-bound nature of decode-phase attention, inflating serving latency while wasting power and die area on idle compute units. The problem is compounded as reasoning and agentic workloads push context lengths toward one million tokens, making attention latency the primary user-facing bottleneck. To address these inefficiencies, we present AMMA, a multi-chiplet, memory-centric architecture for low-latency long-context attention. AMMA replaces GPU compute dies with HBM-PNM cubes, roughly doubling the available memory bandwidth to better serve memory-bound attention workloads. To translate this bandwidth into proportional performance gains, we introduce (i) a logic-die microarchitecture that fully exploits per-cube internal bandwidth for decode attention under a minimal power and area budget, (ii) a two-level hybrid parallelism scheme, and (iii) a reordered collective flow that reduces intra-chip die-to-die communication overhead. We further conduct a design-space exploration over per-cube compute power and intra-chip D2D link bandwidth, providing actionable guidance for hardware designers. Evaluations show that AMMA achieves 15.5X lower attention latency and 6.9X lower energy consumption compared with the NVIDIA H100.

cs.AR↗

arXiv2026

Data Darwinism Part II: DataEvolve -- AI can Autonomously Evolve Pretraining Data Curation

Data Darwinism (Part I) established a ten-level hierarchy for data processing, showing that stronger processing can unlock greater data value. However, that work relied on manually designed strategies for a single category. Modern pretraining corpora comprise hundreds of heterogeneous categories spanning domains and content types, each demanding specialized treatment. At this scale, manual strategy design becomes prohibitive. This raises a key question: can strategies evolve in an automated way? We introduce DataEvolve, a framework that enables strategies to evolve through iterative optimization rather than manual design. For each data category, DataEvolve operates in a closed evolutionary loop: it identifies quality issues, generates candidate strategies, executes them on sampled data, evaluates results, and refines approaches across generations. The process accumulates knowledge through an experience pool of discovered issues and a strategy pool tracking performance across iterations. Applied to 8 categories spanning 672B tokens from Nemotron-CC, DataEvolve produces Darwin-CC, a 504B-token dataset with strategies evolved through 30 iterations per category. Training 3B models on 500B tokens, Darwin-CC outperforms raw data (+3.96 points) and achieves a 44.13 average score across 18 benchmarks, surpassing DCLM, Ultra-FineWeb, and FineWeb-Edu, with strong gains on knowledge-intensive tasks such as MMLU. Analysis shows evolved strategies converge on cleaning-focused approaches: targeted noise removal and format normalization with domain-aware preservation, echoing the L4 (Generative Refinement) principles from Part I. Ablation studies confirm iterative evolution is essential: optimized strategies outperform suboptimal ones by 2.93 points, establishing evolutionary strategy design as feasible and necessary for pretraining-scale data curation.

cs.AI↗

arXiv2026

Data Darwinism Part I: Unlocking the Value of Scientific Data for Pre-training

Data quality determines foundation model performance, yet systematic processing frameworks are lacking. We introduce Data Darwinism, a ten-level taxonomy (L0-L9) that conceptualizes data-model co-evolution: advanced models produce superior data for next-generation systems. We validate this on scientific literature by constructing Darwin-Science, a 900B-token corpus (L0-L5). We identify a learnability gap in raw scientific text, which we bridge via L4 (Generative Refinement) and L5 (Cognitive Completion) using frontier LLMs to explicate reasoning and terminology. To ensure rigorous attribution, we pre-trained daVinci-origin-3B/7B models from scratch, excluding scientific content to create contamination-free baselines. After 600B tokens of continued pre-training, Darwin-Science outperforms baselines by +2.12 (3B) and +2.95 (7B) points across 20+ benchmarks, rising to +5.60 and +8.40 points on domain-aligned tasks. Systematic progression to L5 yields a +1.36 total gain, confirming that higher-level processing unlocks latent data value. We release the Darwin-Science corpus and daVinci-origin models to enable principled, co-evolutionary development.

cs.AI↗

arXiv2026

ChipBench: A Next-Step Benchmark for Evaluating LLM Performance in AI-Aided Chip Design

While Large Language Models (LLMs) show significant potential in hardware engineering, current benchmarks suffer from saturation and limited task diversity, failing to reflect LLMs' performance in real industrial workflows. To address this gap, we propose a comprehensive benchmark for AI-aided chip design that rigorously evaluates LLMs across three critical tasks: Verilog generation, debugging, and reference model generation. Our benchmark features 44 realistic modules with complex hierarchical structures, 89 systematic debugging cases, and 132 reference model samples across Python, SystemC, and CXXRTL. Evaluation results reveal substantial performance gaps, with state-of-the-art Claude-4.5-opus achieving only 30.74\% on Verilog generation and 13.33\% on Python reference model generation, demonstrating significant challenges compared to existing saturated benchmarks where SOTA models achieve over 95\% pass rates. Additionally, to help enhance LLM reference model generation, we provide an automated toolbox for high-quality training data generation, facilitating future research in this underexplored domain. Our code is available at https://github.com/zhongkaiyu/ChipBench.git.

cs.AI↗

arXiv2025

Patterns behind Chaos: Forecasting Data Movement for Efficient Large-Scale MoE LLM Inference

Large-scale Mixture of Experts (MoE) Large Language Models (LLMs) have recently become the frontier open-weight models, achieving remarkable model capability similar to proprietary ones. But their random expert selection mechanism introduces significant data movement overhead that becomes the dominant bottleneck in multi-unit LLM serving systems. To understand the patterns underlying this data movement, we conduct comprehensive data-movement-centric profiling across four state-of-the-art large-scale MoE models released in 2025 (200B-1000B) using over 24,000 requests spanning diverse workloads. We perform systematic analysis from both temporal and spatial perspectives and distill six key insights to guide the design of diverse serving systems. We verify these insights on both future wafer-scale GPU architectures and existing GPU systems. On wafer-scale GPUs, lightweight architectural modifications guided by our insights yield a 6.6$\times$ average speedup across four 200B--1000B models. On existing GPU systems, our insights drive the design of a prefill-aware expert placement algorithm that achieves up to 1.25$\times$ speedup on MoE computation. Our work presents the first comprehensive data-centric analysis of large-scale MoE models together with a concrete design study applying the learned lessons. Our profiling traces are publicly available at \href{https://huggingface.co/datasets/core12345/MoE_expert_selection_trace}{\textcolor{blue}{https://huggingface.co/datasets/core12345/MoE\_expert\_selection\_trace}}.

cs.DC↗

arXiv2024

The Cardinality of Identifying Code Sets for Soccer Ball Graph with Application to Remote Sensing

In the context of satellite monitoring of the earth, we can assume that the surface of the earth is divided into a set of regions. We assume that the impact of a big social/environmental event spills into neighboring regions. Using Identifying Code Sets (ICSes), we can deploy sensors in such a way that the region in which an event takes place can be uniquely identified, even with fewer sensors than regions. As Earth is almost a sphere, we use a soccer ball as a model. We construct a Soccer Ball Graph (SBG), and provide human-oriented, analytical proofs that 1) the SBG has at least 26 ICSes of cardinality ten, implying that there are at least 26 different ways to deploy ten satellites to monitor the Earth and 2) that the cardinality of the minimum Identifying Code Set (MICS) for the SBG is at least nine. We then provide a machine-oriented formal proof that the cardinality of the MICS for the SBG is in fact ten, meaning that one must deploy at least ten satellites to monitor the Earth in the SBG model. We also provide machine-oriented proof that there are exactly 26 ICSes of cardinality ten for the SBG.

cs.AI↗

arXiv2019

A Novel Graph Analytic Approach to Monitor Terrorist Networks

Terrorist attacks all across the world have become a major source of concern for almost all national governments. The United States Department of State's Bureau of Counter-Terrorism, maintains a list of 66 terrorist organizations spanning the entire world. Actively monitoring a large number of organizations and their members, require considerable amounts of resources on the part of law enforcement agencies. Oftentimes, the law enforcement agencies do not have adequate resources to monitor these organizations and their members effectively. On multiple incidences of terrorist attacks in recent times across Europe, it has been observed that the perpetrators of the attack were in the suspect databases of the law enforcement authorities, but weren't under active surveillance at the time of the attack, due to resource limitations on the part of the authorities. As the suspect databases in various countries are very large, and it takes significant amount of technical and human resources to monitor a suspect in the database, monitoring all the suspects in the database may be an impossible task. In this paper, we propose a novel terror network monitoring approach that will significantly reduce the resource requirement of law enforcement authorities, but still provide the capability of uniquely identifying a suspect in case the suspect becomes active in planning a terrorist attack. The approach relies on the assumption that, when an individual becomes active in planning a terrorist attack, his/her friends/associates will have some inkling of the individuals plan. Accordingly, even if the individual is not under active surveillance by the authorities, but the individual's friends/associates are, then the individual planning the attack can be uniquely identified. We apply our techniques on various real-world terror network datasets and show the effectiveness of our approach.

cs.SI↗

arXiv2018

Decentralized Cooperative Planning for Automated Vehicles with Hierarchical Monte Carlo Tree Search

Today's automated vehicles lack the ability to cooperate implicitly with others. This work presents a Monte Carlo Tree Search (MCTS) based approach for decentralized cooperative planning using macro-actions for automated vehicles in heterogeneous environments. Based on cooperative modeling of other agents and Decoupled-UCT (a variant of MCTS), the algorithm evaluates the state-action-values of each agent in a cooperative and decentralized manner, explicitly modeling the interdependence of actions between traffic participants. Macro-actions allow for temporal extension over multiple time steps and increase the effective search depth requiring fewer iterations to plan over longer horizons. Without predefined policies for macro-actions, the algorithm simultaneously learns policies over and within macro-actions. The proposed method is evaluated under several conflict scenarios, showing that the algorithm can achieve effective cooperative planning with learned macro-actions in heterogeneous environments.

cs.AI↗

arXiv2017

CPG-Based Control Scheme for Quadruped Robot to Withstand the Lateral Impact

This paper aims to present a stability control strategy for quadruped robot under lateral impact with the help of lateral trot. We firstly propose five necessary conditions for keeping balance. The classical four-neuron Central Pattern Generator (CPG) network with Hopf oscillators is then extended to eight-neuron network with four more trigger-enabled neurons, which controls the lateral trot. With proper adjustment of network's parameters, such network can coordinate the lateral and longitudinal trot gait. Based on Zero Movement Point (ZMP) theory, the robot is modeled as an inverted pendulum to plan the Center of Gravity (CoG) position and calculate the needed lateral step length. The simulation shows that the lateral acceleration of the quadruped robot after lateral impact regains to the normal range in a short time. Comparison shows that the maximal lateral impact that robot can resist increases about 125% from 0.72g to 1.55g.

cs.RO↗

arXiv2017

On Robustness in Multilayer Interdependent Network

Critical Infrastructures like power and communication networks are highly interdependent on each other for their full functionality. Many significant research have been pursued to model the interdependency and failure analysis of these interdependent networks. However, most of these models fail to capture the complex interdependencies that might actually exist between the infrastructures. The \emph{Implicative Interdependency Model} that utilizes Boolean Logic to capture complex interdependencies was recently proposed which overcome the limitations of the existing models. A number of problems were studies based on this model. In this paper we study the \textit{Robustness} problem in Interdependent Power and Communication Network. The robustness is defined with respect to two parameters $K \in I^{+} \cup \{0\}$ and $ρ\in (0,1]$. We utilized the \emph{Implicative Interdependency Model} model to capture the complex interdependency between the two networks. The model classifies the interdependency relations into four cases. Computational complexity of the problem is analyzed for each of these cases. A polynomial time algorithm is designed for the first case that outputs the optimal solution. All the other cases are proved to be NP-complete. An in-approximability bound is provided for the third case. For the general case we formulate an Integer Linear Program to get the optimal solution and a polynomial time heuristic. The applicability of the heuristic is evaluated using power and communication network data of Maricopa County, Arizona. The experimental results showed that the heuristic almost always produced near optimal value of parameter $K$ for $ρ< 0.42$.

cs.NI↗

arXiv2014

On the Entity Hardening Problem in Multi-layered Interdependent Networks

The power grid and the communication network are highly interdependent on each other for their well being. In recent times the research community has shown significant interest in modeling such interdependent networks and studying the impact of failures on these networks. Although a number of models have been proposed, many of them are simplistic in nature and fail to capture the complex interdependencies that exist between the entities of these networks. To overcome the limitations, recently an Implicative Interdependency Model that utilizes Boolean Logic, was proposed and a number of problems were studied. In this paper we study the entity hardening problem, where by entity hardening we imply the ability of the network operator to ensure that an adversary (be it Nature or human) cannot take a network entity from operative to inoperative state. Given that the network operator with a limited budget can only harden k entities, the goal of the entity hardening problem is to identify the set of k entities whose hardening will ensure maximum benefit for the operator, i.e. maximally reduce the ability of the adversary to degrade the network. We show that the problem is solvable in polynomial time for some cases, whereas for others it is NP-complete. We provide the optimal solution using ILP, and propose a heuristic approach to solve the problem. We evaluate the efficacy of our heuristic using power and communication network data of Maricopa County, Arizona. The experiments show that our heuristic almost always produces near optimal results.

cs.NI↗