arXiv subjects

Zhenyu Gu

Publications and source records attributed to Zhenyu Gu.

16 recordsLinked to original sources

arXiv2026

Instella-MoE Technical Report

In this work, we introduce Instella-MoE, a fully open Mixture-of-Experts (MoE) language model with 16 billion total parameters and 2.8 billion active parameters per token, trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. Instella-MoE combines a sparsely activated MoE design with architectural and system-level innovations, including Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity, enabling efficient large-scale training and inference. The model is developed through a multi-stage pipeline comprising pre-training, mid-training, long-context extension, supervised fine-tuning with feedback-driven data curation, direct preference optimization, and reinforcement learning with Multi-Teacher On-Policy Distillation. Instella-MoE achieves an average score of 76.7 across standard pre-training benchmarks, outperforming prior fully open models including OLMo-3-7B, SmolLM3-3B, and OLMoE-1B-7B, while remaining competitive with open-weight MoE and dense baselines at comparable active-parameter scales, including Moonlight-16B-A3B and Qwen3.5-4B. After post-training, our final Think checkpoint achieves an average score of 73.2 across instruction-following, reasoning, math, coding, and chat benchmarks, outperforming both fully open and open-weight models with comparable or larger active parameter counts in our evaluation. To support transparent and reproducible research, we release the complete Instella-MoE model flow, including model weights, training configurations, data mixtures, and training code. Together, these contributions establish Instella-MoE a strong, fully open foundation for efficient, high-performing MoE models and reproducible research.

cs.CL↗

arXiv2026

Representation Learning in Diffusion and Flow-based Model: An Application Aspect

Diffusion models and flow-based models have recently become the dominant paradigms in generative modeling, largely due to their ability to learn rich, multi-level visual representations through large-scale training. This creates a bidirectional relationship between generative models and representation learning: improving representation learning enhances generation quality, while the learned representations can be leveraged for broader understanding tasks. This survey systematically explores this interplay with a focus on applications. We propose a three-tier progressive framework that organizes existing works from three perspectives: using representation learning to improve generative capabilities, exploiting generative models to extract representations for perception tasks, and ultimately moving toward general-purpose unified applications. We systematically categorize representative methods across a wide range of downstream tasks, including image classification, dense visual prediction, instance-level perception, and annotation-scarce scenarios. By providing a unified taxonomy and identifying key challenges, this survey aims to clarify the underlying logic of current research and suggest promising directions for future exploration. We hope this work can serve as a valuable reference for researchers interested in harnessing the representation power of generative models for applications beyond generation.

cs.CV↗

arXiv2026

FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications

Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and intra-model parallelism. Existing serving systems and auto-parallelism compilers commit to limited transformations and fixed workload assumptions, so achieving high performance on a new application requires hand-crafting an efficient implementation. We present FlashRT, an agent harness that guides coding agents to lift simple developer-written reference implementations into optimized multi-GPU deployments that flexibly weigh target metrics like latency and throughput. Using a new chain-of-program paradigm, FlashRT directs a generic coding agent through a multi-pass transformation process where an agent transforms the reference into an intermediate representation (IR) to capture data dependencies and persistent-state scopes, validates this IR via a sequential interpreter, and performs static analyses to identify candidate transformations. Then, the agent iteratively implements, verifies, and benchmarks each candidate under a measurement-gated optimization loop to produce effective deployments that span different hardware budgets. Across various applications, including video world models and multimodal LLMs, FlashRT converts reference implementations into highly efficient deployments, delivering up to ~70x latency reduction and 2.8x throughput improvement on NVIDIA B200 GPUs. On AMD MI355X GPUs, FlashRT matches the peak latency reduction while increasing peak throughput improvement to 3.6x, demonstrating that agent-driven optimization can be more scalable on platforms with less mature expert optimization. In fact, for Qwen3-Omni text-to-audio inference, FlashRT reduces response latency by 65% compared to the expert vLLM-Omni implementation on AMD MI355X.

cs.LG↗

arXiv2026

Arbor: Tree Search as a Cognition Layer for Autonomous Agents

Arbor is a multi-agent framework that introduces structured tree search as a cognition layer for autonomous agents operating in large, stateful action spaces. Prior autonomous optimization systems operate on isolated targets with stateless evaluation. Arbor instead maintains an explicit search tree of scored hypotheses that serves as the shared working memory across agents, evolving with every measurement, treating failures as diagnostic signal that reshapes subsequent exploration, and expanding as prior successes shift the bottleneck distribution. We validate Arbor on full-stack LLM inference optimization, a domain where achieving peak performance has historically required coordinated effort from engineering teams across the application, framework, compiler, kernel, and hardware stack. Arbor pairs an Orchestrator agent, which drives optimization by delegating to Domain Specialists across the inference stack, with a Critic agent that safeguards stability through root-cause analysis, introspection, and measurement validation -- a checks-and-balances architecture where neither agent can unilaterally drive the system. Agent capabilities are decomposed into hard skills (domain expertise) and soft skills (coordination protocols that determine how contributions compose), enabling fully autonomous multi-day campaigns. Arbor achieves up to 193% inference throughput-latency Pareto improvement over vendor-optimized baselines, while a single agent without the harness plateaus at +33% throughput improvement and crashes irrecoverably within hours. Arbor generalizes to multiple generations of hardware platform, and run-to-run variance is within 2 percentage points demonstrating that the method is hardware-agnostic and reproducible.

cs.AI↗

arXiv2026

AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents

GPU kernel optimization is increasingly critical for efficient deep learning systems, but writing high-performance kernels still requires substantial low-level expertise. Recent AI coding agents can iteratively read code, invoke compilers and profilers, and refine implementations, yet existing kernel benchmarks evaluate single LLM calls rather than full agent workflows, and none include both kernel-to-kernel optimization and unseen-configuration generalization testing. We present AgentKernelArena, an open-source benchmark for measuring AI coding agents on GPU kernel optimization. The benchmark contains 196 tasks spanning HIP-to-HIP optimization, Triton-to-Triton optimization, and PyTorch-to-HIP translation, and evaluates complete agent workflows in isolated workspaces using gated compilation, correctness, and performance checks, centralized scoring and an unseen-configuration generalization protocol that tests whether optimizations transfer to input configurations the agent never observed. Across production agents including Cursor Agent, Claude Code, and Codex Agent, we find near-perfect compilation and high correctness rates on most task categories, with the strongest configurations achieving mean speedups of up to 6.89x on PyTorch-to-HIP, 6.69x on HIP-to-HIP, and 2.13x on Triton-to-Triton tasks. Our unseen-configuration evaluation shows that HIP-to-HIP and Triton-to-Triton optimizations largely transfer to unseen input shapes, while PyTorch-to-HIP exhibits substantial correctness drops, indicating that agents generating kernels from scratch frequently hardcode shape-specific assumptions. AgentKernelArena is designed as a modular, extensible framework for rigorous evaluation of agentic GPU kernel optimization across agents, tasks, and hardware targets.

cs.CL↗

arXiv2026

Uncovering Intra-expert Activation Sparsity for Efficient Mixture-of-Expert Model Execution

Mixture of Experts (MoE) architecture has become the standard for state-of-the-art large language models, owing to its computational efficiency through sparse expert activation. However, sparsity through finer expert granularity is becoming increasingly difficult to achieve due to fundamental training challenges such as expert collapse and load imbalance. In this work, we explore and leverage intra-expert activation sparsity as a complementary and underexplored dimension of sparsity in MoE models. Surprisingly, substantial intra-expert sparsity is readily available in existing pre-trained MoE models, without any modification to the activation function or model parameters, providing up to 90% sparsity within each expert without significant accuracy loss. We explore intra-expert activation sparsity across eight off-the-shelf MoE models ranging from 1B to 400B parameters, and extend the MoE execution pipeline of vLLM to leverage intra-expert activation sparsity by skipping the computations of inactive neurons, on top of its existing optimizations, achieving up to 2.5 times speedup in MoE layer execution and 1.2 times end-to-end speedup compared to the original dense vLLM baseline.

cs.LG↗

arXiv2025

UCCL-EP: Portable Expert-Parallel Communication

Mixture-of-Experts (MoE) workloads rely on expert parallelism (EP) to achieve high GPU efficiency. State-of-the-art EP communication systems such as DeepEP demonstrate strong performance but exhibit poor portability across heterogeneous GPU and NIC platforms. The poor portability is rooted in architecture: GPU-initiated token-level RDMA communication requires tight vertical integration between GPUs and NICs, e.g., GPU writes to NIC driver/MMIO interfaces. We present UCCL-EP, a portable EP communication system that delivers DeepEP-level performance across heterogeneous GPU and NIC hardware. UCCL-EP replaces GPU-initiated RDMA with a high-throughput GPU-CPU control channel: compact token-routing commands are transferred to multithreaded CPU proxies, which then issue GPUDirect RDMA operations on behalf of GPUs. UCCL-EP further emulates various ordering semantics required by specialized EP communication modes using RDMA immediate data, enabling correctness on NICs that lack such ordering, e.g., AWS EFA. We implement UCCL-EP on NVIDIA and AMD GPUs with EFA and Broadcom NICs. On EFA, it outperforms the best existing EP solution by up to $2.1\times$ for dispatch and combine throughput. On NVIDIA-only platform, UCCL-EP achieves comparable performance to the original DeepEP. UCCL-EP also improves token throughput on SGLang by up to 40% on the NVIDIA+EFA platform, and improves DeepSeek-V3 training throughput over the AMD Primus/Megatron-LM framework by up to 45% on a 16-node AMD+Broadcom platform.

cs.DC↗

arXiv2025

Training Foundation Models on a Full-Stack AMD Platform: Compute, Networking, and System Design

We report on the first large-scale mixture-of-experts (MoE) pretraining study on pure AMD hardware, utilizing both MI300X GPUs and Pollara networking. We distill practical guidance for both systems and model design. On the systems side, we deliver a comprehensive cluster and networking characterization: microbenchmarks for all core collectives (all-reduce, reduce-scatter, all-gather, broadcast) across message sizes and GPU counts over Pollara. To our knowledge, this is the first at this scale. We further provide MI300X microbenchmarks on kernel sizing and memory bandwidth to inform model design. On the modeling side, we introduce and apply MI300X-aware transformer sizing rules for attention and MLP blocks and justify MoE widths that jointly optimize training throughput and inference latency. We describe our training stack in depth, including often-ignored utilities such as fault-tolerance and checkpoint-reshaping, as well as detailed information on our training recipe. We also provide a preview of our model architecture and base model - ZAYA1 (760M active, 8.3B total parameters MoE, available at https://huggingface.co/Zyphra/ZAYA1-base) - which will be further improved upon in forthcoming papers. ZAYA1-base achieves performance comparable to leading base models such as Qwen3-4B and Gemma3-12B at its scale and larger, and outperforms models including Llama-3-8B and OLMoE across reasoning, mathematics, and coding benchmarks. Together, these results demonstrate that the AMD hardware, network, and software stack are mature and optimized enough for competitive large-scale pretraining.

cs.CL↗

arXiv2025

Rectified Noise: A Generative Model Using Positive-incentive Noise

Rectified Flow (RF) has been widely used as an effective generative model. Although RF is primarily based on probability flow Ordinary Differential Equations (ODE), recent studies have shown that injecting noise through reverse-time Stochastic Differential Equations (SDE) for sampling can achieve superior generative performance. Inspired by Positive-incentive Noise (pi-noise), we propose an innovative generative algorithm to train pi-noise generators, namely Rectified Noise (RN), which improves the generative performance by injecting pi-noise into the velocity field of pre-trained RF models. After introducing the Rectified Noise pipeline, pre-trained RF models can be efficiently transformed into pi-noise generators. We validate Rectified Noise by conducting extensive experiments across various model architectures on different datasets. Notably, we find that: (1) RF models using Rectified Noise reduce FID from 10.16 to 9.05 on ImageNet-1k. (2) The models of pi-noise generators achieve improved performance with only 0.39% additional training parameters.

cs.LG↗

arXiv2025

RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation

Simulation-based data synthesis has emerged as a powerful paradigm for advancing real-world robotic manipulation. Yet existing datasets remain insufficient for robust bimanual manipulation due to (1) the lack of scalable task generation methods and (2) oversimplified simulation environments. We present RoboTwin 2.0, a scalable framework for automated, large-scale generation of diverse and realistic data, together with unified evaluation protocols for dual-arm manipulation. At its core is RoboTwin-OD, an object library of 731 instances across 147 categories with semantic and manipulation-relevant annotations. Building on this, we design an expert data synthesis pipeline that leverages multimodal language models (MLLMs) and simulation-in-the-loop refinement to automatically generate task-level execution code. To improve sim-to-real transfer, RoboTwin 2.0 applies structured domain randomization along five axes: clutter, lighting, background, tabletop height, and language, enhancing data diversity and policy robustness. The framework is instantiated across 50 dual-arm tasks and five robot embodiments. Empirically, it yields a 10.9% gain in code generation success rate. For downstream policy learning, a VLA model trained with synthetic data plus only 10 real demonstrations achieves a 367% relative improvement over the 10-demo baseline, while zero-shot models trained solely on synthetic data obtain a 228% gain. These results highlight the effectiveness of RoboTwin 2.0 in strengthening sim-to-real transfer and robustness to environmental variations. We release the data generator, benchmark, dataset, and code to support scalable research in robust bimanual manipulation. Project Page: https://robotwin-platform.github.io/, Code: https://github.com/robotwin-Platform/robotwin/.

cs.RO↗

arXiv2024

Yi: Open Foundation Models by 01.AI

We introduce the Yi model family, a series of language and multimodal models that demonstrate strong multi-dimensional capabilities. The Yi model family is based on 6B and 34B pretrained language models, then we extend them to chat models, 200K long context models, depth-upscaled models, and vision-language models. Our base models achieve strong performance on a wide range of benchmarks like MMLU, and our finetuned chat models deliver strong human preference rate on major evaluation platforms like AlpacaEval and Chatbot Arena. Building upon our scalable super-computing infrastructure and the classical transformer architecture, we attribute the performance of Yi models primarily to its data quality resulting from our data-engineering efforts. For pretraining, we construct 3.1 trillion tokens of English and Chinese corpora using a cascaded data deduplication and quality filtering pipeline. For finetuning, we polish a small scale (less than 10K) instruction dataset over multiple iterations such that every single instance has been verified directly by our machine learning engineers. For vision-language, we combine the chat language model with a vision transformer encoder and train the model to align visual representations to the semantic space of the language model. We further extend the context length to 200K through lightweight continual pretraining and demonstrate strong needle-in-a-haystack retrieval performance. We show that extending the depth of the pretrained checkpoint through continual pretraining further improves performance. We believe that given our current results, continuing to scale up model parameters using thoroughly optimized data will lead to even stronger frontier models.

cs.CL↗

arXiv2023

Display object alignment may influence location recall in unexpected ways

There is a presumption in human-computer interaction that laying out menus and most other material in neat rows and columns helps users get work done. The rule has been so implicit in the field of design as to allow for no debate. However, the idea that perfect collinearity benefits creates an advantage for both either search and or recall has rarely been tested. Drawing from separate branches of cognitive literature, we tested a minimal brainstorming interface with either aligned or eccentrically arranged layouts on 96 college students. Incidental exact recall of recently worked locations improved in the eccentric condition. And in both conditions there were frequent near-miss recall errors to neighboring aligned objects and groups of objects. Further analysis found only marginal performance advantages specifically for females with the eccentric design. However, NASA-TLX subjective measures showed that in eccentric, females reported higher performance, less effort, and yet also higher frustration; while males reported lower performance with about the same effort, and lower frustration.

cs.HC↗

arXiv2023

A foolish consistency? Aligning interface objects hinders location recall and may induce collinear errors

During our nearly constant use of digital devices, perhaps our most frequent need is to visually identify icons representing our content and invoke the actions to manipulate them. Almost since the inception of user interface design in the 1970s, with rare exception it has become the tendency for programmers to prescribe the arrangement these things in uniform rectilinear rows and columns. This was imported from theories for print design and ultimately brought into widespread practice for graphical user interfaces (GUIs). Whether consistent rectilinearity actually does better than less rectilinear arrangements to maximize selection efficiency has not been challenged on considerations of speed or any other measure. In a series of four experiments, we explore how alignment may in fact discourage easy recallability of screen object locations and hence increase search intensity. A second objective is to present a methodological model where we deliberately attempt to begin with psychophysical cognitive evidence at the environmental schematic low end (beginning with contextual cueing paradigm), then move progressively upwards in naturalism in the experiments to something that approximates actual human work at the higher end, all along attempting to keep one important environmental property constant. Two experiments using contextual cueing paradigm confirm that collinearly aligned arrays do not encourage recallability of location, while noncollinear arrays appear to create traces that can be recalled automatically. Two other experiments give further demonstration of explicit recollection and location recall, showing that collinear arrangements may in fact induce location recall errors to neighboring collinear objects. We discuss surrounding theoretical, historical, and practical questions.

cs.HC↗

arXiv2022

Energon: Towards Efficient Acceleration of Transformers Using Dynamic Sparse Attention

In recent years, transformer models have revolutionized Natural Language Processing (NLP) and shown promising performance on Computer Vision (CV) tasks. Despite their effectiveness, transformers' attention operations are hard to accelerate due to the complicated data movement and quadratic computational complexity, prohibiting the real-time inference on resource-constrained edge-computing platforms. To tackle this challenge, we propose Energon, an algorithm-architecture co-design approach that accelerates various transformers using dynamic sparse attention. With the observation that attention results only depend on a few important query-key pairs, we propose a Mix-Precision Multi-Round Filtering (MP-MRF) algorithm to dynamically identify such pairs at runtime. We adopt low bitwidth in each filtering round and only use high-precision tensors in the attention stage to reduce overall complexity. By this means, we significantly mitigate the computational cost with negligible accuracy loss. To enable such an algorithm with lower latency and better energy efficiency, we also propose an Energon co-processor architecture. Elaborated pipelines and specialized optimizations jointly boost the performance and reduce power consumption. Extensive experiments on both NLP and CV benchmarks demonstrate that Energon achieves $168\times$ and $8.7\times$ geo-mean speedup and up to $10^4\times$ and $10^3\times$ energy reduction compared with Intel Xeon 5220 CPU and NVIDIA V100 GPU. Compared to state-of-the-art attention accelerators SpAtten and $A^3$, Energon also achieves $1.7\times, 1.25\times$ speedup and $1.6 \times, 1.5\times $ higher energy efficiency.

cs.AR↗

arXiv2021

Distribution Adaptive INT8 Quantization for Training CNNs

Researches have demonstrated that low bit-width (e.g., INT8) quantization can be employed to accelerate the inference process. It makes the gradient quantization very promising since the backward propagation requires approximately twice more computation than forward one. Due to the variability and uncertainty of gradient distribution, a lot of methods have been proposed to attain training stability. However, most of them ignore the channel-wise gradient distributions and the impact of gradients with different magnitudes, resulting in the degradation of final accuracy. In this paper, we propose a novel INT8 quantization training framework for convolutional neural network to address the above issues. Specifically, we adopt Gradient Vectorized Quantization to quantize the gradient, based on the observation that layer-wise gradients contain multiple distributions along the channel dimension. Then, Magnitude-aware Clipping Strategy is introduced by taking the magnitudes of gradients into consideration when minimizing the quantization error, and we present a theoretical derivation to solve the quantization parameters of different distributions. Experimental results on broad range of computer vision tasks, such as image classification, object detection and video classification, demonstrate that the proposed Distribution Adaptive INT8 Quantization training method has achieved almost lossless training accuracy for different backbones, including ResNet, MobileNetV2, InceptionV3, VGG and AlexNet, which is superior to the state-of-the-art techniques. Moreover, we further implement the INT8 kernel that can accelerate the training iteration more than 200% under the latest Turing architecture, i.e., our method excels on both training accuracy and speed.

cs.CV↗

arXiv2018

Predicting Webpage Aesthetics with Heatmap Entropy

Today, eye trackers are extensively used in user interface evaluations. However, it's still hard to analyze and interpret eye tracking data from the aesthetic point of view. To find quantitative links between eye movements and aesthetic experience, we tracked 30 observers' initial landings for 40 web pages (each displayed for 3 seconds). The web pages were also rated based on the observers' subjective aesthetic judgments. Shannon entropy was introduced to analyze the eye-tracking data. The result shows that the heatmap entropy (visual attention entropy, VAE) is highly correlated with the observers' aesthetic judgements of the web pages. Its improved version, relative VAE (rVAE), has a more significant correlation with the perceived aesthetics. (r=-0.65, F= 26.84, P$<$0.0001). This single metric alone can distinguish between good- and bad-looking pages with an approximate 85\% accuracy. Further investigation reveals that the performance of both VAE and rVAE became stable after 1 second. The curves indicate that their performances could be better, if the tracking time was extended beyond 3 seconds.

cs.HC↗