Dan Zhao

Zhiyao Xu, Aoxue Liu, Zhanjie Ding et al.

Sparsely activated Mixture-of-Experts (MoE) models scale capacity via conditional computation, but distributed inference suffers from cross-GPU expert communication and routing-induced load imbalance. Existing placement methods reduce this cost by co-locating frequently co-activated experts; however, they derive a single deployment plan from globally aggregated routing traces, thereby averaging away the heterogeneous, task-specific co-activation patterns that actually drive communication in multi-task serving. We observe that expert co-activation is strongly task-conditioned: pairs tightly coupled in one task family are often uncorrelated in another, so effective deployment should group experts by task-aware co-activation rather than by a task-agnostic average. Based on this insight, we propose \emph{Task-Aware Coactivation Grouping} (TACG), a deployment-time framework that uses family-specific dispatch and co-activation traces to derive per-expert task-family preferences, reweights the co-activation graph so that intra-family locality dominates grouping, and assigns each expert to a primary GPU under exact capacity constraints. To keep the static placement robust under online workload skew, we further introduce \emph{Generic Expert Shared Replication} (GESR), a lightweight companion that identifies generic experts with consistently central co-activation profiles, replicates them across a small set of secondary GPUs, and applies locality- and load-aware selection at serving time. Experiments on three representative open-source MoE models demonstrate that our framework reduces the average communication cost by 31.39\% over the baseline, while preserving an average Jain fairness index of 0.9975. This advantage persists even under severe distribution shifts in the inference data, consistently outperforming strong baselines.

Rogerio Bonatti, Dan Zhao, Francesco Bonacci et al.

Large language models (LLMs) show remarkable potential to act as computer agents, enhancing human productivity and software accessibility in multi-modal tasks that require planning and reasoning. However, measuring agent performance in realistic environments remains a challenge since: (i) most benchmarks are limited to specific modalities or domains (e.g. text-only, web navigation, Q&A, coding) and (ii) full benchmark evaluations are slow (on order of magnitude of days) given the multi-step sequential nature of tasks. To address these challenges, we introduce the Windows Agent Arena: a reproducible, general environment focusing exclusively on the Windows operating system (OS) where agents can operate freely within a real Windows OS and use the same wide range of applications, tools, and web browsers available to human users when solving tasks. We adapt the OSWorld framework (Xie et al., 2024) to create 150+ diverse Windows tasks across representative domains that require agent abilities in planning, screen understanding, and tool usage. Our benchmark is scalable and can be seamlessly parallelized in Azure for a full benchmark evaluation in as little as 20 minutes. To demonstrate Windows Agent Arena's capabilities, we also introduce a new multi-modal agent, Navi. Our agent achieves a success rate of 19.5% in the Windows domain, compared to 74.5% performance of an unassisted human. Navi also demonstrates strong performance on another popular web-based benchmark, Mind2Web. We offer extensive quantitative and qualitative analysis of Navi's performance, and provide insights into the opportunities for future research in agent development and data generation using Windows Agent Arena. Webpage: https://microsoft.github.io/WindowsAgentArena Code: https://github.com/microsoft/WindowsAgentArena

Dan Zhao, Nathan C. Frey, Joseph McDonald et al. · berkeley

As research and practice in artificial intelligence (A.I.) grow in leaps and bounds, the resources necessary to sustain and support their operations also grow at an increasing pace. While innovations and applications from A.I. have brought significant advances, from applications to vision and natural language to improvements to fields like medical imaging and materials engineering, their costs should not be neglected. As we embrace a world with ever-increasing amounts of data as well as research and development of A.I. applications, we are sure to face an ever-mounting energy footprint to sustain these computational budgets, data storage needs, and more. But, is this sustainable and, more importantly, what kind of setting is best positioned to nurture such sustainable A.I. in both research and practice? In this paper, we outline our outlook for Green A.I. -- a more sustainable, energy-efficient and energy-aware ecosystem for developing A.I. across the research, computing, and practitioner communities alike -- and the steps required to arrive there. We present a bird's eye view of various areas for potential changes and improvements from the ground floor of AI's operational and hardware optimizations for datacenters/HPCs to the current incentive structures in the world of A.I. research and practice, and more. We hope these points will spur further discussion, and action, on some of these issues and their potential solutions.

Siddharth Samsi, Dan Zhao, Joseph McDonald et al.

Large language models (LLMs) have exploded in popularity due to their new generative capabilities that go far beyond prior state-of-the-art. These technologies are increasingly being leveraged in various domains such as law, finance, and medicine. However, these models carry significant computational challenges, especially the compute and energy costs required for inference. Inference energy costs already receive less attention than the energy costs of training LLMs -- despite how often these large models are called on to conduct inference in reality (e.g., ChatGPT). As these state-of-the-art LLMs see increasing usage and deployment in various domains, a better understanding of their resource utilization is crucial for cost-savings, scaling performance, efficient hardware usage, and optimal inference strategies. In this paper, we describe experiments conducted to study the computational and energy utilization of inference with LLMs. We benchmark and conduct a preliminary analysis of the inference performance and inference energy costs of different sizes of LLaMA -- a recent state-of-the-art LLM -- developed by Meta AI on two generations of popular GPUs (NVIDIA V100 \& A100) and two datasets (Alpaca and GSM8K) to reflect the diverse set of tasks/benchmarks for LLMs in research and practice. We present the results of multi-node, multi-GPU inference using model sharding across up to 32 GPUs. To our knowledge, our work is the one of the first to study LLM inference performance from the perspective of computational and energy resources at this scale.

Zhiwen Yang, Jiayin Li, Hui Zhang et al.

Transformers have revolutionized medical image restoration, but the quadratic complexity still poses limitations for their application to high-resolution medical images. The recent advent of the Receptance Weighted Key Value (RWKV) model in the natural language processing field has attracted much attention due to its ability to process long sequences efficiently. To leverage its advanced design, we propose Restore-RWKV, the first RWKV-based model for medical image restoration. Since the original RWKV model is designed for 1D sequences, we make two necessary modifications for modeling spatial relations in 2D medical images. First, we present a recurrent WKV (Re-WKV) attention mechanism that captures global dependencies with linear computational complexity. Re-WKV incorporates bidirectional attention as basic for a global receptive field and recurrent attention to effectively model 2D dependencies from various scan directions. Second, we develop an omnidirectional token shift (Omni-Shift) layer that enhances local dependencies by shifting tokens from all directions and across a wide context range. These adaptations make the proposed Restore-RWKV an efficient and effective model for medical image restoration. Even a lightweight variant of Restore-RWKV, with only 1.16 million parameters, achieves comparable or even superior results compared to existing state-of-the-art (SOTA) methods. Extensive experiments demonstrate that the resulting Restore-RWKV achieves SOTA performance across a range of medical image restoration tasks, including PET image synthesis, CT image denoising, MRI image super-resolution, and all-in-one medical image restoration. Code is available at: https://github.com/Yaziwel/Restore-RWKV.

Han Li, Dan Zhao, Jianyang Zeng

Designing accurate deep learning models for molecular property prediction plays an increasingly essential role in drug and material discovery. Recently, due to the scarcity of labeled molecules, self-supervised learning methods for learning generalizable and transferable representations of molecular graphs have attracted lots of attention. In this paper, we argue that there exist two major issues hindering current self-supervised learning methods from obtaining desired performance on molecular property prediction, that is, the ill-defined pre-training tasks and the limited model capacity. To this end, we introduce Knowledge-guided Pre-training of Graph Transformer (KPGT), a novel self-supervised learning framework for molecular graph representation learning, to alleviate the aforementioned issues and improve the performance on the downstream molecular property prediction tasks. More specifically, we first introduce a high-capacity model, named Line Graph Transformer (LiGhT), which emphasizes the importance of chemical bonds and is mainly designed to model the structural information of molecular graphs. Then, a knowledge-guided pre-training strategy is proposed to exploit the additional knowledge of molecules to guide the model to capture the abundant structural and semantic information from large-scale unlabeled molecular graphs. Extensive computational tests demonstrated that KPGT can offer superior performance over current state-of-the-art methods on several molecular property prediction tasks.

Zhiwen Yang, Haowei Chen, Ziniu Qian et al.

Transformer-based methods have demonstrated impressive results in medical image restoration, attributed to the multi-head self-attention (MSA) mechanism in the spatial dimension. However, the majority of existing Transformers conduct attention within fixed and coarsely partitioned regions (\text{e.g.} the entire image or fixed patches), resulting in interference from irrelevant regions and fragmentation of continuous image content. To overcome these challenges, we introduce a novel Region Attention Transformer (RAT) that utilizes a region-based multi-head self-attention mechanism (R-MSA). The R-MSA dynamically partitions the input image into non-overlapping semantic regions using the robust Segment Anything Model (SAM) and then performs self-attention within these regions. This region partitioning is more flexible and interpretable, ensuring that only pixels from similar semantic regions complement each other, thereby eliminating interference from irrelevant regions. Moreover, we introduce a focal region loss to guide our model to adaptively focus on recovering high-difficulty regions. Extensive experiments demonstrate the effectiveness of RAT in various medical image restoration tasks, including PET image synthesis, CT image denoising, and pathological image super-resolution. Code is available at \href{https://github.com/Yaziwel/Region-Attention-Transformer-for-Medical-Image-Restoration.git}{https://github.com/RAT}.

Yaosen Min, Ye Wei, Peizhuo Wang et al.

Accurate prediction of protein-ligand binding affinities is an essential challenge in structure-based drug design. Despite recent advances in data-driven methods for affinity prediction, their accuracy is still limited, partially because they only take advantage of static crystal structures while the actual binding affinities are generally determined by the thermodynamic ensembles between proteins and ligands. One effective way to approximate such a thermodynamic ensemble is to use molecular dynamics (MD) simulation. Here, an MD dataset containing 3,218 different protein-ligand complexes is curated, and Dynaformer, a graph-based deep learning model is further developed to predict the binding affinities by learning the geometric characteristics of the protein-ligand interactions from the MD trajectories. In silico experiments demonstrated that the model exhibits state-of-the-art scoring and ranking power on the CASF-2016 benchmark dataset, outperforming the methods hitherto reported. Moreover, in a virtual screening on heat shock protein 90 (HSP90) using Dynaformer, 20 candidates are identified and their binding affinities are further experimentally validated. Dynaformer displayed promising results in virtual drug screening, revealing 12 hit compounds (two are in the submicromolar range), including several novel scaffolds. Overall, these results demonstrated that the approach offer a promising avenue for accelerating the early drug discovery process.

Zhuoying Zhao, Ziling Tan, Pinghui Mo et al.

This paper proposes a special-purpose system to achieve high-accuracy and high-efficiency machine learning (ML) molecular dynamics (MD) calculations. The system consists of field programmable gate array (FPGA) and application specific integrated circuit (ASIC) working in heterogeneous parallelization. To be specific, a multiplication-less neural network (NN) is deployed on the non-von Neumann (NvN)-based ASIC (SilTerra 180 nm process) to evaluate atomic forces, which is the most computationally expensive part of MD. All other calculations of MD are done using FPGA (Xilinx XC7Z100). It is shown that, to achieve similar-level accuracy, the proposed NvN-based system based on low-end fabrication technologies (180 nm) is 1.6x faster and 10^2-10^3x more energy efficiency than state-of-the-art vN based MLMD using graphics processing units (GPUs) based on much more advanced technologies (12 nm), indicating superiority of the proposed NvN-based heterogeneous parallel architecture.

Tianyi Chen, Sihan Chen, Xiaoyi Qu et al.

Quantization-aware training (QAT) is essential for deploying large models under strict memory and latency constraints, yet achieving stable and robust optimization at ultra-low bitwidths remains challenging. Common approaches based on the straight-through estimator (STE) or soft quantizers often suffer from gradient mismatch, instability, or high computational overhead. As such, we propose StableQAT, a unified and efficient QAT framework that stabilizes training in ultra low-bit settings via a novel, lightweight, and theoretically grounded surrogate for backpropagation derived from a discrete Fourier analysis of the rounding operator. StableQAT strictly generalizes STE as the latter arises as a special case of our more expressive surrogate family, yielding smooth, bounded, and inexpensive gradients that improve QAT training performance and stability across various hyperparameter choices. In experiments, StableQAT exhibits stable and efficient QAT at 2-4 bit regimes, demonstrating improved training stability, robustness, and superior performance with negligible training overhead against standard QAT techniques. Our code is available at https://github.com/microsoft/StableQAT.

Nancy Lau, Louis Sloot, Jyoutir Raj et al.

Large language models (LLMs) are increasingly being deployed as software engineering agents that autonomously contribute to repositories. A major benefit these agents present is their ability to find and patch security vulnerabilities in the codebases they oversee. To estimate the capability of agents in this domain, we introduce ZeroDayBench, a benchmark where LLM agents find and patch 22 novel critical vulnerabilities in open-source codebases. We focus our efforts on three popular frontier agentic LLMs: GPT-5.2, Claude Sonnet 4.5, and Grok 4.1. We find that frontier LLMs are not yet capable of autonomously solving our tasks and observe some behavioral patterns that suggest how these models can be improved in the domain of proactive cyberdefense.

Kaira Samuel, Matthew LaRosa, Kyle McAlpin et al.

Artificial intelligence (AI) has enormous potential to improve Air Force pilot training by providing actionable feedback to pilot trainees on the quality of their maneuvers and enabling instructor-less flying familiarization for early-stage trainees in low-cost simulators. Historically, AI challenges consisting of data, problem descriptions, and example code have been critical to fueling AI breakthroughs. The Department of the Air Force-Massachusetts Institute of Technology AI Accelerator (DAF-MIT AI Accelerator) developed such an AI challenge using real-world Air Force flight simulator data. The Maneuver ID challenge assembled thousands of virtual reality simulator flight recordings collected by actual Air Force student pilots at Pilot Training Next (PTN). This dataset has been publicly released at Maneuver-ID.mit.edu and represents the first of its kind public release of USAF flight training data. Using this dataset, we have applied a variety of AI methods to separate "good" vs "bad" simulator data and categorize and characterize maneuvers. These data, algorithms, and software are being released as baselines of model performance for others to build upon to enable the AI ecosystem for flight simulator training.

Zixin Shu, Rui Hua, Dengying Yan et al.

Symptom phenotypes are one of the key types of manifestations for diagnosis and treatment of various disease conditions. However, the diversity of symptom terminologies is one of the major obstacles hindering the analysis and knowledge sharing of various types of symptom-related medical data particularly in the fields of Traditional Chinese Medicine (TCM). Objective: This study aimed to construct an Integrated Ontology of symptom phenotypes (ISPO) to support the data mining of Chinese EMRs and real-world study in TCM field. Methods: To construct an integrated ontology of symptom phenotypes (ISPO), we manually annotated classical TCM textbooks and large-scale Chinese electronic medical records (EMRs) to collect symptom terms with support from a medical text annotation system. Furthermore, to facilitate the semantic interoperability between different terminologies, we incorporated public available biomedical vocabularies by manual mapping between Chinese terms and English terms with cross-references to source vocabularies. In addition, we evaluated the ISPO using independent clinical EMRs to provide a high-usable medical ontology for clinical data analysis. Results: By integrating 78,696 inpatient cases of EMRs, 5 biomedical vocabularies, 21 TCM books and dictionaries, ISPO provides 3,147 concepts, 23,475 terms, and 55,552 definition or contextual texts. Adhering to the taxonomical structure of the related anatomical systems of symptom phenotypes, ISPO provides 12 top-level categories and 79 middle-level sub-categories. The validation of data analysis showed the ISPO has a coverage rate of 95.35%, 98.53% and 92.66% for symptom terms with occurrence rates of 0.5% in additional three independent curated clinical datasets, which can demonstrate the significant value of ISPO in mapping clinical terms to ontologies.

Zheng Hui, Yinheng Li, Dan zhao et al.

Graphical User Interface (GUI) tasks are vital for automating workflows such as software testing, user interface navigation. For users, the GUI is the most intuitive platform for interacting with a computer. Previous work identified a key challenge in developing visual GUI agents: GUI grounding - the ability to accurately locate screen elements based on instructions. However, most existing GUI agents rely on structured data formats like DOM or HTML files in training or inferencing, which are inaccessible across all applications, particular in a general desktop environments such as Windows OS. To address this, we introduce WinClick, a novel visual GUI agent developed in Windows platform. WinClick leverages screenshots to detect actionable regions. To overcome the challenge of GUI grounding, we enhance WinClick with GUI grounding pre-training and propose an LLM-based method for aligning GUI grounding data. Additionally, we introduce WinSpot, the first comprehensive benchmark for GUI grounding on Windows. Our experiments demonstrate that WinClick, combined with GUI grounding pre-training, significantly outperforms existing baselines, offering a scalable solution for GUI automation in desktop environments. WinSpot is publicly available at https://github.com/zackhuiiiii/WinSpot.

Dan Zhao

Works on implicit regularization have studied gradient trajectories during the optimization process to explain why deep networks favor certain kinds of solutions over others. In deep linear networks, it has been shown that gradient descent implicitly regularizes toward low-rank solutions on matrix completion/factorization tasks. Adding depth not only improves performance on these tasks but also acts as an accelerative pre-conditioning that further enhances this bias towards low-rankedness. Inspired by this, we propose an explicit penalty to mirror this implicit bias which only takes effect with certain adaptive gradient optimizers (e.g. Adam). This combination can enable a degenerate single-layer network to achieve low-rank approximations with generalization error comparable to deep linear networks, making depth no longer necessary for learning. The single-layer network also performs competitively or out-performs various approaches for matrix completion over a range of parameter and data regimes despite its simplicity. Together with an optimizer's inductive bias, our findings suggest that explicit regularization can play a role in designing different, desirable forms of regularization and that a more nuanced understanding of this interplay may be necessary.

Hanling Wang, Qing Li, Dan Zhao et al.

The rapid emergence of AI-powered applications is reshaping the role of the Internet. Users increasingly rely on the network to obtain intelligence services derived from large foundation models, rather than merely to reach remote endpoints or retrieve specific content. Today's dominant deployment paradigm for AI services remains cloud-centric, where user requests are transmitted to remote data centers for centralized inference. Although operationally convenient, this paradigm suffers from latency and jitter, heavy wide-area traffic, limited utilization of distributed heterogeneous compute resources, and growing privacy and governance concerns. In this paper, we propose the Intelligence Delivery Network (IDN), an Internet architecture that treats AI capabilities as deliverable network services. The key idea is to position, select, reuse, and verify intelligence across cloud, regional, edge, and local environments according to demand locality, resource availability, and policy constraints. We present the system assumptions of IDN, define its core architectural mechanisms, and discuss how capability abstraction, compute resource integration, demand-driven deployment, service routing, state-aware caching, and trust management can jointly support distributed AI services. We believe that IDN provides a practical path toward an Internet architecture for the AI age, making AI capabilities more accessible, efficient, trustworthy, and responsive to diverse application needs.

Zhiyao Xu, Dan Zhao, Qingsong Zou et al.

As smart homes become increasingly prevalent, intelligent models are widely used for tasks such as anomaly detection and behavior prediction. These models are typically trained on static datasets, making them brittle to behavioral drift caused by seasonal changes, lifestyle shifts, or evolving routines. However, collecting new behavior data for retraining is often impractical due to its slow pace, high cost, and privacy concerns. In this paper, we propose SmartGen, an LLM-based framework that synthesizes context-aware user behavior data to support continual adaptation of downstream smart home models. SmartGen consists of four key components. First, we design a Time and Semantic-aware Split module to divide long behavior sequences into manageable, semantically coherent subsequences under dual time-span constraints. Second, we propose Semantic-aware Sequence Compression to reduce input length while preserving representative semantics by clustering behavior mapping in latent space. Third, we introduce Graph-guided Sequence Synthesis, which constructs a behavior relationship graph and encodes frequent transitions into prompts, guiding the LLM to generate data aligned with contextual changes while retaining core behavior patterns. Finally, we design a Two-stage Outlier Filter to identify and remove implausible or semantically inconsistent outputs, aiming to improve the factual coherence and behavioral validity of the generated sequences. Experiments on three real-world datasets demonstrate that SmartGen significantly enhances model performance on anomaly detection and behavior prediction tasks under behavioral drift, with anomaly detection improving by 85.43% and behavior prediction by 70.51% on average. The code is available at https://github.com/horizonsinzqs/SmartGen.

Huihan Wang, Zhiwen Yang, Hui Zhang et al.

Synthesizing high-quality dynamic medical videos remains a significant challenge due to the need for modeling both spatial consistency and temporal dynamics. Existing Transformer-based approaches face critical limitations, including insufficient channel interactions, high computational complexity from self-attention, and coarse denoising guidance from timestep embeddings when handling varying noise levels. In this work, we propose FEAT, a full-dimensional efficient attention Transformer, which addresses these issues through three key innovations: (1) a unified paradigm with sequential spatial-temporal-channel attention mechanisms to capture global dependencies across all dimensions, (2) a linear-complexity design for attention mechanisms in each dimension, utilizing weighted key-value attention and global channel attention, and (3) a residual value guidance module that provides fine-grained pixel-level guidance to adapt to different noise levels. We evaluate FEAT on standard benchmarks and downstream tasks, demonstrating that FEAT-S, with only 23\% of the parameters of the state-of-the-art model Endora, achieves comparable or even superior performance. Furthermore, FEAT-L surpasses all comparison methods across multiple datasets, showcasing both superior effectiveness and scalability. Code is available at https://github.com/Yaziwel/FEAT.

Sihan Chen, Dan Zhao, Jongwoo Ko et al.

The growing computational demands of large language models (LLMs) make efficient inference and activation strategies increasingly critical. While recent approaches, such as Mixture-of-Experts (MoE), leverage selective activation but require specialized training, training-free sparse activation methods offer broader applicability and superior resource efficiency through their plug-and-play design. However, many existing methods rely solely on hidden state magnitudes to determine activation, resulting in high approximation errors and suboptimal inference accuracy. To address these limitations, we propose WINA (Weight Informed Neuron Activation), a novel, simple, and training-free sparse activation framework that jointly considers hidden state magnitudes and the column-wise $\ell_2$-norms of weight matrices. We show that this leads to a sparsification strategy that obtains optimal approximation error bounds with theoretical guarantees tighter than existing techniques. Empirically, WINA also outperforms state-of-the-art methods (e.g., TEAL) by up to $2.94\%$ in average performance at the same sparsity levels, across a diverse set of LLM architectures and datasets. These results position WINA as a new performance frontier for training-free sparse activation in LLM inference, advancing training-free sparse activation methods and setting a robust baseline for efficient inference. The source code is available at https://github.com/microsoft/wina.

Varshitha Chennamsetti, Laiba Mehnaz, Dan Zhao et al.

In this paper, we report the performance benchmarking results of deep learning models on MLCommons' Science cloud-masking benchmark using a high-performance computing cluster at New York University (NYU): NYU Greene. MLCommons is a consortium that develops and maintains several scientific benchmarks that can benefit from developments in AI. We provide a description of the cloud-masking benchmark task, updated code, and the best model for this benchmark when using our selected hyperparameter settings. Our benchmarking results include the highest accuracy achieved on the NYU system as well as the average time taken for both training and inference on the benchmark across several runs/seeds. Our code can be found on GitHub. MLCommons team has been kept informed about our progress and may use the developed code for their future work.

Oscar Skean, Md Rifat Arefin, Dan Zhao et al.

From extracting features to generating text, the outputs of large language models (LLMs) typically rely on the final layers, following the conventional wisdom that earlier layers capture only low-level cues. However, our analysis shows that intermediate layers can encode even richer representations, often improving performance on a range of downstream tasks. To explain and quantify these hidden-layer properties, we propose a unified framework of representation quality metrics based on information theory, geometry, and invariance to input perturbations. Our framework highlights how each layer balances information compression and signal preservation, revealing why mid-depth embeddings can exceed the last layer's performance. Through extensive experiments on 32 text-embedding tasks across various architectures (transformers, state-space models) and domains (language, vision), we demonstrate that intermediate layers consistently provide stronger features, challenging the standard view on final-layer embeddings and opening new directions on using mid-layer representations for more robust and accurate representations.

Dan Zhao, Siddharth Samsi, Joseph McDonald et al.

As research and deployment of AI grows, the computational burden to support and sustain its progress inevitably does too. To train or fine-tune state-of-the-art models in NLP, computer vision, etc., some form of AI hardware acceleration is virtually a requirement. Recent large language models require considerable resources to train and deploy, resulting in significant energy usage, potential carbon emissions, and massive demand for GPUs and other hardware accelerators. However, this surge carries large implications for energy sustainability at the HPC/datacenter level. In this paper, we study the aggregate effect of power-capping GPUs on GPU temperature and power draw at a research supercomputing center. With the right amount of power-capping, we show significant decreases in both temperature and power draw, reducing power consumption and potentially improving hardware life-span with minimal impact on job performance. While power-capping reduces power draw by design, the aggregate system-wide effect on overall energy consumption is less clear; for instance, if users notice job performance degradation from GPU power-caps, they may request additional GPU-jobs to compensate, negating any energy savings or even worsening energy consumption. To our knowledge, our work is the first to conduct and make available a detailed analysis of the effects of GPU power-capping at the supercomputing scale. We hope our work will inspire HPCs/datacenters to further explore, evaluate, and communicate the impact of power-capping AI hardware accelerators for more sustainable AI.

Lawrence Jang, Yinheng Li, Dan Zhao et al.

Videos are often used to learn or extract the necessary information to complete tasks in ways different than what text and static imagery alone can provide. However, many existing agent benchmarks neglect long-context video understanding, instead focusing on text or static image inputs. To bridge this gap, we introduce VideoWebArena (VideoWA), a benchmark for evaluating the capabilities of long-context multimodal agents for video understanding. VideoWA consists of 2,021 web agent tasks based on manually crafted video tutorials, which total almost four hours of content. For our benchmark, we define a taxonomy of long-context video-based agent tasks with two main areas of focus: skill retention and factual retention. While skill retention tasks evaluate whether an agent can use a given human demonstration to complete a task efficiently, the factual retention task evaluates whether an agent can retrieve instruction-relevant information from a video to complete a task. We find that the best model achieves 13.3% success on factual retention tasks and 45.8% on factual retention QA pairs, far below human performance at 73.9% and 79.3%, respectively. On skill retention tasks, long-context models perform worse with tutorials than without, exhibiting a 5% performance decrease in WebArena tasks and a 10.3% decrease in VisualWebArena tasks. Our work highlights the need to improve the agentic abilities of long-context multimodal models and provides a testbed for future development with long-context video agents.

Zhengxin Zhang, Dan Zhao, Xupeng Miao et al.

Finetuning large language models (LLMs) has been empirically effective on a variety of downstream tasks. Existing approaches to finetuning an LLM either focus on parameter-efficient finetuning, which only updates a small number of trainable parameters, or attempt to reduce the memory footprint during the training phase of the finetuning. Typically, the memory footprint during finetuning stems from three contributors: model weights, optimizer states, and intermediate activations. However, existing works still require considerable memory and none can simultaneously mitigate memory footprint for all three sources. In this paper, we present Quantized Side Tuing (QST), which enables memory-efficient and fast finetuning of LLMs by operating through a dual-stage process. First, QST quantizes an LLM's model weights into 4-bit to reduce the memory footprint of the LLM's original weights; QST also introduces a side network separated from the LLM, which utilizes the hidden states of the LLM to make task-specific predictions. Using a separate side network avoids performing backpropagation through the LLM, thus reducing the memory requirement of the intermediate activations. Furthermore, QST leverages several low-rank adaptors and gradient-free downsample modules to significantly reduce the trainable parameters, so as to save the memory footprint of the optimizer states. Experiments show that QST can reduce the total memory footprint by up to 2.3 $\times$ and speed up the finetuning process by up to 3 $\times$ while achieving competent performance compared with the state-of-the-art. When it comes to full finetuning, QST can reduce the total memory footprint up to 7 $\times$.

Maximilian Puelma Touzel, Sneheel Sarangi, Austin Welch et al.

The rise of AI-driven manipulation poses significant risks to societal trust and democratic processes. Yet, studying these effects in real-world settings at scale is ethically and logistically impractical, highlighting a need for simulation tools that can model these dynamics in controlled settings to enable experimentation with possible defenses. We present a simulation environment designed to address this. We elaborate upon the Concordia framework that simulates offline, `real life' activity by adding online interactions to the simulation through social media with the integration of a Mastodon server. We improve simulation efficiency and information flow, and add a set of measurement tools, particularly longitudinal surveys. We demonstrate the simulator with a tailored example in which we track agents' political positions and show how partisan manipulation of agents can affect election results.

Oliver Knitter, Dan Zhao, James Stokes et al.

Neural-network quantum states (NQS) has emerged as a powerful application of quantum-inspired deep learning for variational Monte Carlo methods, offering a competitive alternative to existing techniques for identifying ground states of quantum problems. A significant advancement toward improving the practical scalability of NQS has been the incorporation of autoregressive models, most recently transformers, as variational ansatze. Transformers learn sequence information with greater expressiveness than recurrent models, but at the cost of increased time complexity with respect to sequence length. We explore the use of the retentive network (RetNet), a recurrent alternative to transformers, as an ansatz for solving electronic ground state problems in $\textit{ab initio}$ quantum chemistry. Unlike transformers, RetNets overcome this time complexity bottleneck by processing data in parallel during training, and recurrently during inference. We give a simple computational cost estimate of the RetNet and directly compare it with similar estimates for transformers, establishing a clear threshold ratio of problem-to-model size past which the RetNet's time complexity outperforms that of the transformer. Though this efficiency can comes at the expense of decreased expressiveness relative to the transformer, we overcome this gap through training strategies that leverage the autoregressive structure of the model -- namely, variational neural annealing. Our findings support the RetNet as a means of improving the time complexity of NQS without sacrificing accuracy. We provide further evidence that the ablative improvements of neural annealing extend beyond the RetNet architecture, suggesting it would serve as an effective general training strategy for autoregressive NQS.

Shaona Ghosh, Barnaby Simkin, Kyriacos Shiarlis et al.

This paper introduces a dynamic and actionable framework for securing agentic AI systems in enterprise deployment. We contend that safety and security are not merely fixed attributes of individual models but also emergent properties arising from the dynamic interactions among models, orchestrators, tools, and data within their operating environments. We propose a new way of identification of novel agentic risks through the lens of user safety. Although, for traditional LLMs and agentic models in isolation, safety and security has a clear separation, through the lens of safety in agentic systems, they appear to be connected. Building on this foundation, we define an operational agentic risk taxonomy that unifies traditional safety and security concerns with novel, uniquely agentic risks, including tool misuse, cascading action chains, and unintended control amplification among others. At the core of our approach is a dynamic agentic safety and security framework that operationalizes contextual agentic risk management by using auxiliary AI models and agents, with human oversight, to assist in contextual risk discovery, evaluation, and mitigation. We further address one of the most challenging aspects of safety and security of agentic systems: risk discovery through sandboxed, AI-driven red teaming. We demonstrate the framework effectiveness through a detailed case study of NVIDIA flagship agentic research assistant, AI-Q Research Assistant, showcasing practical, end-to-end safety and security evaluations in complex, enterprise-grade agentic workflows. This risk discovery phase finds novel agentic risks that are then contextually mitigated. We also release the dataset from our case study, containing traces of over 10,000 realistic attack and defense executions of the agentic workflow to help advance research in agentic safety.

Aurélien Bück-Kaeffer, Je Qin Chooi, Dan Zhao et al.

Large language models (LLMs) offer promising capabilities for simulating social media dynamics at scale, enabling studies that would be ethically or logistically challenging with human subjects. However, the field lacks standardized data resources for fine-tuning and evaluating LLMs as realistic social media agents. We address this gap by introducing SIMPACT, the SIMulation-oriented Persona and Action Capture Toolkit, a privacy respecting framework for constructing behaviorally-grounded social media datasets suitable for training agent models. We formulate next-action prediction as a task for training and evaluating LLM-based agents and introduce metrics at both the cluster and population levels to assess behavioral fidelity and stylistic realism. As a concrete implementation, we release BluePrint, a large-scale dataset built from public Bluesky data focused on political discourse. BluePrint clusters anonymized users into personas of aggregated behaviours, capturing authentic engagement patterns while safeguarding privacy through pseudonymization and removal of personally identifiable information. The dataset includes a sizable action set of 12 social media interaction types (likes, replies, reposts, etc.), each instance tied to the posting activity preceding it. This supports the development of agents that use context-dependence, not only in the language, but also in the interaction behaviours of social media to model social media users. By standardizing data and evaluation protocols, SIMPACT provides a foundation for advancing rigorous, ethically responsible social media simulations. BluePrint serves as both an evaluation benchmark for political discourse modeling and a template for building domain specific datasets to study challenges such as misinformation and polarization.

Oliver Knitter, Dan Zhao, Stefan Leichenauer et al.

Scaling laws have been used to describe how large language model (LLM) performance scales with model size, training data size, or amount of computational resources. Motivated by the fact that neural quantum states (NQS) has increasingly adopted LLM-based components, we seek to understand NQS scaling laws, thereby shedding light on the scalability and optimal performance--resource trade-offs of NQS ansatze. In particular, we identify scaling laws that predict the performance, as measured by absolute error and V-score, for transformer-based NQS as a function of problem size in second-quantized quantum chemistry applications. By performing analogous compute-constrained optimization of the obtained parametric curves, we find that the relationship between model size and training time is highly dependent on loss metric and ansatz, and does not follow the approximately linear relationship found for language models.

Wei Du, Branislav Kisacanin, George Armstrong et al.

Reasoning-capable language models achieve state-of-the-art performance in diverse complex tasks by generating long, explicit Chain-of-Thought (CoT) traces. While recent works show that base models can acquire such reasoning traces via reinforcement learning or distillation from stronger models like DeepSeek-R1, previous works demonstrate that even short CoT prompting without fine-tuning is able to improve reasoning. We ask whether long CoT can be induced in a base model using only prompting or minimal tuning. Using just 20 long CoT examples from the reasoning model \texttt{QwQ-32B-Preview}, we lightly fine-tune the base model \texttt{Qwen2.5-32B}. The resulting model outperforms the much larger \texttt{Qwen2.5-Math-72B-Instruct}, showing that a handful of high-quality examples can unlock strong reasoning capabilities. We further explore using CoT data from non-reasoning models and human annotators, enhanced with prompt engineering, multi-pass editing, and structural guidance. However, neither matches the performance of reasoning model traces, suggesting that certain latent qualities of expert CoT are difficult to replicate. We analyze key properties of reasoning data, such as problem difficulty, diversity, and answer length, that influence reasoning distillation. While challenges remain, we are optimistic that carefully curated human-written CoT, even in small quantities, can activate reasoning behaviors in base models. We release our human-authored dataset across refinement stages and invite further investigation into what makes small-scale reasoning supervision so effective.

Fucheng Guo, Zeyu Luan, Qing Li et al.

Federated Learning (FL) has emerged as an essential framework for distributed machine learning, especially with its potential for privacy-preserving data processing. However, existing FL frameworks struggle to address statistical and model heterogeneity, which severely impacts model performance. While Heterogeneous Federated Learning (HtFL) introduces prototype-based strategies to address the challenges, current approaches face limitations in achieving optimal separation of prototypes. This paper presents FedORGP, a novel HtFL algorithm designed to improve global prototype separation through orthogonality regularization, which not only encourages intra-class prototype similarity but also significantly expands the inter-class angular separation. With the guidance of the global prototype, each client keeps its embeddings aligned with the corresponding prototype in the feature space, promoting directional independence that integrates seamlessly with the cross-entropy (CE) loss. We provide theoretical proof of FedORGP's convergence under non-convex conditions. Extensive experiments demonstrate that FedORGP outperforms seven state-of-the-art baselines, achieving up to 10.12\% accuracy improvement in scenarios where statistical and model heterogeneity coexist.

Zhiyao Xu, Dan Zhao, Qingsong Zou et al.

In recent years, as smart home systems have become more widespread, security concerns within these environments have become a growing threat. Currently, most smart home security solutions, such as anomaly detection and behavior prediction models, are trained using fixed datasets that are precollected. However, the process of dataset collection is time-consuming and lacks the flexibility needed to adapt to the constantly evolving smart home environment. Additionally, the collection of personal data raises significant privacy concerns for users. Lately, large language models (LLMs) have emerged as a powerful tool for a wide range of tasks across diverse application domains, thanks to their strong capabilities in natural language processing, reasoning, and problem-solving. In this paper, we propose an LLM-based synthetic dataset generation IoTGen framework to enhance the generalization of downstream smart home intelligent models. By generating new synthetic datasets that reflect changes in the environment, smart home intelligent models can be retrained to overcome the limitations of fixed and outdated data, allowing them to better align with the dynamic nature of real-world home environments. Specifically, we first propose a Structure Pattern Perception Compression (SPPC) method tailored for IoT behavior data, which preserves the most informative content in the data while significantly reducing token consumption. Then, we propose a systematic approach to create prompts and implement data generation to automatically generate IoT synthetic data with normative and reasonable properties, assisting task models in adaptive training to improve generalization and real-world performance.

Jingyu Xiao, Zhiyao Xu, Qingsong Zou et al.

Smart homes, powered by the Internet of Things, offer great convenience but also pose security concerns due to abnormal behaviors, such as improper operations of users and potential attacks from malicious attackers. Several behavior modeling methods have been proposed to identify abnormal behaviors and mitigate potential risks. However, their performance often falls short because they do not effectively learn less frequent behaviors, consider temporal context, or account for the impact of noise in human behaviors. In this paper, we propose SmartGuard, an autoencoder-based unsupervised user behavior anomaly detection framework. First, we design a Loss-guided Dynamic Mask Strategy (LDMS) to encourage the model to learn less frequent behaviors, which are often overlooked during learning. Second, we propose a Three-level Time-aware Position Embedding (TTPE) to incorporate temporal information into positional embedding to detect temporal context anomaly. Third, we propose a Noise-aware Weighted Reconstruction Loss (NWRL) that assigns different weights for routine behaviors and noise behaviors to mitigate the interference of noise behaviors during inference. Comprehensive experiments on three datasets with ten types of anomaly behaviors demonstrates that SmartGuard consistently outperforms state-of-the-art baselines and also offers highly interpretable results.

Constantijn van der Poel, Dan Zhao

This paper examines the use of tensor networks, which can efficiently represent high-dimensional quantum states, in language modeling. It is a distillation and continuation of the work done in (van der Poel, 2023). To do so, we will abstract the problem down to modeling Motzkin spin chains, which exhibit long-range correlations reminiscent of those found in language. The Matrix Product State (MPS), also known as the tensor train, has a bond dimension which scales as the length of the sequence it models. To combat this, we use the factored core MPS, whose bond dimension scales sub-linearly. We find that the tensor models reach near perfect classifying ability, and maintain a stable level of performance as the number of valid training examples is decreased.

Nathan C. Frey, Baolin Li, Joseph McDonald et al.

Deep learning (DL) workflows demand an ever-increasing budget of compute and energy in order to achieve outsized gains. Neural architecture searches, hyperparameter sweeps, and rapid prototyping consume immense resources that can prevent resource-constrained researchers from experimenting with large models and carry considerable environmental impact. As such, it becomes essential to understand how different deep neural networks (DNNs) and training leverage increasing compute and energy resources -- especially specialized computationally-intensive models across different domains and applications. In this paper, we conduct over 3,400 experiments training an array of deep networks representing various domains/tasks -- natural language processing, computer vision, and chemistry -- on up to 424 graphics processing units (GPUs). During training, our experiments systematically vary compute resource characteristics and energy-saving mechanisms such as power utilization and GPU clock rate limits to capture and illustrate the different trade-offs and scaling behaviors each representative model exhibits under various resource and energy-constrained regimes. We fit power law models that describe how training time scales with available compute resources and energy constraints. We anticipate that these findings will help inform and guide high-performance computing providers in optimizing resource utilization, by selectively reducing energy consumption for different deep learning tasks/workflows with minimal impact on training.

Dan Zhao, Guizhi Xu, Zhenghua XU et al.

Computer-Aided Diagnosis and Treatment of Tumors is a hot topic of deep learning in recent years, which constitutes a series of medical tasks, such as detection of tumor markers, the outline of tumor leisures, subtypes and stages of tumors, prediction of therapeutic effect, and drug development. Meanwhile, there are some deep learning models with precise positioning and excellent performance produced in mainstream task scenarios. Thus follow to introduce deep learning methods from task-orient, mainly focus on the improvements for medical tasks. Then to summarize the recent progress in four stages of tumor diagnosis and treatment, which named In-Vitro Diagnosis (IVD), Imaging Diagnosis (ID), Pathological Diagnosis (PD), and Treatment Planning (TP). According to the specific data types and medical tasks of each stage, we present the applications of deep learning in the Computer-Aided Diagnosis and Treatment of Tumors and analyzing the excellent works therein. This survey concludes by discussing research issues and suggesting challenges for future improvement.