Jingwei Zuo

Weize Chen, Yusheng Su, Jingwei Zuo et al. · tsinghua

Autonomous agents empowered by Large Language Models (LLMs) have undergone significant improvements, enabling them to generalize across a broad spectrum of tasks. However, in real-world scenarios, cooperation among individuals is often required to enhance the efficiency and effectiveness of task accomplishment. Hence, inspired by human group dynamics, we propose a multi-agent framework \framework that can collaboratively and dynamically adjust its composition as a greater-than-the-sum-of-its-parts system. Our experiments demonstrate that \framework framework can effectively deploy multi-agent groups that outperform a single agent. Furthermore, we delve into the emergence of social behaviors among individual agents within a group during collaborative task accomplishment. In view of these behaviors, we discuss some possible strategies to leverage positive ones and mitigate negative ones for improving the collaborative potential of multi-agent groups. Our codes for \framework will soon be released at \url{https://github.com/OpenBMB/AgentVerse}.

Jingwei Zuo, Wenbin Li, Michele Baldo et al.

Air quality forecasting has garnered significant attention recently, with data-driven models taking center stage due to advancements in machine learning and deep learning models. However, researchers face challenges with complex data acquisition and the lack of open-sourced datasets, hindering efficient model validation. This paper introduces PurpleAirSF, a comprehensive and easily accessible dataset collected from the PurpleAir network. With its high temporal resolution, various air quality measures, and diverse geographical coverage, this dataset serves as a useful tool for researchers aiming to develop novel forecasting models, study air pollution patterns, and investigate their impacts on health and the environment. We present a detailed account of the data collection and processing methods employed to build PurpleAirSF. Furthermore, we conduct preliminary experiments using both classic and modern spatio-temporal forecasting models, thereby establishing a benchmark for future air quality forecasting tasks.

Jingwei Zuo, Karine Zeitouni, Yehia Taher et al.

Traffic forecasting has attracted widespread attention recently. In reality, traffic data usually contains missing values due to sensor or communication errors. The Spatio-temporal feature in traffic data brings more challenges for processing such missing values, for which the classic techniques (e.g., data imputations) are limited: 1) in temporal axis, the values can be randomly or consecutively missing; 2) in spatial axis, the missing values can happen on one single sensor or on multiple sensors simultaneously. Recent models powered by Graph Neural Networks achieved satisfying performance on traffic forecasting tasks. However, few of them are applicable to such a complex missing-value context. To this end, we propose GCN-M, a Graph Convolutional Network model with the ability to handle the complex missing values in the Spatio-temporal context. Particularly, we jointly model the missing value processing and traffic forecasting tasks, considering both local Spatio-temporal features and global historical patterns in an attention-based memory network. We propose as well a dynamic graph learning module based on the learned local-global features. The experimental results on real-life datasets show the reliability of our proposed method.

Guangxuan Xiao, Jiaming Tang, Jingwei Zuo et al. · mit

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.

Jingwei Zuo, George Arvanitakis, Hakim Hacid

Human activity recognition (HAR) has been a classic research problem. In particular, with recent machine learning (ML) techniques, the recognition task has been largely investigated by companies and integrated into their products for customers. However, most of them apply a predefined activity set and conduct the learning process on the cloud, hindering specific personalizations from end users (i.e., edge devices). Even though recent progress in Incremental Learning allows learning new-class data on the fly, the learning process is generally conducted on the cloud, requiring constant data exchange between cloud and edge devices, thus leading to data privacy issues. In this paper, we propose PILOTE, which pushes the incremental learning process to the extreme edge, while providing reliable data privacy and practical utility, e.g., low processing latency, personalization, etc. In particular, we consider the practical challenge of extremely limited data during the incremental learning process on edge, where catastrophic forgetting is required to be handled in a practical way. We validate PILOTE with extensive experiments on human activity data collected from mobile sensors. The results show PILOTE can work on edge devices with extremely limited resources while providing reliable performance.

Jingwei Zuo, Wenbin Li, Michele Baldo et al.

Air Quality Monitoring and Forecasting has been a popular research topic in recent years. Recently, data-driven approaches for air quality forecasting have garnered significant attention, owing to the availability of well-established data collection facilities in urban areas. Fixed infrastructures, typically deployed by national institutes or tech giants, often fall short in meeting the requirements of diverse personalized scenarios, e.g., forecasting in areas without any existing infrastructure. Consequently, smaller institutes or companies with limited budgets are compelled to seek tailored solutions by introducing more flexible infrastructures for data collection. In this paper, we propose an expandable graph attention network (EGAT) model, which digests data collected from existing and newly-added infrastructures, with different spatial structures. Additionally, our proposal can be embedded into any air quality forecasting models, to apply to the scenarios with evolving spatial structures. The proposal is validated over real air quality data from PurpleAir.

Jingwei Zuo, Maksim Velikanov, Ilyas Chahed et al.

In this report, we introduce Falcon-H1, a new series of large language models (LLMs) featuring hybrid architecture designs optimized for both high performance and efficiency across diverse use cases. Unlike earlier Falcon models built solely on Transformer or Mamba architectures, Falcon-H1 adopts a parallel hybrid approach that combines Transformer-based attention with State Space Models (SSMs), known for superior long-context memory and computational efficiency. We systematically revisited model design, data strategy, and training dynamics, challenging conventional practices in the field. Falcon-H1 is released in multiple configurations, including base and instruction-tuned variants at 0.5B, 1.5B, 1.5B-deep, 3B, 7B, and 34B parameters. Quantized instruction-tuned models are also available, totaling over 30 checkpoints on Hugging Face Hub. Falcon-H1 models demonstrate state-of-the-art performance and exceptional parameter and training efficiency. The flagship Falcon-H1-34B matches or outperforms models up to 70B scale, such as Qwen3-32B, Qwen2.5-72B, and Llama3.3-70B, while using fewer parameters and less data. Smaller models show similar trends: the Falcon-H1-1.5B-Deep rivals current leading 7B-10B models, and Falcon-H1-0.5B performs comparably to typical 7B models from 2024. These models excel across reasoning, mathematics, multilingual tasks, instruction following, and scientific knowledge. With support for up to 256K context tokens and 18 languages, Falcon-H1 is suitable for a wide range of applications. All models are released under a permissive open-source license, underscoring our commitment to accessible and impactful AI research.

Maksim Velikanov, Ilyas Chahed, Jingwei Zuo et al.

Applying weight decay (WD) to matrix layers is standard practice in large-language-model pretraining. Prior work suggests that stochastic gradient noise induces a Brownian-like expansion of the weight matrices W, whose growth is counteracted by WD, leading to a WD-noise equilibrium with a certain weight norm ||W||. In this work, we view the equilibrium norm as a harmful artifact of the training procedure, and address it by introducing learnable multipliers to learn the optimal scale. First, we attach a learnable scalar multiplier to W and confirm that the WD-noise equilibrium norm is suboptimal: the learned scale adapts to data and improves performance. We then argue that individual row and column norms are similarly constrained, and free their scale by introducing learnable per-row and per-column multipliers. Our method can be viewed as a learnable, more expressive generalization of muP multipliers. It outperforms a well-tuned muP baseline, reduces the computational overhead of multiplier tuning, and surfaces practical questions such as forward-pass symmetries and the width-scaling of the learned multipliers. Finally, we validate learnable multipliers with both Adam and Muon optimizers, where it shows improvement in downstream evaluations matching the improvement of the switching from Adam to Muon.

George Arvanitakis, Jingwei Zuo, Mthandazo Ndhlovu et al.

Edge Machine Learning (Edge ML), which shifts computational intelligence from cloud-based systems to edge devices, is attracting significant interest due to its evident benefits including reduced latency, enhanced data privacy, and decreased connectivity reliance. While these advantages are compelling, they introduce unique challenges absent in traditional cloud-based approaches. In this paper, we delve into the intricacies of Edge-based learning, examining the interdependencies among: (i) constrained data storage on Edge devices, (ii) limited computational power for training, and (iii) the number of learning classes. Through experiments conducted using our MAGNETO system, that focused on learning human activities via data collected from mobile sensors, we highlight these challenges and offer valuable perspectives on Edge ML.

Jingwei Zuo, Xinze Feng, Zien Liu et al.

Low-Rank Adaptation (LoRA) is now the dominant method for parameter-efficient fine-tuning of large language models, but achieving a high-quality adapter often requires systematic hyperparameter tuning because LoRA performance is highly sensitive to configuration choices. In practice, this leads to many concurrent LoRA jobs, often spanning heterogeneous tasks in multi-tenant environments. Existing systems largely handle these jobs independently, which both wastes computation on weak candidates and leaves GPUs underutilized. We present ALTO (Adaptive LoRA Tuning and Orchestration), a co-designed training system that accelerates LoRA hyperparameter tuning while enabling efficient cluster sharing across heterogeneous tasks. The central insight behind ALTO is that when multiple tuning jobs run concurrently over a shared frozen backbone, they expose optimization opportunities that single-job designs cannot exploit. Building on this, ALTO monitors loss trajectories to terminate unpromising configurations early, uses fused grouped GEMM together with a new rank-local adapter parallelism to co-locate surviving adapters and reclaim freed GPU capacity, and combines intra-task and inter-task scheduling to improve multi-task placement by leveraging the predictable duration of LoRA jobs. Extensive evaluation shows that ALTO achieves up to $13.8\times$ speedup over state-of-the-art without sacrificing adapter quality.

Jingwei Zuo, George Arvanitakis, Mthandazo Ndhlovu et al.

Human activity recognition (HAR) is a well-established field, significantly advanced by modern machine learning (ML) techniques. While companies have successfully integrated HAR into consumer products, they typically rely on a predefined activity set, which limits personalizations at the user level (edge devices). Despite advancements in Incremental Learning for updating models with new data, this often occurs on the Cloud, necessitating regular data transfers between cloud and edge devices, thus leading to data privacy issues. In this paper, we propose MAGNETO, an Edge AI platform that pushes HAR tasks from the Cloud to the Edge. MAGNETO allows incremental human activity learning directly on the Edge devices, without any data exchange with the Cloud. This enables strong privacy guarantees, low processing latency, and a high degree of personalization for users. In particular, we demonstrate MAGNETO in an Android device, validating the whole pipeline from data collection to result visualization.

Fanjiang Ye, Zhangke Li, Xinrui Zhong et al.

Diffusion models have emerged as the prevailing approach for text-to-image (T2I) and text-to-video (T2V) generation, yet production platforms must increasingly serve both modalities on shared GPU clusters while meeting stringent latency SLOs. Co-serving such heterogeneous workloads is challenging: T2I and T2V requests exhibit vastly different compute demands, parallelism characteristics, and latency requirements, leading to significant SLO violations in existing serving systems. We present GENSERVE, a co-serving system that leverages the inherent predictability of the diffusion process to optimize serving efficiency. A central insight is that diffusion inference proceeds in discrete, predictable steps and is naturally preemptible at step boundaries, opening a new design space for heterogeneity-aware resource management. GENSERVE introduces step-level resource adaptation through three coordinated mechanisms: intelligent video preemption, elastic sequence parallelism with dynamic batching, and an SLO-aware scheduler that jointly optimizes resource allocation across all concurrent requests. Experimental results show that GENSERVE improves the SLO attainment rate by up to 44% over the strongest baseline across diverse configurations.

Mouadh Yagoubi, Yasser Dahou, Billel Mokeddem et al.

Existing benchmarks have proven effective for assessing the performance of fully trained large language models. However, we find striking differences in the early training stages of small models, where benchmarks often fail to provide meaningful or discriminative signals. To explore how these differences arise, this competition tackles the challenge of designing scientific knowledge evaluation tasks specifically tailored for measuring early training progress of language models. Participants are invited to develop novel evaluation methodologies or adapt existing benchmarks to better capture performance differences among language models. To support this effort, we provide three pre-trained small models (0.5B, 1B, and 3B parameters), along with intermediate checkpoints sampled during training up to 200B tokens. All experiments and development work can be run on widely available free cloud-based GPU platforms, making participation accessible to researchers with limited computational resources. Submissions will be evaluated based on three criteria: the quality of the performance signal they produce, the consistency of model rankings at 1 trillion tokens of training, and their relevance to the scientific knowledge domain. By promoting the design of tailored evaluation strategies for early training, this competition aims to attract a broad range of participants from various disciplines, including those who may not be machine learning experts or have access to dedicated GPU resources. Ultimately, this initiative seeks to make foundational LLM research more systematic and benchmark-informed from the earliest phases of model development.

Jingwei Zuo, Hakim Hacid

Human Activity Recognition (HAR) has been studied for decades, from data collection, learning models, to post-processing and result interpretations. However, the inherent hierarchy in the activities remains relatively under-explored, despite its significant impact on model performance and interpretation. In this paper, we propose H-HAR, by rethinking the HAR tasks from a fresh perspective by delving into their intricate global label relationships. Rather than building multiple classifiers separately for multi-layered activities, we explore the efficacy of a flat model enhanced with graph-based label relationship modeling. Being hierarchy-aware, the graph-based label modeling enhances the fundamental HAR model, by incorporating intricate label relationships into the model. We validate the proposal with a multi-label classifier on complex human activity data. The results highlight the advantages of the proposal, which can be vertically integrated into advanced HAR models to further enhance their performances.

Jingwei Zuo, Karine Zeitouni, Yehia Taher

Learning from Multivariate Time Series (MTS) has attracted widespread attention in recent years. In particular, label shortage is a real challenge for the classification task on MTS, considering its complex dimensional and sequential data structure. Unlike self-training and positive unlabeled learning that rely on distance-based classifiers, in this paper, we propose SMATE, a novel semi-supervised model for learning the interpretable Spatio-Temporal representation from weakly labeled MTS. We validate empirically the learned representation on 30 public datasets from the UEA MTS archive. We compare it with 13 state-of-the-art baseline methods for fully supervised tasks and four baselines for semi-supervised tasks. The results show the reliability and efficiency of our proposed method.