arXiv subjects

Dongqi Cai

Publications and source records attributed to Dongqi Cai.

At least 19 recordsLinked to original sources

arXiv2026

PhyAI: Real-Time Physical AI at the Edge, Scalable Rollouts in the Cloud

Physical AI policies require inference throughout their lifecycle, including model evaluation, cloud reinforcement learning rollout, edge GPU serving, and onboard deployment. Although these settings share the same checkpoint and action semantics, they often rely on separate inference programs. To unify them, we build PhyAI, a Physical AI inference engine with a single runtime that keeps architecture-specific conditioning, solver, cache, and output logic in model adapters while sharing graph execution, kernels, memory management, and parallel services. The same codebase runs vision-language-action (VLA) models and world-action models (WAMs) on single or multiple GPUs across onboard, edge, and cloud deployments. We used the adapter interface to add MiniCPM-Robot on the day of its release. PhyAI achieves 1.40x-4.65x speedups over the official implementations of pi0, pi0.5, GR00T N1.7, and MiniCPM-Robot. On Cosmos3-Nano-Policy-DROID it reduces latency from 2.46 to 1.18 s on eight H20 GPUs (CFG=2, TP=4), a 2.08x speedup. Specialized runtimes remain faster in several configurations, so our goal is one runtime with competitive latency rather than the fastest result in every case. Detailed profiles reveal why different models need different execution policies: on a Hopper-series GPU at batch size one, the pi0.5 action expert accounts for 8.8% of FLOPs but 57.2% of latency; at batch size 32 its share drops to 13.5% and throughput reaches about 100 samples/s. Cosmos3 remains generation-dominated and gains only 14.3% throughput as batch size increases from 1 to 16. We further introduce the control-time Roofline, which distinguishes inference-bound from environment-bound control; the measured pi0.5 points on four LIBERO suites are environment-bound while Cosmos3 stays inference-bound. Code and benchmarks: https://github.com/mingti-org/phyai.

cs.AI↗

arXiv2026

Beyond Scaling: Agents Are Heading to the Edge

The bottleneck of useful agentic intelligence has shifted from compressing world knowledge into a single model to executing a coordinated system. This position paper argues that personal-agent architecture must move to the edge because the core properties of agentic intelligence tasks, particularly their structural coupling with high-fidelity local context and the need for zero-latency execution loops, do not sit well with cloud-centric designs. We develop this claim through three structural shifts. First, the Prefrontal Turn: the main marginal lever of capability has moved from pre-training scale to framework-level executive control. Such control must remain physically close to the environment of action if the agent is to preserve cognitive alignment. Second, the Data-Geography Paradox, the ``dark matter'' of agentic data (local file hierarchies, real-time sensor streams, and transient OS states) degrades, disappears, or loses meaning once prepared for cloud transmission, thereby cutting the agent off from ground-truth context. Third, the interaction-alignment loop, the only economically and ecologically sustainable source of agentic refinement data is the high-fidelity implicit preference signal produced through real-time local interaction. Third, the interaction-alignment loop, the only economically and ecologically sustainable source of agentic refinement data is the high-fidelity implicit preference signal produced through real-time local interaction. We conclude with falsifiable predictions for the next deployment cycle of personal agents.

cs.LG↗

arXiv2025

MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs

Large language models (LLMs) are deployed on mobile devices to power killer applications such as intelligent assistants. LLMs pre-trained on general corpora often hallucinate when handling personalized or unseen queries, leading to incorrect or outdated responses. Knowledge editing addresses this by identifying and adjusting a small crucial portion of model weights, without compromising the general knowledge. However, prior knowledge editing methods are impractical to run on local devices due to the resource-heavy backpropagation (BP) needed for updates. We present MobiEdit, the first mobile knowledge editing framework that enables efficient LLM personalization on commercial off-the-shelf (COTS) mobile devices. MobiEdit replaces full-precision BP with quantized forward-only gradient estimation, thus compatible with the energy-efficient mobile neural processing units (NPUs). MobiEdit replaces full-precision backpropagation with quantized forward-only gradient estimation, making it compatible with energy-efficient mobile NPUs. To further improve gradient estimation efficiency, we introduce two optimizations: an early stoping mechanism that adaptively terminates editing upon success and a prefix cache that reuses computation across steps. Our approach enables real-time editing of a 3B-parameter model (Qwen2.5-3B-Instruct) on COTS mobile devices with 7.6$\times$ less memory, 14.7 $\times$ less energy and 3.6$\times$ less latency compared to previous knowledge editing methods.

cs.LG↗

arXiv2025

Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?

Large language Model (LLM) unlearning, i.e., selectively removing information from LLMs, is vital for responsible model deployment. Differently, LLM knowledge editing aims to modify LLM knowledge instead of removing it. Though editing and unlearning seem to be two distinct tasks, we find there is a tight connection between them. In this paper, we conceptualize unlearning as a special case of editing where information is modified to a refusal or "empty set" $\emptyset$ response, signifying its removal. This paper thus investigates if knowledge editing techniques are strong baselines for LLM unlearning. We evaluate state-of-the-art (SOTA) editing methods (e.g., ROME, MEMIT, GRACE, WISE, and AlphaEdit) against existing unlearning approaches on pretrained and finetuned knowledge. Results show certain editing methods, notably WISE and AlphaEdit, are effective unlearning baselines, especially for pretrained knowledge, and excel in generating human-aligned refusal answers. To better adapt editing methods for unlearning applications, we propose practical recipes including self-improvement and query merging. The former leverages the LLM's own in-context learning ability to craft a more human-aligned unlearning target, and the latter enables ROME and MEMIT to perform well in unlearning longer sample sequences. We advocate for the unlearning community to adopt SOTA editing methods as baselines and explore unlearning from an editing perspective for more holistic LLM memory control.

cs.LG↗

arXiv2025

DEPT: Decoupled Embeddings for Pre-training Language Models

Language Model pre-training uses broad data mixtures to enhance performance across domains and languages. However, training on such heterogeneous text corpora requires extensive and expensive efforts. Since these data sources vary significantly in lexical, syntactic, and semantic aspects, they cause negative interference or the ``curse of multilinguality''. To address these challenges we propose a communication-efficient pre-training framework, DEPT. Our method decouples embeddings from the transformer body while simultaneously training the latter on multiple data sources without requiring a shared vocabulary. DEPT can: (1) train robustly and effectively under significant data heterogeneity, (2) minimize token embedding parameters to only what the data source vocabulary requires, while cutting communication costs in direct proportion to both the communication frequency and the reduction in parameters, (3) enhance transformer body plasticity and generalization, improving both average perplexity (up to 20%) and downstream task performance, and (4) enable training with custom optimized vocabularies per data source. We demonstrate DEPT's potential via the first vocabulary-agnostic federated pre-training of billion-scale models, reducing communication costs by orders of magnitude and embedding memory by 4-5x.

cs.LG↗

arXiv2025

Small Language Models: Survey, Measurements, and Insights

Small language models (SLMs), despite their widespread adoption in modern smart devices, have received significantly less academic attention compared to their large language model (LLM) counterparts, which are predominantly deployed in data centers and cloud environments. While researchers continue to improve the capabilities of LLMs in the pursuit of artificial general intelligence, SLM research aims to make machine intelligence more accessible, affordable, and efficient for everyday tasks. Focusing on transformer-based, decoder-only language models with 100M-5B parameters, we survey 70 state-of-the-art open-source SLMs, analyzing their technical innovations across three axes: architectures, training datasets, and training algorithms. In addition, we evaluate their capabilities in various domains, including commonsense reasoning, mathematics, in-context learning, and long context. To gain further insight into their on-device runtime costs, we benchmark their inference latency and memory footprints. Through in-depth analysis of our benchmarking data, we offer valuable insights to advance research in this field.

cs.CL↗

arXiv2025

ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents

Recent advancements in integrating large language models (LLMs) with application programming interfaces (APIs) have gained significant interest in both academia and industry. Recent work demonstrates that these API-based agents exhibit relatively strong autonomy and planning capabilities. However, their ability to handle multi-dimensional difficulty levels, diverse task types, and real-world demands remains unknown. In this paper, we introduce \textsc{ShortcutsBench}, a large-scale benchmark for the comprehensive evaluation of API-based agents in solving real-world complex tasks. \textsc{ShortcutsBench} includes a wealth of real APIs from Apple Inc., refined user queries, human-annotated high-quality action sequences, detailed parameter filling values, and parameters requesting necessary input from the system or user. We revealed how existing benchmarks~/~datasets struggle to accommodate the advanced reasoning capabilities of existing more intelligent LLMs. Moreover, our extensive evaluation of agents built with $5$ leading open-source (size $\geq$ 57B) and $5$ closed-source LLMs (e.g. Gemini-1.5-Pro and GPT-4o-mini) with varying intelligence level reveals significant limitations of existing API-based agents in the whole process of handling complex queries related to API selection, parameter filling, and requesting necessary input from the system and the user. These findings highlight the great challenges that API-based agents face in effectively fulfilling real and complex user queries. All datasets, code, experimental logs, and results are available at \url{https://github.com/EachSheep/ShortcutsBench}.

cs.SE↗

arXiv2024

Photon: Federated LLM Pre-Training

Scaling large language models (LLMs) demands extensive data and computing resources, which are traditionally constrained to data centers by the high-bandwidth requirements of distributed training. Low-bandwidth methods like federated learning (FL) could enable collaborative training of larger models across weakly-connected GPUs if they can effectively be used for pre-training. To achieve this, we introduce Photon, the first complete system for federated end-to-end LLM training, leveraging cross-silo FL for global-scale training with minimal communication overheads. Using Photon, we train the first federated family of decoder-only LLMs from scratch. We show that: (1) Photon can train model sizes up to 7B in a federated fashion while reaching an even better perplexity than centralized pre-training; (2) Photon model training time decreases with available compute, achieving a similar compute-time trade-off to centralized; and (3) Photon outperforms the wall-time of baseline distributed training methods by 35% via communicating 64x-512xless. Our proposal is robust to data heterogeneity and converges twice as fast as previous methods like DiLoCo. This surprising data efficiency stems from a unique approach combining small client batch sizes with extremely high learning rates, enabled by federated averaging's robustness to hyperparameters. Photon thus represents the first economical system for global internet-wide LLM pre-training.

cs.LG↗

arXiv2024

A Survey of Resource-efficient LLM and Multimodal Foundation Models

Large foundation models, including large language models (LLMs), vision transformers (ViTs), diffusion, and LLM-based multimodal models, are revolutionizing the entire machine learning lifecycle, from training to deployment. However, the substantial advancements in versatility and performance these models offer come at a significant cost in terms of hardware resources. To support the growth of these large models in a scalable and environmentally sustainable way, there has been a considerable focus on developing resource-efficient strategies. This survey delves into the critical importance of such research, examining both algorithmic and systemic aspects. It offers a comprehensive analysis and valuable insights gleaned from existing literature, encompassing a broad array of topics from cutting-edge model architectures and training/serving algorithms to practical system designs and implementations. The goal of this survey is to provide an overarching understanding of how current approaches are tackling the resource challenges posed by large foundation models and to potentially inspire future breakthroughs in this field.

cs.LG↗

arXiv2024

Recall: Empowering Multimodal Embedding for Edge Devices

Human memory is inherently prone to forgetting. To address this, multimodal embedding models have been introduced, which transform diverse real-world data into a unified embedding space. These embeddings can be retrieved efficiently, aiding mobile users in recalling past information. However, as model complexity grows, so do its resource demands, leading to reduced throughput and heavy computational requirements that limit mobile device implementation. In this paper, we introduce RECALL, a novel on-device multimodal embedding system optimized for resource-limited mobile environments. RECALL achieves high-throughput, accurate retrieval by generating coarse-grained embeddings and leveraging query-based filtering for refined retrieval. Experimental results demonstrate that RECALL delivers high-quality embeddings with superior throughput, all while operating unobtrusively with minimal memory and energy consumption.

cs.IR↗

arXiv2024

FedMoE: Personalized Federated Learning via Heterogeneous Mixture of Experts

As Large Language Models (LLMs) push the boundaries of AI capabilities, their demand for data is growing. Much of this data is private and distributed across edge devices, making Federated Learning (FL) a de-facto alternative for fine-tuning (i.e., FedLLM). However, it faces significant challenges due to the inherent heterogeneity among clients, including varying data distributions and diverse task types. Towards a versatile FedLLM, we replace traditional dense model with a sparsely-activated Mixture-of-Experts (MoE) architecture, whose parallel feed-forward networks enable greater flexibility. To make it more practical in resource-constrained environments, we present FedMoE, the efficient personalized FL framework to address data heterogeneity, constructing an optimal sub-MoE for each client and bringing the knowledge back to global MoE. FedMoE is composed of two fine-tuning stages. In the first stage, FedMoE simplifies the problem by conducting a heuristic search based on observed activation patterns, which identifies a suboptimal submodel for each client. In the second stage, these submodels are distributed to clients for further training and returned for server aggregating through a novel modular aggregation strategy. Meanwhile, FedMoE progressively adjusts the submodels to optimal through global expert recommendation. Experimental results demonstrate the superiority of our method over previous personalized FL methods.

cs.LG↗

arXiv2024

Mobile Foundation Model as Firmware

In today's landscape, smartphones have evolved into hubs for hosting a multitude of deep learning models aimed at local execution. A key realization driving this work is the notable fragmentation among these models, characterized by varied architectures, operators, and implementations. This fragmentation imposes a significant burden on the comprehensive optimization of hardware, system settings, and algorithms. Buoyed by the recent strides in large foundation models, this work introduces a pioneering paradigm for mobile AI: a collaborative management approach between the mobile OS and hardware, overseeing a foundational model capable of serving a broad spectrum of mobile AI tasks, if not all. This foundational model resides within the NPU and remains impervious to app or OS revisions, akin to firmware. Concurrently, each app contributes a concise, offline fine-tuned "adapter" tailored to distinct downstream tasks. From this concept emerges a concrete instantiation known as \sys. It amalgamates a curated selection of publicly available Large Language Models (LLMs) and facilitates dynamic data flow. This concept's viability is substantiated through the creation of an exhaustive benchmark encompassing 38 mobile AI tasks spanning 50 datasets, including domains such as Computer Vision (CV), Natural Language Processing (NLP), audio, sensing, and multimodal inputs. Spanning this benchmark, \sys unveils its impressive performance. It attains accuracy parity in 85\% of tasks, demonstrates improved scalability in terms of storage and memory, and offers satisfactory inference speed on Commercial Off-The-Shelf (COTS) mobile devices fortified with NPU support. This stands in stark contrast to task-specific models tailored for individual applications.

cs.AI↗

arXiv2024

FedRDMA: Communication-Efficient Cross-Silo Federated LLM via Chunked RDMA Transmission

Communication overhead is a significant bottleneck in federated learning (FL), which has been exaggerated with the increasing size of AI models. In this paper, we propose FedRDMA, a communication-efficient cross-silo FL system that integrates RDMA into the FL communication protocol. To overcome the limitations of RDMA in wide-area networks (WANs), FedRDMA divides the updated model into chunks and designs a series of optimization techniques to improve the efficiency and robustness of RDMA-based communication. We implement FedRDMA atop the industrial federated learning framework and evaluate it on a real-world cross-silo FL scenario. The experimental results show that \sys can achieve up to 3.8$\times$ speedup in communication efficiency compared to traditional TCP/IP-based FL systems.

cs.LG↗

arXiv2024

Lightweight Protection for Privacy in Offloaded Speech Understanding

Speech is a common input method for mobile embedded devices, but cloud-based speech recognition systems pose privacy risks. Disentanglement-based encoders, designed to safeguard user privacy by filtering sensitive information from speech signals, unfortunately require substantial memory and computational resources, which limits their use in less powerful devices. To overcome this, we introduce a novel system, XXX, optimized for such devices. XXX is built on the insight that speech understanding primarily relies on understanding the entire utterance's long-term dependencies, while privacy concerns are often linked to short-term details. Therefore, XXX focuses on selectively masking these short-term elements, preserving the quality of long-term speech understanding. The core of XXX is an innovative differential mask generator, grounded in interpretable learning, which fine-tunes the masking process. We tested XXX on the STM32H7 microcontroller, assessing its performance in various potential attack scenarios. The results show that XXX maintains speech understanding accuracy and privacy at levels comparable to existing encoders, but with a significant improvement in efficiency, achieving up to 53.3$\times$ faster processing and a 134.1$\times$ smaller memory footprint.

cs.SD↗

arXiv2024

Accelerating Vertical Federated Learning

Privacy, security and data governance constraints rule out a brute force process in the integration of cross-silo data, which inherits the development of the Internet of Things. Federated learning is proposed to ensure that all parties can collaboratively complete the training task while the data is not out of the local. Vertical federated learning is a specialization of federated learning for distributed features. To preserve privacy, homomorphic encryption is applied to enable encrypted operations without decryption. Nevertheless, together with a robust security guarantee, homomorphic encryption brings extra communication and computation overhead. In this paper, we analyze the current bottlenecks of vertical federated learning under homomorphic encryption comprehensively and numerically. We propose a straggler-resilient and computation-efficient accelerating system that reduces the communication overhead in heterogeneous scenarios by 65.26% at most and reduces the computation overhead caused by homomorphic encryption by 40.66% at most. Our system can improve the robustness and efficiency of the current vertical federated learning framework without loss of security.

cs.CR↗

arXiv2024

FwdLLM: Efficient FedLLM using Forward Gradient

Large Language Models (LLMs) are transforming the landscape of mobile intelligence. Federated Learning (FL), a method to preserve user data privacy, is often employed in fine-tuning LLMs to downstream mobile tasks, an approach known as FedLLM. Though recent efforts have addressed the network issue induced by the vast model size, they have not practically mitigated vital challenges concerning integration with mobile devices, such as significant memory consumption and sluggish model convergence. In response to these challenges, this work introduces FwdLLM, an innovative FL protocol designed to enhance the FedLLM efficiency. The key idea of FwdLLM to employ backpropagation (BP)-free training methods, requiring devices only to execute ``perturbed inferences''. Consequently, FwdLLM delivers way better memory efficiency and time efficiency (expedited by mobile NPUs and an expanded array of participant devices). FwdLLM centers around three key designs: (1) it combines BP-free training with parameter-efficient training methods, an essential way to scale the approach to the LLM era; (2) it systematically and adaptively allocates computational loads across devices, striking a careful balance between convergence speed and accuracy; (3) it discriminatively samples perturbed predictions that are more valuable to model convergence. Comprehensive experiments with five LLMs and three NLP tasks illustrate FwdLLM's significant advantages over conventional methods, including up to three orders of magnitude faster convergence and a 14.6x reduction in memory footprint. Uniquely, FwdLLM paves the way for federated learning of billion-parameter LLMs such as LLaMA on COTS mobile devices -- a feat previously unattained.

cs.AI↗

arXiv2023

Federated Few-Shot Learning for Mobile NLP

Natural language processing (NLP) sees rich mobile applications. To support various language understanding tasks, a foundation NLP model is often fine-tuned in a federated, privacy-preserving setting (FL). This process currently relies on at least hundreds of thousands of labeled training samples from mobile clients; yet mobile users often lack willingness or knowledge to label their data. Such an inadequacy of data labels is known as a few-shot scenario; it becomes the key blocker for mobile NLP applications. For the first time, this work investigates federated NLP in the few-shot scenario (FedFSL). By retrofitting algorithmic advances of pseudo labeling and prompt learning, we first establish a training pipeline that delivers competitive accuracy when only 0.05% (fewer than 100) of the training data is labeled and the remaining is unlabeled. To instantiate the workflow, we further present a system FeS, addressing the high execution cost with novel designs. (1) Curriculum pacing, which injects pseudo labels to the training workflow at a rate commensurate to the learning progress; (2) Representational diversity, a mechanism for selecting the most learnable data, only for which pseudo labels will be generated; (3) Co-planning of a model's training depth and layer capacity. Together, these designs reduce the training delay, client energy, and network traffic by up to 46.0$\times$, 41.2$\times$ and 3000.0$\times$, respectively. Through algorithm/system co-design, FFNLP demonstrates that FL can apply to challenging settings where most training samples are unlabeled.

cs.LG↗

arXiv2023

Towards Practical Few-shot Federated NLP

Transformer-based pre-trained models have emerged as the predominant solution for natural language processing (NLP). Fine-tuning such pre-trained models for downstream tasks often requires a considerable amount of labeled private data. In practice, private data is often distributed across heterogeneous mobile devices and may be prohibited from being uploaded. Moreover, well-curated labeled data is often scarce, presenting an additional challenge. To address these challenges, we first introduce a data generator for federated few-shot learning tasks, which encompasses the quantity and skewness of scarce labeled data in a realistic setting. Subsequently, we propose AUG-FedPrompt, a prompt-based federated learning system that exploits abundant unlabeled data for data augmentation. Our experiments indicate that AUG-FedPrompt can perform on par with full-set fine-tuning with a limited amount of labeled data. However, such competitive performance comes at a significant system cost.

cs.CL↗