Zexi Li

Chenrui Wu, Zexi Li, Fangxin Wang et al.

Federated learning (FL) is a distributed framework for collaboratively training with privacy guarantees. In real-world scenarios, clients may have Non-IID data (local class imbalance) with poor annotation quality (label noise). The co-existence of label noise and class imbalance in FL's small local datasets renders conventional FL methods and noisy-label learning methods both ineffective. To address the challenges, we propose FedCNI without using an additional clean proxy dataset. It includes a noise-resilient local solver and a robust global aggregator. For the local solver, we design a more robust prototypical noise detector to distinguish noisy samples. Further to reduce the negative impact brought by the noisy samples, we devise a curriculum pseudo labeling method and a denoise Mixup training strategy. For the global aggregator, we propose a switching re-weighted aggregation method tailored to different learning periods. Extensive experiments demonstrate our method can substantially outperform state-of-the-art solutions in mix-heterogeneous FL environments.

Didi Zhu, Yincuan Li, Junkun Yuan et al. · tencent-ai

Universal domain adaptation (UniDA) aims to transfer knowledge from the source domain to the target domain without any prior knowledge about the label set. The challenge lies in how to determine whether the target samples belong to common categories. The mainstream methods make judgments based on the sample features, which overemphasizes global information while ignoring the most crucial local objects in the image, resulting in limited accuracy. To address this issue, we propose a Universal Attention Matching (UniAM) framework by exploiting the self-attention mechanism in vision transformer to capture the crucial object information. The proposed framework introduces a novel Compressive Attention Matching (CAM) approach to explore the core information by compressively representing attentions. Furthermore, CAM incorporates a residual-based measurement to determine the sample commonness. By utilizing the measurement, UniAM achieves domain-wise and category-wise Common Feature Alignment (CFA) and Target Class Separation (TCS). Notably, UniAM is the first method utilizing the attention in vision transformer directly to perform classification tasks. Extensive experiments show that UniAM outperforms the current state-of-the-art methods on various benchmark datasets.

Bowen Tian, Caixue He, Jiemin Wu et al.

Editing complex, long-form knowledge in Large Language Models remains a significant challenge due to the difficulty of maintaining generation coherence. Existing autoregressive methods like AnyEdit alleviate length constraints but rely on Fixed-window Chunking, which disregards logical structure and compromises consistency. To address this, we present AnyEdit++, a structure-aware framework incorporating Bayes-Chunk, an adaptive segmentation mechanism that dynamically identifies semantic boundaries based on Bayesian Surprise. We underpin this approach with a theoretical framework establishing two key principles: (1) Structural Independence: we prove that cross-segment interference is minimized when anchor keys are geometrically orthogonal (a condition naturally satisfied by our surprisal-based boundaries but violated by fixed windows), and (2) Causal Locality: we demonstrate that updates injected at these semantic peaks yield strictly superior control compared to arbitrary split points. Extensive experiments across mathematical reasoning, code generation, and narrative tasks demonstrate that AnyEdit++ achieves superior performance and robustness compared to state-of-the-art baselines, validating that structural awareness is critical for effective long-form knowledge editing.

Zexi Li, Xinyi Shang, Rui He et al.

Data heterogeneity is an inherent challenge that hinders the performance of federated learning (FL). Recent studies have identified the biased classifiers of local models as the key bottleneck. Previous attempts have used classifier calibration after FL training, but this approach falls short in improving the poor feature representations caused by training-time classifier biases. Resolving the classifier bias dilemma in FL requires a full understanding of the mechanisms behind the classifier. Recent advances in neural collapse have shown that the classifiers and feature prototypes under perfect training scenarios collapse into an optimal structure called simplex equiangular tight frame (ETF). Building on this neural collapse insight, we propose a solution to the FL's classifier bias problem by utilizing a synthetic and fixed ETF classifier during training. The optimal classifier structure enables all clients to learn unified and optimal feature representations even under extremely heterogeneous data. We devise several effective modules to better adapt the ETF structure in FL, achieving both high generalization and personalization. Extensive experiments demonstrate that our method achieves state-of-the-art performances on CIFAR-10, CIFAR-100, and Tiny-ImageNet.

Zexi Li, Tao Lin, Xinyi Shang et al.

In federated learning (FL), weighted aggregation of local models is conducted to generate a global model, and the aggregation weights are normalized (the sum of weights is 1) and proportional to the local data sizes. In this paper, we revisit the weighted aggregation process and gain new insights into the training dynamics of FL. First, we find that the sum of weights can be smaller than 1, causing global weight shrinking effect (analogous to weight decay) and improving generalization. We explore how the optimal shrinking factor is affected by clients' data heterogeneity and local epochs. Second, we dive into the relative aggregation weights among clients to depict the clients' importance. We develop client coherence to study the learning dynamics and find a critical point that exists. Before entering the critical point, more coherent clients play more essential roles in generalization. Based on the above insights, we propose an effective method for Federated Learning with Learnable Aggregation Weights, named as FedLAW. Extensive experiments verify that our method can improve the generalization of the global model by a large margin on different datasets and models.

Zexi Li, Jiaxun Lu, Shuang Luo et al.

In federated learning (FL), clients may have diverse objectives, and merging all clients' knowledge into one global model will cause negative transfer to local performance. Thus, clustered FL is proposed to group similar clients into clusters and maintain several global models. In the literature, centralized clustered FL algorithms require the assumption of the number of clusters and hence are not effective enough to explore the latent relationships among clients. In this paper, without assuming the number of clusters, we propose a peer-to-peer (P2P) FL algorithm named PANM. In PANM, clients communicate with peers to adaptively form an effective clustered topology. Specifically, we present two novel metrics for measuring client similarity and a two-stage neighbor matching algorithm based Monte Carlo method and Expectation Maximization under the Gaussian Mixture Model assumption. We have conducted theoretical analyses of PANM on the probability of neighbor estimation and the error gap to the clustered optimum. We have also implemented extensive experiments under both synthetic and real-world clustered heterogeneity. Theoretical analysis and empirical experiments show that the proposed algorithm is superior to the P2P FL counterparts, and it achieves better performance than the centralized cluster FL method. PANM is effective even under extremely low communication budgets.

Zexi Li, Qunwei Li, Yi Zhou et al.

Federated learning (FL) is a popular way of edge computing that doesn't compromise users' privacy. Current FL paradigms assume that data only resides on the edge, while cloud servers only perform model averaging. However, in real-life situations such as recommender systems, the cloud server has the ability to store historical and interactive features. In this paper, our proposed Edge-Cloud Collaborative Knowledge Transfer Framework (ECCT) bridges the gap between the edge and cloud, enabling bi-directional knowledge transfer between both, sharing feature embeddings and prediction logits. ECCT consolidates various benefits, including enhancing personalization, enabling model heterogeneity, tolerating training asynchronization, and relieving communication burdens. Extensive experiments on public and industrial datasets demonstrate ECCT's effectiveness and potential for use in academia and industry.

Ailin Huang, Ang Li, Aobo Kong et al.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

Weijie Shi, Yanxi Chen, Zexi Li et al.

Reinforcement learning drives recent advances in LLM reasoning and agentic capabilities, yet current approaches struggle with both exploration and exploitation. Exploration suffers from low success rates on difficult tasks and high costs of repeated rollouts from scratch. Exploitation suffers from coarse credit assignment and training instability: Trajectory-level rewards penalize valid prefixes for later errors, and failure-dominated groups overwhelm the few positive signals, leaving optimization without constructive direction. To this end, we propose R$^3$L, Reflect-then-Retry Reinforcement Learning with Language-Guided Exploration, Pivotal Credit, and Positive Amplification. To synthesize high-quality trajectories, R$^3$L shifts from stochastic sampling to active synthesis via reflect-then-retry, leveraging language feedback to diagnose errors, transform failed attempts into successful ones, and reduce rollout costs by restarting from identified failure points. With errors diagnosed and localized, Pivotal Credit Assignment updates only the diverging suffix where contrastive signals exist, excluding the shared prefix from gradient update. Since failures dominate on difficult tasks and reflect-then-retry produces off-policy data, risking training instability, Positive Amplification upweights successful trajectories to ensure positive signals guide the optimization process. Experiments on agentic and reasoning tasks demonstrate 5\% to 52\% relative improvements over baselines while maintaining training stability. Our code is released at https://github.com/shiweijiezero/R3L.

Didi Zhu, Zexi Li, Min Zhang et al. · tsinghua

Large-scale vision-language (V-L) models have demonstrated remarkable generalization capabilities for downstream tasks through prompt tuning. However, the mechanisms behind the learned text representations are unknown, limiting further generalization gains, especially under class imbalance scenarios. Recent advances in the neural collapse (NC) phenomenon of vision-only models suggest that the optimal representation structure is the simplex ETF, which paves the way to study representations in V-L models. In this paper, we make the first attempt to use NC for examining the representations in V-L models via prompt tuning. It is found that NC optimality of text-to-image representations shows a positive correlation with downstream generalizability, which is more severe under class imbalance settings. To improve the representations, we propose Neural-collapse-anchored Prompt Tuning (NPT), a novel method that learns prompts with text and image representations that satisfy the same simplex ETF. NPT incorporates two regularization terms: language-modality collapse and multi-modality isomorphism; and it is compatible with other prompt tuning methods. Extensive experiments show that NPT can consistently help to improve existing prompt tuning techniques across 11 datasets for both balanced and imbalanced settings.

Ziyu Zhao, Tao Shen, Didi Zhu et al.

Low-Rank Adaptation (LoRA) has emerged as a popular technique for fine-tuning large language models (LLMs) to various domains due to its modular design and widespread availability on platforms like Huggingface. This modularity has sparked interest in combining multiple LoRAs to enhance LLM capabilities. However, existing methods for LoRA composition primarily focus on task-specific adaptations that require additional training, and current model merging techniques often fail to fully leverage LoRA's modular nature, leading to parameter interference and performance degradation. In this paper, we investigate the feasibility of disassembling and reassembling multiple LoRAs at a finer granularity, analogous to assembling LEGO blocks. We introduce the concept of Minimal Semantic Units (MSUs), where the parameters corresponding to each rank in LoRA function as independent units. These MSUs demonstrate permutation invariance and concatenation-summation equivalence properties, enabling flexible combinations to create new LoRAs. Building on these insights, we propose the LoRA-LEGO framework. This framework conducts rank-wise parameter clustering by grouping MSUs from different LoRAs into $k$ clusters. The centroid of each cluster serves as a representative MSU, enabling the assembly of a merged LoRA with an adjusted rank of $k$. Additionally, we apply a dual reweighting strategy to optimize the scale of the merged LoRA. Experiments across various benchmarks demonstrate that our method outperforms existing approaches in LoRA merging.

Rui Ye, Wenhao Wang, Jingyi Chai et al.

Trained on massive publicly available data, large language models (LLMs) have demonstrated tremendous success across various fields. While more data contributes to better performance, a disconcerting reality is that high-quality public data will be exhausted in a few years. In this paper, we offer a potential next step for contemporary LLMs: collaborative and privacy-preserving LLM training on the underutilized distributed private data via federated learning (FL), where multiple data owners collaboratively train a shared model without transmitting raw data. To achieve this, we build a concise, integrated, and research-friendly framework/codebase, named OpenFedLLM. It covers federated instruction tuning for enhancing instruction-following capability, federated value alignment for aligning with human values, and 7 representative FL algorithms. Besides, OpenFedLLM supports training on diverse domains, where we cover 8 training datasets; and provides comprehensive evaluations, where we cover 30+ evaluation metrics. Through extensive experiments, we observe that all FL algorithms outperform local training on training LLMs, demonstrating a clear performance improvement across a variety of settings. Notably, in a financial benchmark, Llama2-7B fine-tuned by applying any FL algorithm can outperform GPT-4 by a significant margin while the model obtained through individual training cannot, demonstrating strong motivation for clients to participate in FL. The code is available at https://github.com/rui-ye/OpenFedLLM.

Tao Shen, Didi Zhu, Ziyu Zhao et al.

The remarkable success of foundation models has been driven by scaling laws, demonstrating that model performance improves predictably with increased training data and model size. However, this scaling trajectory faces two critical challenges: the depletion of high-quality public data, and the prohibitive computational power required for larger models, which have been monopolized by tech giants. These two bottlenecks pose significant obstacles to the further development of AI. In this position paper, we argue that leveraging massive distributed edge devices can break through these barriers. We reveal the vast untapped potential of data and computational resources on massive edge devices, and review recent technical advancements in distributed/federated learning that make this new paradigm viable. Our analysis suggests that by collaborating on edge devices, everyone can participate in training large language models with small edge devices. This paradigm shift towards distributed training on edge has the potential to democratize AI development and foster a more inclusive AI community.

Peng Wang, Zexi Li, Ningyu Zhang et al.

Large language models (LLMs) need knowledge updates to meet the ever-growing world facts and correct the hallucinated responses, facilitating the methods of lifelong model editing. Where the updated knowledge resides in memories is a fundamental question for model editing. In this paper, we find that editing either long-term memory (direct model parameters) or working memory (non-parametric knowledge of neural network activations/representations by retrieval) will result in an impossible triangle -- reliability, generalization, and locality can not be realized together in the lifelong editing settings. For long-term memory, directly editing the parameters will cause conflicts with irrelevant pretrained knowledge or previous edits (poor reliability and locality). For working memory, retrieval-based activations can hardly make the model understand the edits and generalize (poor generalization). Therefore, we propose WISE to bridge the gap between memories. In WISE, we design a dual parametric memory scheme, which consists of the main memory for the pretrained knowledge and a side memory for the edited knowledge. We only edit the knowledge in the side memory and train a router to decide which memory to go through when given a query. For continual editing, we devise a knowledge-sharding mechanism where different sets of edits reside in distinct subspaces of parameters, and are subsequently merged into a shared memory without conflicts. Extensive experiments show that WISE can outperform previous model editing methods and overcome the impossible triangle under lifelong model editing of question answering, hallucination, and out-of-distribution settings across trending LLM architectures, e.g., GPT, LLaMA, and Mistral. Code is available at https://github.com/zjunlp/EasyEdit.

Lingzhi Gao, Zexi Li, Yang Lu et al.

Personalized federated learning (pFL) enables collaborative training among multiple clients to enhance the capability of customized local models. In pFL, clients may have heterogeneous (also known as non-IID) data, which poses a key challenge in how to decouple the data knowledge into generic knowledge for global sharing and personalized knowledge for preserving local personalization. A typical way of pFL focuses on label distribution skew, and they adopt a decoupling scheme where the model is split into a common feature extractor and two prediction heads (generic and personalized). However, such a decoupling scheme cannot solve the essential problem of feature skew heterogeneity, because a common feature extractor cannot decouple the generic and personalized features. Therefore, in this paper, we rethink the architecture decoupling design for feature-skew pFL and propose an effective pFL method called FediOS. In FediOS, we reformulate the decoupling into two feature extractors (generic and personalized) and one shared prediction head. Orthogonal projections are used for clients to map the generic features into one common subspace and scatter the personalized features into different subspaces to achieve decoupling for them. In addition, a shared prediction head is trained to balance the importance of generic and personalized features during inference. Extensive experiments on four vision datasets demonstrate our method reaches state-of-the-art pFL performances under feature skew heterogeneity.

Chen Hu, Haikuo Du, Heng Wang et al.

As LLMs shift toward autonomous agents, Deep Research has emerged as a pivotal metric. However, existing academic benchmarks like BrowseComp often fail to meet real-world demands for open-ended research, which requires robust skills in intent recognition, long-horizon decision-making, and cross-source verification. To address this, we introduce Step-DeepResearch, a cost-effective, end-to-end agent. We propose a Data Synthesis Strategy Based on Atomic Capabilities to reinforce planning and report writing, combined with a progressive training path from agentic mid-training to SFT and RL. Enhanced by a Checklist-style Judger, this approach significantly improves robustness. Furthermore, to bridge the evaluation gap in the Chinese domain, we establish ADR-Bench for realistic deep research scenarios. Experimental results show that Step-DeepResearch (32B) scores 61.4% on Scale AI Research Rubrics. On ADR-Bench, it significantly outperforms comparable models and rivals SOTA closed-source models like OpenAI and Gemini DeepResearch. These findings prove that refined training enables medium-sized models to achieve expert-level capabilities at industry-leading cost-efficiency.

Haohao Luo, Zexi Li, Yuexiang Xie et al.

Deep Research (DR) agents extend Large Language Models (LLMs) beyond parametric knowledge by autonomously retrieving and synthesizing evidence from large web corpora into long-form reports, enabling a long-horizon agentic paradigm. However, unlike real-time conversational assistants, DR is computationally expensive and time-consuming, creating an autonomy-interaction dilemma: high autonomy on ambiguous user queries often leads to prolonged execution with unsatisfactory outcomes. To address this, we propose IntentRL, a framework that trains proactive agents to clarify latent user intents before starting long-horizon research. To overcome the scarcity of open-ended research data, we introduce a scalable pipeline that expands a few seed samples into high-quality dialogue turns via a shallow-to-deep intent refinement graph. We further adopt a two-stage reinforcement learning (RL) strategy: Stage I applies RL on offline dialogues to efficiently learn general user-interaction behavior, while Stage II uses the trained agent and a user simulator for online rollouts to strengthen adaptation to diverse user feedback. Extensive experiments show that IntentRL significantly improves both intent hit rate and downstream task performance, outperforming the built-in clarify modules of closed-source DR agents and proactive LLM baselines.

Ruiyuan Zhang, Jiaxiang Liu, Zexi Li et al.

Geometric fracture assembly presents a challenging practical task in archaeology and 3D computer vision. Previous methods have focused solely on assembling fragments based on semantic information, which has limited the quantity of objects that can be effectively assembled. Therefore, there is a need to develop a scalable framework for geometric fracture assembly without relying on semantic information. To improve the effectiveness of assembling geometric fractures without semantic information, we propose a co-creation space comprising several assemblers capable of gradually and unambiguously assembling fractures. Additionally, we introduce a novel loss function, i.e., the geometric-based collision loss, to address collision issues during the fracture assembly process and enhance the results. Our framework exhibits better performance on both PartNet and Breaking Bad datasets compared to existing state-of-the-art frameworks. Extensive experiments and quantitative comparisons demonstrate the effectiveness of our proposed framework, which features linear computational complexity, enhanced abstraction, and improved generalization. Our code is publicly available at https://github.com/Ruiyuan-Zhang/CCS.

Yan Gao, Massimo Roberto Scamarcia, Javier Fernandez-Marques et al.

Large Language Models (LLMs) have achieved state-of-the-art results across diverse domains, yet their development remains reliant on vast amounts of publicly available data, raising concerns about data scarcity and the lack of access to domain-specific, sensitive information. Federated Learning (FL) presents a compelling framework to address these challenges by enabling decentralized fine-tuning on pre-trained LLMs without sharing raw data. However, the compatibility and performance of pre-trained LLMs in FL settings remain largely under explored. We introduce the FlowerTune LLM Leaderboard, a first-of-its-kind benchmarking suite designed to evaluate federated fine-tuning of LLMs across four diverse domains: general NLP, finance, medical, and coding. Each domain includes federated instruction-tuning datasets and domain-specific evaluation metrics. Our results, obtained through a collaborative, open-source and community-driven approach, provide the first comprehensive comparison across 26 pre-trained LLMs with different aggregation and fine-tuning strategies under federated settings, offering actionable insights into model performance, resource constraints, and domain adaptation. This work lays the foundation for developing privacy-preserving, domain-specialized LLMs for real-world applications.

Zexi Li, Jie Lin, Zhiqi Li et al.

Federated learning (FL) involves multiple heterogeneous clients collaboratively training a global model via iterative local updates and model fusion. The generalization of FL's global model has a large gap compared with centralized training, which is its bottleneck for broader applications. In this paper, we study and improve FL's generalization through a fundamental ``connectivity'' perspective, which means how the local models are connected in the parameter region and fused into a generalized global model. The term ``connectivity'' is derived from linear mode connectivity (LMC), studying the interpolated loss landscape of two different solutions (e.g., modes) of neural networks. Bridging the gap between LMC and FL, in this paper, we leverage fixed anchor models to empirically and theoretically study the transitivity property of connectivity from two models (LMC) to a group of models (model fusion in FL). Based on the findings, we propose FedGuCci(+), improving group connectivity for better generalization. It is shown that our methods can boost the generalization of FL under client heterogeneity across various tasks (4 CV datasets and 6 NLP datasets) and model architectures (e.g., ViTs and PLMs). The code is available here: \href{https://github.com/ZexiLee/fedgucci}{\faGithub~FedGuCci Codebase}.

Bowen Tian, Wenshuo Chen, Zexi Li et al.

How far are we really from automatically generating neural networks? While neural network weight generation shows promise, current approaches struggle with generalization to unseen tasks and practical application exploration. To address this, we propose T2W, a diffusion transformer framework that generates task-specific weights conditioned on natural language descriptions. T2W hierarchically processes network parameters into uniform blocks, integrates text embeddings from CLIP via a prior attention mechanism, and employs adversarial training with weight-space augmentation to enhance generalization. Experiments on Cifar100, Caltech256, and TinyImageNet demonstrate T2W's ability to produce high-quality weights for unseen tasks, outperforming optimization-based initialization and enabling novel applications such as weight enhancement and text-guided model fusion. Our work bridges textual semantics with weight-space dynamics, supported by an open-source dataset of text-weight pairs, advancing the practicality of generative models in neural network parameter synthesis. Our code is available on Github.

Didi Zhu, Zhongyi Sun, Zexi Li et al.

Catastrophic forgetting emerges as a critical challenge when fine-tuning multi-modal large language models (MLLMs), where improving performance on unseen tasks often leads to a significant performance drop on the original tasks. This paper presents a comprehensive analysis of catastrophic forgetting in MLLMs and introduces a post-training adjustment method called Model Tailor. Our method primarily preserves the pre-trained parameters while replacing a small number ($\leq$ 10\%) of fine-tuned parameters, maintaining $\sim$ 99\% effectiveness on original tasks versus pre-training, and achieving $\sim$ 97\% on new tasks compared to standard fine-tuning. Specifically, we derive a sparse mask to identify the "model patch", based on a fusion strategy that integrates salience and sensitivity analysis. Subsequently, a compensation mechanism is introduced to "decorate the patch", enhancing the model's performance on both target and original tasks. Additionally, our method is adaptable to multi-task scenarios. Through extensive experiments on InstructBLIP and LLaVA-1.5 in both image captioning and visual question answering tasks, our approach demonstrates significant task adaptability while preserving inherent pre-trained capabilities.

Jinge Wu, Hang Dong, Zexi Li et al.

Rare diseases pose significant challenges in diagnosis and treatment due to their low prevalence and heterogeneous clinical presentations. Unstructured clinical notes contain valuable information for identifying rare diseases, but manual curation is time-consuming and prone to subjectivity. This study aims to develop a hybrid approach combining dictionary-based natural language processing (NLP) tools with large language models (LLMs) to improve rare disease identification from unstructured clinical reports. We propose a novel hybrid framework that integrates the Orphanet Rare Disease Ontology (ORDO) and the Unified Medical Language System (UMLS) to create a comprehensive rare disease vocabulary. The proposed hybrid approach demonstrates superior performance compared to traditional NLP systems and standalone LLMs. Notably, the approach uncovers a significant number of potential rare disease cases not documented in structured diagnostic records, highlighting its ability to identify previously unrecognized patients.

Lorenzo Sani, Alex Iacob, Zeyu Cao et al.

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.

Ziyu Zhao, Yixiao Zhou, Zhi Zhang et al.

Low-Rank Adaptation (LoRA) is widely used for adapting large language models (LLMs) to specific domains due to its efficiency and modularity. Meanwhile, vanilla LoRA struggles with task conflicts in multi-task scenarios. Recent works adopt Mixture of Experts (MoE) by treating each LoRA module as an expert, thereby mitigating task interference through multiple specialized LoRA modules. While effective, these methods often isolate knowledge within individual tasks, failing to fully exploit the shared knowledge across related tasks. In this paper, we establish a connection between single LoRA and multi-LoRA MoE, integrating them into a unified framework. We demonstrate that the dynamic routing of multiple LoRAs is functionally equivalent to rank partitioning and block-level activation within a single LoRA. We further empirically demonstrate that finer-grained LoRA partitioning, within the same total and activated parameter constraints, leads to better performance gains across heterogeneous tasks. Building on these findings, we propose Single-ranked Mixture of Experts LoRA (\textbf{SMoRA}), which embeds MoE into LoRA by \textit{treating each rank as an independent expert}. With a \textit{dynamic rank-wise activation} mechanism, SMoRA promotes finer-grained knowledge sharing while mitigating task conflicts. Experiments demonstrate that SMoRA activates fewer parameters yet achieves better performance in multi-task scenarios.

Zexi Li, Lingzhi Gao, Chao Wu

Generative artificial intelligence (GenAI) has made significant progress in understanding world knowledge and generating content from human languages across various modalities, like text-to-text large language models, text-to-image stable diffusion, and text-to-video Sora. While in this paper, we investigate the capability of GenAI for text-to-model generation, to see whether GenAI can comprehend hyper-level knowledge embedded within AI itself parameters. Specifically, we study a practical scenario termed train-once-for-all personalization, aiming to generate personalized models for diverse end-users and tasks using text prompts. Inspired by the recent emergence of neural network diffusion, we present Tina, a text-conditioned neural network diffusion for train-once-for-all personalization. Tina leverages a diffusion transformer model conditioned on task descriptions embedded using a CLIP model. Despite the astronomical number of potential personalized tasks (e.g., $1.73\times10^{13}$), by our design, Tina demonstrates remarkable in-distribution and out-of-distribution generalization even trained on small datasets ($\sim 1000$). We further verify whether and how \Tina understands world knowledge by analyzing its capabilities under zero-shot/few-shot image prompts, different numbers of personalized classes, prompts of natural language descriptions, and predicting unseen entities.

Zexi Li, Zhiqi Li, Jie Lin et al.

Model fusion aims to integrate several deep neural network (DNN) models' knowledge into one by fusing parameters, and it has promising applications, such as improving the generalization of foundation models and parameter averaging in federated learning. However, models under different settings (data, hyperparameter, etc.) have diverse neuron permutations; in other words, from the perspective of loss landscape, they reside in different loss basins, thus hindering model fusion performances. To alleviate this issue, previous studies highlighted the role of permutation invariance and have developed methods to find correct network permutations for neuron alignment after training. Orthogonal to previous attempts, this paper studies training-time neuron alignment, improving model fusion without the need for post-matching. Training-time alignment is cheaper than post-alignment and is applicable in various model fusion scenarios. Starting from fundamental hypotheses and theorems, a simple yet lossless algorithm called TNA-PFN is introduced. TNA-PFN utilizes partially fixed neuron weights as anchors to reduce the potential of training-time permutations, and it is empirically validated in reducing the barriers of linear mode connectivity and multi-model fusion. It is also validated that TNA-PFN can improve the fusion of pretrained models under the setting of model soup (vision transformers) and ColD fusion (pretrained language models). Based on TNA-PFN, two federated learning methods, FedPFN and FedPNU, are proposed, showing the prospects of training-time neuron alignment. FedPFN and FedPNU reach state-of-the-art performances in federated learning under heterogeneous settings and can be compatible with the server-side algorithm.

Shanshan Yan, Zexi Li, Chao Wu et al.

Data heterogeneity, stemming from local non-IID data and global long-tailed distributions, is a major challenge in federated learning (FL), leading to significant performance gaps compared to centralized learning. Previous research found that poor representations and biased classifiers are the main problems and proposed neural-collapse-inspired synthetic simplex ETF to help representations be closer to neural collapse optima. However, we find that the neural-collapse-inspired methods are not strong enough to reach neural collapse and still have huge gaps to centralized training. In this paper, we rethink this issue from a self-bootstrap perspective and propose FedYoYo (You Are Your Own Best Teacher), introducing Augmented Self-bootstrap Distillation (ASD) to improve representation learning by distilling knowledge between weakly and strongly augmented local samples, without needing extra datasets or models. We further introduce Distribution-aware Logit Adjustment (DLA) to balance the self-bootstrap process and correct biased feature representations. FedYoYo nearly eliminates the performance gap, achieving centralized-level performance even under mixed heterogeneity. It enhances local representation learning, reducing model drift and improving convergence, with feature prototypes closer to neural collapse optimality. Extensive experiments show FedYoYo achieves state-of-the-art results, even surpassing centralized logit adjustment methods by 5.4\% under global long-tailed settings.

Zhenyan Lu, Daliang Xu, Dongqi Cai et al. · cambridge

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.

Tao Shen, Zexi Li, Didi Zhu et al.

Federated learning (FL) is a machine learning paradigm that allows multiple clients to collaboratively train a shared model without exposing their private data. Data heterogeneity is a fundamental challenge in FL, which can result in poor convergence and performance degradation. Client drift has been recognized as one of the factors contributing to this issue resulting from the multiple local updates in FedAvg. However, in cross-device FL, a different form of drift arises due to the partial client participation, but it has not been studied well. This drift, we referred as period drift, occurs as participating clients at each communication round may exhibit distinct data distribution that deviates from that of all clients. It could be more harmful than client drift since the optimization objective shifts with every round. In this paper, we investigate the interaction between period drift and client drift, finding that period drift can have a particularly detrimental effect on cross-device FL as the degree of data heterogeneity increases. To tackle these issues, we propose a predict-observe framework and present an instantiated method, FedEve, where these two types of drift can compensate each other to mitigate their overall impact. We provide theoretical evidence that our approach can reduce the variance of model updates. Extensive experiments demonstrate that our method outperforms alternatives on non-iid data in cross-device settings.

Yuxiang Xiao, Yang Hu, Bin Li et al.

Pathology foundation models (PFMs) have demonstrated strong representational capabilities through self-supervised pre-training on large-scale, unannotated histopathology image datasets. However, their diverse yet opaque pretraining contexts, shaped by both data-related and structural/training factors, introduce latent biases that hinder generalisability and transparency in downstream applications. In this paper, we propose AdaFusion, a novel prompt-guided inference framework that, to our knowledge, is among the very first to dynamically integrate complementary knowledge from multiple PFMs. Our method compresses and aligns tile-level features from diverse models and employs a lightweight attention mechanism to adaptively fuse them based on tissue phenotype context. We evaluate AdaFusion on three real-world benchmarks spanning treatment response prediction, tumour grading, and spatial gene expression inference. Our approach consistently surpasses individual PFMs across both classification and regression tasks, while offering interpretable insights into each model's biosemantic specialisation. These results highlight AdaFusion's ability to bridge heterogeneous PFMs, achieving both enhanced performance and interpretability of model-specific inductive biases.

Shuang Luo, Didi Zhu, Zexi Li et al.

Despite federated learning endows distributed clients with a cooperative training mode under the premise of protecting data privacy and security, the clients are still vulnerable when encountering adversarial samples due to the lack of robustness. The adversarial samples can confuse and cheat the client models to achieve malicious purposes via injecting elaborate noise into normal input. In this paper, we introduce a novel Ensemble Federated Adversarial Training Method, termed as EFAT, that enables an efficacious and robust coupled training mechanism. Our core idea is to enhance the diversity of adversarial examples through expanding training data with different disturbances generated from other participated clients, which helps adversarial training perform well in Non-IID settings. Experimental results on different Non-IID situations, including feature distribution skew and label distribution skew, show that our proposed method achieves promising results compared with solely combining federated learning with adversarial approaches.