Federated Customization of Large Models: Approaches, Experiments, and Insights

In this article, we explore federated customization of large models and highlight the key challenges it poses within the federated learning framework. We review several popular large model customization techniques, including full fine-tuning, efficient fine-tuning, prompt engineering, prefix-tuning, knowledge distillation, and retrieval-augmented generation. Then, we discuss how these techniques can be implemented within the federated learning framework. Moreover, we conduct experiments on federated prefix-tuning, which, to the best of our knowledge, is the first trial to apply prefix-tuning in the federated learning setting. The conducted experiments validate its feasibility with performance close to centralized approaches. Further comparison with three other federated customization methods demonstrated its competitive performance, satisfactory efficiency, and consistent robustness.

NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents

Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks fail to rigorously evaluate the long-horizon capabilities required to build complete software systems. Most prior evaluations focus on localized code generation, scaffolded completion, or short-term repair tasks, leaving open the question of whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we present NL2Repo Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation ability of coding agents. Given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, implement multi-module logic, and produce a fully installable Python library. Our experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved: even the strongest agents achieve below 40% average test pass rates and rarely complete an entire repository correctly. Detailed analysis uncovers fundamental long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. NL2Repo Bench establishes a rigorous, verifiable testbed for measuring sustained agentic competence and highlights long-horizon reasoning as a central bottleneck for the next generation of autonomous coding agents.

Virtual Width Networks

We introduce Virtual Width Networks (VWN), a framework that delivers the benefits of wider representations without incurring the quadratic cost of increasing the hidden size. VWN decouples representational width from backbone width, expanding the embedding space while keeping backbone compute nearly constant. In our large-scale experiment, an 8-times expansion accelerates optimization by over 2 times for next-token and 3 times for next-2-token prediction. The advantage amplifies over training as both the loss gap grows and the convergence-speedup ratio increases, showing that VWN is not only token-efficient but also increasingly effective with scale. Moreover, we identify an approximately log-linear scaling relation between virtual width and loss reduction, offering an initial empirical basis and motivation for exploring virtual-width scaling as a new dimension of large-model efficiency.

Hierarchical Federated Learning for Social Network with Mobility

Federated Learning (FL) offers a decentralized solution that allows collaborative local model training and global aggregation, thereby protecting data privacy. In conventional FL frameworks, data privacy is typically preserved under the assumption that local data remains absolutely private, whereas the mobility of clients is frequently neglected in explicit modeling. In this paper, we propose a hierarchical federated learning framework based on the social network with mobility namely HFL-SNM that considers both data sharing among clients and their mobility patterns. Under the constraints of limited resources, we formulate a joint optimization problem of resource allocation and client scheduling, which objective is to minimize the energy consumption of clients during the FL process. In social network, we introduce the concepts of Effective Data Coverage Rate and Redundant Data Coverage Rate. We analyze the impact of effective data and redundant data on the model performance through preliminary experiments. We decouple the optimization problem into multiple sub-problems, analyze them based on preliminary experimental results, and propose Dynamic Optimization in Social Network with Mobility (DO-SNM) algorithm. Experimental results demonstrate that our algorithm achieves superior model performance while significantly reducing energy consumption, compared to traditional baseline algorithms.

Assessing GPT Performance in a Proof-Based University-Level Course Under Blind Grading

As large language models (LLMs) advance, their role in higher education, particularly in free-response problem-solving, requires careful examination. This study assesses the performance of GPT-4o and o1-preview under realistic educational conditions in an undergraduate algorithms course. Anonymous GPT-generated solutions to take-home exams were graded by teaching assistants unaware of their origin. Our analysis examines both coarse-grained performance (scores) and fine-grained reasoning quality (error patterns). Results show that GPT-4o consistently struggles, failing to reach the passing threshold, while o1-preview performs significantly better, surpassing the passing score and even exceeding the student median in certain exercises. However, both models exhibit issues with unjustified claims and misleading arguments. These findings highlight the need for robust assessment strategies and AI-aware grading policies in education.

Privacy Meets Explainability: Managing Confidential Data and Transparency Policies in LLM-Empowered Science

As Large Language Models (LLMs) become integral to scientific workflows, concerns over the confidentiality and ethical handling of confidential data have emerged. This paper explores data exposure risks through LLM-powered scientific tools, which can inadvertently leak confidential information, including intellectual property and proprietary data, from scientists' perspectives. We propose "DataShield", a framework designed to detect confidential data leaks, summarize privacy policies, and visualize data flow, ensuring alignment with organizational policies and procedures. Our approach aims to inform scientists about data handling practices, enabling them to make informed decisions and protect sensitive information. Ongoing user studies with scientists are underway to evaluate the framework's usability, trustworthiness, and effectiveness in tackling real-world privacy challenges.

Adaptive Clipping for Privacy-Preserving Few-Shot Learning: Enhancing Generalization with Limited Data

In the era of data-driven machine-learning applications, privacy concerns and the scarcity of labeled data have become paramount challenges. These challenges are particularly pronounced in the domain of few-shot learning, where the ability to learn from limited labeled data is crucial. Privacy-preserving few-shot learning algorithms have emerged as a promising solution to address such pronounced challenges. However, it is well-known that privacy-preserving techniques often lead to a drop in utility due to the fundamental trade-off between data privacy and model performance. To enhance the utility of privacy-preserving few-shot learning methods, we introduce a novel approach called Meta-Clip. This technique is specifically designed for meta-learning algorithms, including Differentially Private (DP) model-agnostic meta-learning, DP-Reptile, and DP-MetaSGD algorithms, with the objective of balancing data privacy preservation with learning capacity maximization. By dynamically adjusting clipping thresholds during the training process, our Adaptive Clipping method provides fine-grained control over the disclosure of sensitive information, mitigating overfitting on small datasets and significantly improving the generalization performance of meta-learning models. Through comprehensive experiments on diverse benchmark datasets, we demonstrate the effectiveness of our approach in minimizing utility degradation, showcasing a superior privacy-utility trade-off compared to existing privacy-preserving techniques. The adoption of Adaptive Clipping represents a substantial step forward in the field of privacy-preserving few-shot learning, empowering the development of secure and accurate models for real-world applications, especially in scenarios where there are limited data availability.

Multi-Objective Optimization for Privacy-Utility Balance in Differentially Private Federated Learning

Federated learning (FL) enables collaborative model training across distributed clients without sharing raw data, making it a promising approach for privacy-preserving machine learning. However, ensuring differential privacy (DP) in FL presents challenges due to the trade-off between model utility and privacy protection. Clipping gradients before aggregation is a common strategy to limit privacy loss, but selecting an optimal clipping norm is non-trivial, as excessively high values compromise privacy, while overly restrictive clipping degrades model performance. In this work, we propose an adaptive clipping mechanism that dynamically adjusts the clipping norm using a multi-objective optimization framework. By integrating privacy and utility considerations into the optimization objective, our approach balances privacy preservation with model accuracy. We theoretically analyze the convergence properties of our method and demonstrate its effectiveness through extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 datasets. Our results show that adaptive clipping consistently outperforms fixed-clipping baselines, achieving improved accuracy under the same privacy constraints. This work highlights the potential of dynamic clipping strategies to enhance privacy-utility trade-offs in differentially private federated learning.

Do Fairness Interventions Come at the Cost of Privacy: Evaluations for Binary Classifiers

While in-processing fairness approaches show promise in mitigating biased predictions, their potential impact on privacy leakage remains under-explored. We aim to address this gap by assessing the privacy risks of fairness-enhanced binary classifiers via membership inference attacks (MIAs) and attribute inference attacks (AIAs). Surprisingly, our results reveal that enhancing fairness does not necessarily lead to privacy compromises. For example, these fairness interventions exhibit increased resilience against MIAs and AIAs. This is because fairness interventions tend to remove sensitive information among extracted features and reduce confidence scores for the majority of training data for fairer predictions. However, during the evaluations, we uncover a potential threat mechanism that exploits prediction discrepancies between fair and biased models, leading to advanced attack results for both MIAs and AIAs. This mechanism reveals potent vulnerabilities of fair models and poses significant privacy risks of current fairness methods. Extensive experiments across multiple datasets, attack methods, and representative fairness approaches confirm our findings and demonstrate the efficacy of the uncovered mechanism. Our study exposes the under-explored privacy threats in fairness studies, advocating for thorough evaluations of potential security vulnerabilities before model deployments.

VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation

We present a general strategy to aligning visual generation models -- both image and video generation -- with human preference. To start with, we build VisionReward -- a fine-grained and multi-dimensional reward model. We decompose human preferences in images and videos into multiple dimensions, each represented by a series of judgment questions, linearly weighted and summed to an interpretable and accurate score. To address the challenges of video quality assessment, we systematically analyze various dynamic features of videos, which helps VisionReward surpass VideoScore by 17.2% and achieve top performance for video preference prediction. Based on VisionReward, we develop a multi-objective preference learning algorithm that effectively addresses the issue of confounding factors within preference data. Our approach significantly outperforms existing image and video scoring methods on both machine metrics and human evaluation. All code and datasets are provided at https://github.com/THUDM/VisionReward.