The visual complexity of mobile interfaces presents a significant challenge to enhancing information retrieval efficiency and user experience among elderly users, as it influences their visual behavior. Using the main interface of a mobile audiobook recommendation application as a case study, this research categorizes interface visual complexity into two types: quantitative visual complexity and layout visual complexity. An evaluation model, combining both objective data and subjective assessments, was developed through eye-tracking experiments to examine the impact of visual complexity on the visual behavior of elderly users. The findings indicate that under medium quantitative visual complexity, elderly users exhibit the highest proficiency in obtaining interface information, while under low to medium layout visual complexity, they demonstrate shorter search completion times and higher information processing efficiency. Additionally, the scanpath angles are most pronounced in the range of (0, 45] degrees, exhibiting the highest number of fixation points, which indicates greater visual attention from users.

Medical gene expression feature selection is challenged by the high-dimensional small-sample regime and strong co-expression redundancy, which often yields unstable subsets and brittle trade-offs between predictive performance and compactness. This paper proposes a two-stage framework, Class-guided Joint Hybrid Multi-objective Optimizer (CJHMO), that integrates structural candidate generation with wrapper-based multi-objective optimization. In stage one, a Multi-Scale Co-Expression Attention Network (MSCANet) constructs correlation graphs under multiple thresholds and extracts connected components as co-expression modules. Each module is summarized by its eigengene (the first principal component), and the correlation ratio is used to quantify the association between eigengenes and class labels, producing supervised module scores. These scores are converted into attention weights and propagated to genes for candidate ranking and screening. In stage two, we develop a Dual-Population Heterogeneous Multi-Objective optimizer (DPHMO), where a decomposition-based population emphasizes Pareto-front coverage and global exploration, while an elite-guided particle swarm focuses on local exploitation and refinement. The two populations share an external elite archive (EP) for cross-population information exchange and non-dominated solution maintenance, jointly minimizing classification error rate and feature selection rate. Experiments on multiple public benchmarks with several classifiers demonstrate that CJHMO achieves superior performance-compression trade-offs over representative multi-objective baselines, with improved Pareto quality reflected by HV and IGD.

Facilitated by mobile edge computing, client-edge-cloud hierarchical federated learning (HFL) enables communication-efficient model training in a widespread area but also incurs additional security and privacy challenges from intermediate model aggregations and remains vulnerable to the single point of failure issue. To tackle these challenges, we propose a blockchain-based HFL (BHFL) system that operates a permissioned blockchain among edge servers for model aggregation without the need for a centralized cloud server. The employment of blockchain, however, introduces additional overhead. To enable a compact and efficient workflow, we design a novel lightweight consensus algorithm, named Proof of Federated Edge Learning (PoFEL), to reuse computational work performed for local model training. Specifically, the leader node is selected by evaluating the intermediate FEL models from all edge servers instead of other additional mechanisms used solely for leader elections. This design thus improves the system efficiency compared with traditional BHFL frameworks. To prevent model plagiarism and bribery voting during the consensus process, we propose Hash-based Commitment and Digital Signature (HCDS) and Bayesian Truth Serum-based Voting (BTSV) schemes. Finally, we devise an incentive mechanism to motivate continuous contributions from clients to the learning task. Experimental results demonstrate that our proposed BHFL system with the corresponding consensus protocol and incentive mechanism achieves effectiveness, low computational cost, and fairness.

Entity alignment (EA) is crucial for integrating knowledge graphs (KGs) constructed from diverse sources. Conventional unsupervised EA approaches attempt to eliminate human intervention but often suffer from accuracy limitations. With the rise of large language models (LLMs), leveraging their capabilities for EA presents a promising direction. However, it introduces new challenges: formulating the LLM-based EA problem and extracting the background knowledge in LLMs to realize EA without human intervention. This paper proposes HLMEA, a novel hybrid language model-based unsupervised EA method. HLMEA formulates the EA task into a filtering and single-choice problem and synergistically integrates small language models (SLMs) and LLMs. Specifically, SLMs filter candidate entities based on textual representations generated from KG triples. Then, LLMs refine this selection to identify the most semantically aligned entities. An iterative self-training mechanism allows SLMs to distill knowledge from LLM outputs, enhancing the EA ability of hybrid language models in subsequent rounds cooperatively. We also conducted extensive experiments on benchmark datasets to evaluate HLMEA's performance. The results demonstrate that HLMEA significantly outperforms unsupervised and even supervised EA baselines, proving its potential for scalable and effective EA across large KGs. The code and data are available at \url{https://github.com/xnjin-ai/HLMEA}.

Large language models (LLMs) have achieved remarkable success in machine translation, demonstrating impressive performance across diverse languages. However, translationese—characterized by overly literal and unnatural translations—remains a persistent challenge in LLM-based translation systems. Despite their pre-training on vast corpora of natural utterances, LLMs exhibit translationese errors and generate unexpected unnatural translations, stemming from biases introduced during supervised fine-tuning (SFT). In this work, we systematically evaluate the prevalence of translationese in LLM-generated translations and investigate its roots during supervised training. We introduce methods to mitigate these biases, including polishing golden references and filtering unnatural training instances. Empirical evaluations demonstrate that these approaches significantly reduce translationese while improving translation naturalness, validated by human evaluations and automatic metrics. Our findings highlight the need for training-aware adjustments to optimize LLM translation outputs, paving the way for more fluent and target-language-consistent translations.

Dynamical systems theory provides a framework for analyzing iterative processes and evolution over time.Within such systems, repetitive transformations can lead to stable configurations, known as attractors, including fixed points and limit cycles.Applying this perspective to large language models (LLMs), which iteratively map input text to output text, provides a principled approach to characterizing long-term behaviors.Successive paraphrasing serves as a compelling testbed for exploring such dynamics, as paraphrases re-express the same underlying meaning with linguistic variation.Although LLMs are expected to explore a diverse set of paraphrases in the text space, our study reveals that successive paraphrasing converges to stable periodic states, such as 2period attractor cycles, limiting linguistic diversity.This phenomenon is attributed to the selfreinforcing nature of LLMs, as they iteratively favour and amplify certain textual forms over others.This pattern persists with increasing generation randomness or alternating prompts and LLMs.These findings underscore inherent constraints in LLM generative capability, while offering a novel dynamical systems perspective for studying their expressive potential.Our code is available here.

Zhilin Wang, Jiaqi Zeng, Olivier Delalleau, Daniel Egert, Ellie Evans, Hoo-Chang Shin, Felipe Soares, Yi Dong, Oleksii Kuchaiev. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Prototype network-based methods have made substantial progress in few-shot relation extraction (FSRE) by enhancing relation prototypes with relation descriptions. However, the distribution of relations and instances in distinct representation spaces isolates the constraints of relations on instances, making relation prototypes biased. In this paper, we propose an end-to-end partial order-centered hyperbolic representation learning (PO-HRL) framework, which imposes the constraints of relations on instances by modeling partial order in hyperbolic space, so as to effectively learn the distribution of instance representations. Specifically, we develop the hyperbolic supervised contrastive learning based on Lorentzian cosine similarity to align representations of relations and instances, and model the partial order by constraining instances to reside within the Lorentzian entailment cone of their respective relation. Experiments on three benchmark datasets show that PO-HRL outperforms the strong baselines, especially in 1-shot settings lacking relation descriptions.

Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. This reward model can then be used with REINFORCE algorithm (RLHF) to align an Instruct model to reach 85.0 on Arena Hard, which is No. 1 as of 1 Oct 2024. We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2#preferences-new---1-oct-2024 and openly release the trained Reward and Instruct models at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward and https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct .

We examine diverging preferences in human-labeled preference datasets. We develop a taxonomy of disagreement sources spanning ten categories across four high-level classes and find that the majority of disagreements are due to factors such as task underspecification or response style. Our findings challenge a standard assumption in reward modeling methods that annotator disagreements can be attributed to simple noise. We then explore how these findings impact two areas of LLM development: reward modeling training and evaluation. In our experiments, we demonstrate how standard reward modeling (e.g., Bradley-Terry) and LLM-as-Judge evaluation methods fail to account for divergence between annotators. These findings highlight challenges in LLM evaluations, which are greatly influenced by divisive features like response style, and in developing pluralistically aligned LLMs. To address these issues, we develop methods for identifying diverging preferences to mitigate their influence in evaluations and during LLM training.

In recent years, the scale and computational demand for deep neural network models have been continuously increasing, leading to a growing need for efficient model deployment methods. Model quantization technology is an efficient way for model compression. However, low-bit quantization of models may result in diminished model accuracy. Quantization-aware training (QAT) is a representative training method for alleviating the decline of model accuracy. Nevertheless, most QAT methods require a large amount of labeled datasets and lack generalizability for tasks with limited data. We propose a new approach by using self-training, a semi-supervised method, to meet the dataset requirements of QAT. Based on the analysis of the characteristics of self-training and quantization-aware training, we use a meta-learning module for pseudo-label dataset selection and employ quantization operations to introduce noise into the model to enhance training effectiveness. We conducted experimental evaluations of traditional quantization training methods and our method on four network models (ResNet-50, ResNet-101, MobileNet-v2, MobileNet-v3) with the Cifar100 dataset. Our method enhances the training performance of both full-precision and quantized models compared to traditional quantization training methods. Utilizing 30% of the labeled dataset, compared to QAT, our method demonstrates an average accuracy enhancement of 2.19% after model training, while the average accuracy decline during the transition from a full-precision model to an 8-bit model is diminished by 0.55%.

Preference datasets are essential for training general-domain, instruction-following language models with Reinforcement Learning from Human Feedback (RLHF). Each subsequent data release raises expectations for future data collection, meaning there is a constant need to advance the quality and diversity of openly available preference data. To address this need, we introduce HelpSteer3-Preference, a permissively licensed (CC-BY-4.0), high-quality, human-annotated preference dataset comprising of over 40,000 samples. These samples span diverse real-world applications of large language models (LLMs), including tasks relating to STEM, coding and multilingual scenarios. Using HelpSteer3-Preference, we train Reward Models (RMs) that achieve top performance on RM-Bench (82.4%) and JudgeBench (73.7%). This represents a substantial improvement (~10% absolute) over the previously best-reported results from existing RMs. We demonstrate HelpSteer3-Preference can also be applied to train Generative RMs and how policy models can be aligned with RLHF using our RMs.

Is it secure to measure the reliability of local models by similarity in federated learning (FL)? This paper delves into an unexplored security threat concerning applying similarity metrics, such as the$L_{2}$norm, Euclidean distance, and cosine similarity, in protecting FL. We first uncover the deficiencies of similarity metrics that high-dimensional local models, including benign and poisoned models, may be evaluated to have the same similarity while being significantly different in the parameter values. We then leverage this finding to devise a novel untargeted model poisoning attack, Faker, which launches the attack by simultaneously maximizing the evaluated similarity of the poisoned local model and the difference in the parameter values. Experimental results based on seven datasets and eight defenses show that Faker outperforms the state-of-the-art benchmark attacks by1.1-9.0Xin reducing accuracy and1.2-8.0Xin saving time cost, which even holds for the case of a single malicious client with limited knowledge about the FL system. Moreover, Faker can degrade the performance of the global model by attacking only once. We also preliminarily explore extending Faker to other attacks, such as backdoor attacks and Sybil attacks. Lastly, we provide a model evaluation strategy, called the similarity of partial parameters (SPP), to defend against Faker. Given that numerous mechanisms in FL utilize similarity metrics to assess local models, this work suggests that we should be vigilant regarding the potential risks of using these metrics. The code will be released soon.

Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, Yue Zhang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Jupinder Parmar, Shrimai Prabhumoye, Joseph Jennings, Bo Liu, Aastha Jhunjhunwala, Zhilin Wang, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Zhilin Wang, Yi Dong, Jiaqi Zeng, Virginia Adams, Makesh Narsimhan Sreedhar, Daniel Egert, Olivier Delalleau, Jane Scowcroft, Neel Kant, Aidan Swope, Oleksii Kuchaiev. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

High-quality preference datasets are essential for training reward models that can effectively guide large language models (LLMs) in generating high-quality responses aligned with human preferences.As LLMs become stronger and better aligned, permissively licensed preference datasets, such as Open Assistant, HH-RLHF, and HelpSteer need to be updated to remain effective for reward modeling.Methods that distil preference data from proprietary LLMs such as GPT-4 have restrictions on commercial usage imposed by model providers.To improve upon both generated responses and attribute labeling quality, we release HelpSteer2, a permissively licensed preference dataset (CC-BY-4.0). Using a powerful Nemotron-4-340B base model trained on HelpSteer2, we are able to achieve the SOTA score (92.0%) on Reward-Bench's primary dataset, outperforming currently listed open and proprietary models, as of June 12th, 2024.Notably, HelpSteer2 consists of only ten thousand response pairs, an order of magnitude fewer than existing preference datasets (e.g., HH-RLHF), which makes it highly efficient for training reward models. Our extensive experiments demonstrate that reward models trained with HelpSteer2 are effective in aligning LLMs. Additionally, we propose SteerLM 2.0, a model alignment approach that can effectively make use of the rich multi-attribute score predicted by our reward models. HelpSteer2 is available at https://huggingface.co/datasets/nvidia/HelpSteer2 and code is available at https://github.com/NVIDIA/NeMo-Aligner

With the booming of mobile edge computing (MEC) and blockchain-based blockchain-based federated learning (BCFL), more studies suggest deploying BCFL on edge servers. In this case, edge servers with restricted resources face the dilemma of serving both mobile devices for their offloading tasks and the BCFL system for model training and blockchain consensus without sacrificing the service quality to any side. To address this challenge, this article proposes a resource allocation scheme for edge servers to provide optimal services at the minimum cost. Specifically, we first analyze the energy consumption of the MEC and BCFL tasks, considering the completion time of each task as the service quality constraint. Then, we model the resource allocation challenge into a multivariate, multiconstraint, and convex optimization problem. While solving the problem in a progressive manner, we design two algorithms based on the alternating direction method of multipliers (ADMMs) in both homogeneous and heterogeneous situations, where equal and on-demand resource distribution strategies are, respectively, adopted. The validity of our proposed algorithms is proved via rigorous theoretical analysis. Moreover, the convergence and efficiency of our proposed resource allocation schemes are evaluated through extensive experiments.

Sarcasm, as a form of irony conveying mockery and contempt, has been widespread in social media such as Twitter and Weibo, where the sarcastic text is commonly characterized as an incongruity between the surface positive and negative situation.Naturally, it has an urgent demand to automatically identify sarcasm from social media, so as to illustrate people's real views toward specific targets.In this paper, we develop a novel sarcasm detection method, namely Sarcasm Detector with Augmentation of Potential Result and Reaction (SD-APRR).Inspired by the direct access view, we treat each sarcastic text as an incomplete version without latent content associated with implied negative situations, including the result and human reaction caused by its observable content.To fill the latent content, we estimate the potential result and human reaction for each given training sample by [xEffect] and [xReact] relations inferred by the pre-trained commonsense reasoning tool COMET, and integrate the sample with them as an augmented one.We can then employ those augmented samples to train the sarcasm detector, whose encoder is a graph neural network with a denoising module.We conduct extensive empirical experiments to evaluate the effectiveness of SD-APRR.The results demonstrate that SD-APRR can outperform strong baselines on benchmark datasets.

Mobile crowdsensing (MCS) counting on the mobility of massive workers helps the requestor accomplish various sensing tasks with more flexibility and lower cost. However, for the conventional MCS, the large consumption of communication resources for raw data transmission and high requirements on data storage and computing capability hinder potential requestors with limited resources from using MCS. To facilitate the widespread application of MCS, we propose a novelMCS learning frameworkleveraging on blockchain technology and the new concept of edge intelligence based on federated learning (FL), which involves four major entities, including requestors, blockchain, edge servers, and mobile devices as workers. Even though there exist several studies on blockchain-based MCS and blockchain-based FL, they cannot solve the essential challenges of MCS with respect to accommodating resource-constrained requestors or deal with the privacy concerns brought by the involvement of requestors and workers in the learning process. To fill the gaps, four main procedures, i.e., task publication, data sensing and submission, learning to return final results, and payment settlement and allocation, are designed to address major challenges brought by both internal and external threats, such as malicious edge servers and dishonest requestors. Specifically, a mechanism design-based data submission rule is proposed to guarantee the data privacy of mobile devices being truthfully preserved at edge servers; consortium blockchain-based FL is elaborated to secure the distributed learning process; and a cooperation-enforcing control strategy is devised to elicit full payment from the requestor. Extensive simulations are carried out to evaluate the performance of our designed schemes.

As edge computing faces increasingly severe data security and privacy issues of edge devices, a framework called federated edge learning (FEL) has recently been proposed to enable machine learning (ML) model training at the edge, ensuring communication efficiency and data privacy protection for edge devices. In this paradigm, the training efficiency has long been challenged by the heterogeneity of communication conditions, computing capabilities, and available data sets at devices. Currently, researchers focus on solving this challenge via device selection from the perspective of optimizing energy consumption or convergence speed. However, the consideration of any one of them is insufficient to guarantee the long-term system efficiency and stability. To fill the gap, we propose an optimization problem to simultaneously minimize the total energy consumption of selected devices and maximize the convergence speed of the global model for device selection in FEL, under the constraints of training data amount and time consumption. For the accurate calculation of energy consumption, we deploy online bandit learning to estimate the CPU-cycle frequency availability of each device, based on an efficient algorithm, named fast-convergent energy-efficient device selection (FCE2DS), is proposed to solve the optimization problem with a low level of time complexity. Through a series of comparative experiments, we evaluate the performance of the proposed FCE2DS scheme, verifying its high training accuracy and energy efficiency.

Blockchain-based federated learning (BCFL) has recently gained tremendous attention because of its advantages, such as decentralization and privacy protection of raw data. However, there has been few studies focusing on the allocation of resources for the participated devices (i.e., clients) in the BCFL system. Especially, in the BCFL framework where the FL clients are also the blockchain miners, clients have to train the local models, broadcast the trained model updates to the blockchain network, and then perform mining to generate new blocks. Since each client has a limited amount of computing resources, the problem of allocating computing resources to training and mining needs to be carefully addressed. In this paper, we design an incentive mechanism to help the model owner (MO) (i.e., the BCFL task publisher) assign each client appropriate rewards for training and mining, and then the client will determine the amount of computing power to allocate for each subtask based on these rewards using the two-stage Stackelberg game. After analyzing the utilities of the MO and clients, we transform the game model into two optimization problems, which are sequentially solved to derive the optimal strategies for both the MO and clients. Further, considering the fact that local training related information of each client may not be known by others, we extend the game model with analytical solutions to the incomplete information scenario. Extensive experimental results demonstrate the validity of our proposed schemes.

Advances in distributed machine learning can empower future communications and networking. The emergence of federated learning (FL) has provided an efficient framework for distributed machine learning, which, however, still faces many security challenges. Among them, model poisoning attacks have a significant impact on the security and performance of FL. Given that there have been many studies focusing on defending against model poisoning attacks, it is necessary to survey the existing work and provide insights to inspire future research. In this paper, we first classify defense mechanisms for model poisoning attacks into two categories: evaluation methods for local model updates and aggregation methods for the global model. Then, we analyze some of the existing defense strategies in detail. We also discuss some potential challenges and future research directions. To the best of our knowledge, we are the first to survey defense methods for model poisoning attacks in FL.

The importance of transaction fees in maintaining blockchain security and sustainability has been confirmed by extensive research, although they are not mandatory in most current blockchain systems. To enhance blockchain in the long term, it is crucial to design effective transaction pricing mechanisms. Different from the existing schemes based on auctions with more consideration about the profit of miners, we resort to game theory and propose a correlated equilibrium based transaction pricing mechanism through solving a pricing game among users with transactions, which can achieve both the individual and global optimum. To avoid the computational complexity exponentially increasing with the number of transactions, we further improve the game-theoretic solution with an approximate algorithm, which can derive almost the same results as the original one but costs significantly reduced time. We also propose a truthful assessment model for pricing mechanism to collect the feedback of users regarding the price suggestion. Extensive experimental results demonstrate the effectiveness and efficiency of our proposed mechanism.

In this paper, we propose an end-to-end feature fusion at-tention network (FFA-Net) to directly restore the haze-free image. The FFA-Net architecture consists of three key components:1) A novel Feature Attention (FA) module combines Channel Attention with Pixel Attention mechanism, considering that different channel-wise features contain totally different weighted information and haze distribution is uneven on the different image pixels. FA treats different features and pixels unequally, which provides additional flexibility in dealing with different types of information, expanding the representational ability of CNNs. 2) A basic block structure consists of Local Residual Learning and Feature Attention, Local Residual Learning allowing the less important information such as thin haze region or low-frequency to be bypassed through multiple local residual connections, let main network architecture focus on more effective information. 3) An Attention-based different levels Feature Fusion (FFA) structure, the feature weights are adaptively learned from the Feature Attention (FA) module, giving more weight to important features. This structure can also retain the information of shallow layers and pass it into deep layers.The experimental results demonstrate that our proposed FFA-Net surpasses previous state-of-the-art single image dehazing methods by a very large margin both quantitatively and qualitatively, boosting the best published PSNR metric from 30.23 dB to 36.39 dB on the SOTS indoor test dataset. Code has been made available at GitHub.