Reasoning in Large Language Models (LLMs) has recently shown strong potential in enhancing generative recommendation through deep understanding of complex user preference. Existing approaches follow a reason-then-recommend paradigm, where LLMs perform step-by-step reasoning before item generation. However, this paradigm inevitably suffers from reasoning degradation (i.e., homogeneous or error-accumulated reasoning) due to the lack of intermediate verification, thus undermining the recommendation. To bridge this gap, we propose a novel reason-verify-recommend paradigm, which interleaves reasoning with verification to provide reliable feedback, guiding the reasoning process toward more faithful user preference understanding. To enable effective verification, we establish two key principles for verifier design: 1) reliability ensures accurate evaluation of reasoning correctness and informative guidance generation; and 2) multi-dimensionality emphasizes comprehensive verification across multi-dimensional user preferences. Accordingly, we propose an effective implementation called VRec. It employs a mixture of verifiers to ensure multi-dimensionality, while leveraging a proxy prediction objective to pursue reliability. Experiments on four real-world datasets demonstrate that VRec substantially enhances recommendation effectiveness and scalability without compromising efficiency.

Optimizing recommender systems for objectives beyond accuracy, such as diversity, novelty, and personalization, is crucial for long-term user satisfaction. To this end, industrial practitioners have accumulated vast amounts of structured domain knowledge, which we term human priors (e.g., item taxonomies, temporal patterns). This knowledge is typically applied through post-hoc adjustments during ranking or post-ranking. However, this approach remains decoupled from the core model learning, which is particularly undesirable as the industry shifts to end-to-end generative recommendation foundation models. On the other hand, many methods targeting these beyond-accuracy objectives often require architecture-specific modifications and discard these valuable human priors by learning user intent in a fully unsupervised manner. Instead of discarding the human priors accumulated over years of practice, we introduce a backbone-agnostic framework that seamlessly integrates these human priors directly into the end-to-end training of generative recommenders. With lightweight, prior-conditioned adapter heads inspired by efficient LLM decoding strategies, our approach guides the model to disentangle user intent along human-understandable axes (e.g., interaction types, long- vs. short-term interests). We also introduce a hierarchical composition strategy for modeling complex interactions across different prior types. Extensive experiments on three large-scale datasets demonstrate that our method significantly enhances both accuracy and beyond-accuracy objectives. We also show that human priors allow the backbone model to more effectively leverage longer context lengths and larger model sizes.

Peiqi Wang, ShengYun Peng, Xuewen Zhang, Hanchao Yu, Yibo Yang, Lifu Huang, Fujun Liu, Qifan Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

How well can Multimodal Large Language Models (MLLMs) understand composite images? Composite images (CIs) are synthetic visuals created by merging multiple visual elements, such as charts, posters, or screenshots, rather than being captured directly by a camera. While CIs are prevalent in real-world applications, recent MLLM developments have primarily focused on interpreting natural images (NIs). Our research reveals that current MLLMs face significant challenges in accurately understanding CIs, often struggling to extract information or perform complex reasoning based on these images. We find that existing training data for CIs are mostly formatted for question-answer tasks (e.g., in datasets like ChartQA and ScienceQA), while high-quality image-caption datasets, critical for robust vision-language alignment, are only available for NIs. To bridge this gap, we introduce Composite Captions (CompCap), a flexible framework that leverages Large Language Models (LLMs) and automation tools to synthesize CIs with accurate and detailed captions. Using CompCap, we curate CompCap-118K, a dataset containing 118K image-caption pairs across six CI types. We validate the effectiveness of CompCap-118K by supervised fine-tuning MLLMs of three sizes: xGen-MM-inst.-4B and LLaVA-NeXT-Vicuna-7B/13B. Empirical results show that CompCap-118K significantly enhances MLLMs' understanding of CIs, yielding average gains of 1.7%, 2.0%, and 2.9% across eleven benchmarks, respectively.

Recent years have witnessed success of sequential modeling, generative recommender, and large language model for recommendation. Though the scaling law has been validated for sequential models, it showed inefficiency in computational capacity when considering real-world applications like recommendation, due to the non-linear(quadratic) increasing nature of the transformer model. To improve the efficiency of the sequential model, we introduced a novel approach to sequential recommendation that leverages personalization techniques to enhance efficiency and performance. Our method compresses long user interaction histories into learnable tokens, which are then combined with recent interactions to generate recommendations. This approach significantly reduces computational costs while maintaining high recommendation accuracy. Our method could be applied to existing transformer based recommendation models, e.g., HSTU and HLLM. Extensive experiments on multiple sequential models demonstrate its versatility and effectiveness. Source code is available at https://github.com/facebookresearch/PerSRec.

The task of item-to-item (I2I) retrieval is to identify a set of relevant and highly engaging items based on a given trigger item. It is a crucial component in modern recommendation systems, where users' previously engaged items serve as trigger items to retrieve relevant content for future engagement. However, existing I2I retrieval models in industry are primarily built on co-engagement data and optimized using the recall measure, which overly emphasizes co-engagement patterns while failing to capture semantic relevance. This often leads to overfitting short-term co-engagement trends at the expense of long-term benefits such as discovering novel interests and promoting content diversity. To address this challenge, we propose MTMH, a Multi-Task and Multi-Head I2I retrieval model that achieves both high recall and semantic relevance. Our model consists of two key components: 1) a multi-task learning loss for formally optimizing the trade-off between recall and semantic relevance, and 2) a multi-head I2I retrieval architecture for retrieving both highly co-engaged and semantically relevant items. We evaluate MTMH using proprietary data from a commercial platform serving billions of users and demonstrate that it can improve recall by up to 14.4% and semantic relevance by up to 56.6% compared with prior state-of-the-art models. We also conduct live experiments to verify that MTMH can enhance both short-term consumption metrics and long-term user-experience-related metrics. Our work provides a principled approach for jointly optimizing I2I recall and semantic relevance, which has significant implications for improving the overall performance of recommendation systems.

AI-aided dermatological diagnosis models require large amounts of high-quality data for effective training. While mobile devices make data collection easier, the lack of professional annotations and privacy concerns hinder their use in model training. Semi-supervised Federated Learning allows for collaborative training on both labeled and unlabeled data while protecting privacy. However, existing pseudo-labeling methods struggle with non-IID dermatosis data from different clients. To address this, we propose a multi-stage hierarchical semi-supervised federated learning framework (HSSFL), enabling labeled clients to assist unlabeled ones in generating pseudo-labels and maximizing data use. HSSFL consists of three stages: federated client clustering, iterative self-training, and opportunistic fine-tuning. Using correlation awareness for federated client clustering, followed by an ally-prioritized approach for self-training, enhances the reliability of pseudo-labels. Opportunistic fine-tuning further minimizes label errors. Experiments on both general and real-world dermatological datasets demonstrate that HSSFL outperforms state-of-the-art semi-supervised federated learning methods.

Federated learning (FL) enables multiple clients to collaboratively train a global model while keeping local data decentralized. Data heterogeneity (non-IID) across clients has imposed significant challenges to FL, which makes local models re-optimize towards their own local optima and forget the global knowledge, resulting in performance degradation and convergence slowdown. Many existing works have attempted to address the non-IID issue by adding an extra global-model-based regularizing item to the local training but without an adaption scheme, which is not efficient enough to achieve high performance with deep learning models. In this paper, we propose a Selective Self-Distillation method for Federated learning (FedSSD), which imposes adaptive constraints on the local updates by self-distilling the global model’s knowledge and selectively weighting it by evaluating the credibility at both the class and sample level. The convergence guarantee of FedSSD is theoretically analyzed and extensive experiments are conducted on three public benchmark datasets, which demonstrates that FedSSD achieves better generalization and robustness in fewer communication rounds, compared with other state-of-the-art FL methods.

Federated learning (FL) has been an effective way to train a machine learning model distributedly, holding local data without exchanging them. However, due to the inaccessibility of local data, FL with label noise would be more challenging. Most existing methods assume only open-set or closed-set noise and correspondingly propose filtering or correction solutions, ignoring that label noise can be mixed in real-world scenarios. In this article, we propose a novel FL method to discriminate the type of noise and make the FL mixed noise-robust, named FedMIN. FedMIN employs a composite framework that captures local-global differences in multiparticipant distributions to model generalized noise patterns. By determining adaptive thresholds for identifying mixed label noise in each client and assigning appropriate weights during model aggregation, FedMIN enhances the performance of the global model. Furthermore, FedMIN incorporates a loss alignment mechanism using local and global Gaussian mixture models (GMMs) to mitigate the risk of revealing samplewise loss. Extensive experiments are conducted on several public datasets, which include the simulated FL testbeds, i.e., CIFAR-10, CIFAR-100, and SVHN, and the real-world ones, i.e., Camelyon17 and multiorgan nuclei challenge (MoNuSAC). Compared to FL benchmarks, FedMIN improves model accuracy by up to 9.9% due to its superior noise estimation capabilities.

Tiezheng Yu, Hanchao Yu, Davis Liang, Yuning Mao, Shaoliang Nie, Po-Yao Huang, Madian Khabsa, Pascale Fung, Yi-Chia Wang. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Qifan Wang, Yuning Mao, Jingang Wang, Hanchao Yu, Shaoliang Nie, Sinong Wang, Fuli Feng, Lifu Huang, Xiaojun Quan, Zenglin Xu, Dongfang Liu. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Recently deep neural networks have been applied to skin lesion recognition to learn feature representations, among which segmentation masks have been proved effective in enhancing the recognition performance by making the classifier network focus on the region of interest. However, high-quality segmentations for skin images require expert doctors and are a cumbersome task in terms of time and labor. Although some open-sourced segmentation models for similar tasks could help produce segmentation masks, employing them directly would bring unpredictable mismatches because of the concept drift effect. To tackle this issue, we propose an Uncertainty-based Fusion Network (UFN) for skin lesion diagnosis, which tries to robustly fuse a trained or open-sourced segmentation model with the target skin lesion classification task. To mitigate the negative influence of the segmentation mismatch, UFN uses the Dirichlet distribution to model the uncertainty and probabilities of multi-level representations. Then modified Dempster-Shafer theory is introduced to fuse the segmentation and classification representations adaptively. Moreover, the multi-level fused representations are unequally weighted by an attention mechanism. Furthermore, UFN integrates the label supervision for every fused representation as well as the weighted representations. The extensive experiments on two public benchmark datasets demonstrated the superiority of UFN compared with state-of-the-art methods.

Federated learning (FL) is a novel distributed learning framework where multiple participants collaboratively train a global model without sharing any raw data to preserve privacy. However, data quality may vary among the participants, the most typical of which is label noise. The incorrect label would significantly damage the performance of the global model. In FL, the inaccessibility of raw data makes this issue more challenging. Previously published studies are limited to using a task-specific benchmark-trained model to evaluate the relevance between the benchmark dataset in the server and the local one on the participants’ side. However, such approaches have failed to exploit the cooperative nature of FL itself and are not practical. This paper proposes a Consensus-based Label Correction approach (CLC) in FL, which tries to correct the noisy labels using the developed consensus method among the FL participants. The consensus-defined class-wise information is used to identify the noisy labels and correct them with pseudo-labels. Extensive experiments are conducted on several public datasets in various settings. The experimental results prove the advantage over the state-of-art methods. The link to the source code is https://github.com/bixiao-zeng/CLC.git .

Human Activity Recognition (HAR) is an important application of smart wearable/mobile systems for many human-centric problems such as healthcare. The multi-sensor synchronous measurement has shown better performance for HAR than a single sensor. However, the multi-sensor setting increases the costs of data transmission, computation and energy. Therefore, the efficient sensor selection to balance recognition accuracy and sensor cost is the critical challenge. In this paper, we propose an Instance-wise Dynamic Sensor Selection (IDSS) method for HAR. Firstly, we formalize this problem as minimizing both activity classification loss and sensor number by dynamically selecting a sparse subset for each instance. Then, IDSS solves the above minimization problem via Markov Decision Process whose policy for sensor selection is learned by exploiting the instance-wise states using Imitation Learning. In order to optimize the parameters of the activity classification model and the sensor selection policy, an algorithm named Mutual DAgger is proposed to alternatively enhance their learning process. To evaluate the performance of IDSS, we conduct experiments on three real-world HAR datasets. The experimental results show that IDSS can effectively reduce the overall sensor number without losing accuracy and outperforms the state-of-the-art methods regarding the combined measurement of accuracy and sensor number.

Motion estimation of cardiac MRI videos is crucial for the evaluation of human heart anatomy and function. Recent researches show promising results with deep learning-based methods. In clinical deployment, however, they suffer dramatic performance drops due to mismatched distributions between training and testing datasets, commonly encountered in the clinical environment. On the other hand, it is arguably impossible to collect all representative datasets and to train a universal tracker before deployment. In this context, we proposed a novel fast online adaptive learning (FOAL) framework: an online gradient descent based optimizer that is optimized by a meta-learner. The meta-learner enables the online optimizer to perform a fast and robust adaptation. We evaluated our method through extensive experiments on two public clinical datasets. The results showed the superior performance of FOAL in accuracy compared to the offline-trained tracking method. On average, the FOAL took only 0.4 second per video for online optimization.

Discriminating pathologic cognitive decline from the expected decline of normal aging is an important research topic for elderly care and health monitoring. However, most cognitive assessment methods only work when data distributions of the training set and testing set are consistent. Enabling existing cognitive assessment models to adapt to the data in new cognitive assessment tasks is a significant challenge. In this paper, we propose a novel domain adaptation method, namely the Fine-Grained Adaptation Random Forest (FAT), to bridge the cognitive assessment gap when the data distribution is changed. FAT is composed of two essential parts 1) information gain based model evaluation strategy (IGME) and 2) domain adaptation tree growing mechanism (DATG). IGME is used to evaluate every individual tree, and DATG is used to transfer the source model to the target domain. To evaluate the performance of FAT, we conduct experiments in real clinical environments. Experimental results demonstrate that FAT is significantly more accurate and efficient compared with other state-of-the-art methods.

Surface electromyography (sEMG) armband-based gesture recognition is an active research topic that aims to identify hand gestures with a single row of sEMG electrodes. As a typical type of biological signal, sEMG on one channel is nonstationary temporally and related to multiple adjacent muscles spatially, which hinders the effective representation in gesture recognition. To tackle these aspects, we propose a spatial-temporal features-based gesture recognition method (STF-GR) in this article. Specifically, STF-GR first decomposes the nonstationary multichannel sEMG by multivariate empirical mode decomposition, which jointly transforms each channel into a series of stationary subsignals. It can keep the temporal stationarity within-channel as well as the spatial independence across-channel. Then, by the convolutional recurrent neural network, STF-GR extracts and merges spatial-temporal features of decomposed sEMG signal. Finally, a negative log-likelihood-based cost function is used to make the final gesture decision. To evaluate the performance of STF-GR, we conduct experiments on three data sets, noninvasive adaptive hand prosthetic (NinaPro), CapgMyo, and BandMyo. The first two are publicly available, and BandMyo is collected by ourselves. Experimental evaluations with within-subject tests show that STF-GR exceeds the performance of other state-of-the-art methods, including deep learning algorithms that are not focused on spatial-temporal features and traditional machine learning algorithms that use handcrafted features.

In clinical settings, the assessment of Parkinson's disease (PD) mainly depends on the doctor's experience and observation. Such assessment often lacks of unified standards and may result in varied diagnoses among different doctors. To cope with this problem, we propose an objective and quantified symptom assessment tool of PD on mobile devices, based on the Unified Parkinson's Disease Rating Scale (UPDRS). The mobile PD assessment tool, PdAssit, assesses PD symptoms by actively delivering six tasks and generates scores equivalent to UPDRS using machines learning models. PdAssit is applied in three essential applications, including medication response detection, symptom self-tracking and diagnostic assistance. The feasibility and effectiveness of PdAssit is demonstrated through clinical experiments.

The practical application of Computed Tomography (CT) faces the dilemma between higher image resolution and less X-ray exposure for patients, motivating the research on CT super-resolution (SR). In this paper, we apply state-of-the-art SR techniques to reconstruct CT images using two proposed advanced CT SR models based on Convolutional Neural Networks (CNNs) and residual learning: a single-slice CT SR network (S-CTSRN), and a multi-slice CT SR network (M-CTSRN). S-CTSRN improves the high-frequency feature extraction by incorporating the residual learning strategy, while M-CTSRN further utilizes the coherence between neighboring CT slices for better SR reconstruction. We evaluate both models on a large-scale CT dataset1, and obtain competitive results both quantitatively and qualitatively.

In order to inspire and assist designers to create novel 3D animation modelings, an improved cellular genetic algorithm with evolutionary rules is proposed in this paper and applied in 3D animation modeling design. In this algorithm, ACIS rule expressions are used for deforming the initial 3D animation modeling created by Maya or 3D Max, and human-computer interaction which used expert knowledge to determine fitness function values of individuals is adopted to evaluate generated modelings. Tree-structure encoding is used for the generation and evolution of ACIS rule expressions and an evolutionary rule based on expert knowledge is introduced to reduce the number of human-computer interactions. Besides, a prototype system, called 3D Animation Modeling Design System, is developed based on the proposed algorithm. The experimental results show that the proposed algorithm can automatically and efficiently generate a series of creative 3D animation modelings and reduce the number of human-computer interactions to some extent.

This paper presents a multi-agent cooperative system for crowd animation. It analyses related work about crowd animation first. Then, a multi-agent crowd animation system architecture is introduced, which offers a promising framework for dynamically creating and managing agent communities in widely distributed environments. Next, a role modeling approach based on dynamic self-adaptive genetic algorithm and NURBS (Non Uniform Relational B Splines) technology is presented. Following, a group of fishes modelling example is illustrated for showing the modeling process in the system. Finally, the current work is summarised and an outlook for the future work is given.

In order to generate novel 3D animation modellings automatically, an interactive genetic algorithm HAIGA based on C/S mode was proposed. In HAIGA, HSF synergy technology and ACIS rule were brought in. The ACIS rule expression was expressed by three binary trees, which were used to scale 3D entities unevenly in the x-axis, y-axis and z-axis direction separately. New rule expressions, which were used to generate new modellings by evolving automatically based on the existing 3D animation modellings , were generated by selection, crossover, mutation on binary trees. A prototype system in which 3D animation modellings could evolve automatically was developed. Experiments in the system show that the proposed method can support cooperative design effectively to generate a series of novel 3D animation modellings.