Feng Yuan

Ling Team, Binwei Zeng, Chao Huang et al.

In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled Bǎilíng in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI.

Yifan Gao, Haoyue Li, Feng Yuan et al.

Foundation models pre-trained on large-scale natural image datasets offer a powerful paradigm for medical image segmentation. However, effectively transferring their learned representations for precise clinical applications remains a challenge. In this work, we propose Dino U-Net, a novel encoder-decoder architecture designed to exploit the high-fidelity dense features of the DINOv3 vision foundation model. Our architecture introduces an encoder built upon a frozen DINOv3 backbone, which employs a specialized adapter to fuse the model's rich semantic features with low-level spatial details. To preserve the quality of these representations during dimensionality reduction, we design a new fidelity-aware projection module (FAPM) that effectively refines and projects the features for the decoder. We conducted extensive experiments on seven diverse public medical image segmentation datasets. Our results show that Dino U-Net achieves state-of-the-art performance, consistently outperforming previous methods across various imaging modalities. Our framework proves to be highly scalable, with segmentation accuracy consistently improving as the backbone model size increases up to the 7-billion-parameter variant. The findings demonstrate that leveraging the superior, dense-pretrained features from a general-purpose foundation model provides a highly effective and parameter-efficient approach to advance the accuracy of medical image segmentation. The code is available at https://github.com/yifangao112/DinoUNet.

Feng Yuan, Yifan Gao, Wenbin Wu et al.

Accurate multi-modal medical image translation requires ha-rmonizing global anatomical semantics and local structural fidelity, a challenge complicated by intermodality information loss and structural distortion. We propose ABS-Mamba, a novel architecture integrating the Segment Anything Model 2 (SAM2) for organ-aware semantic representation, specialized convolutional neural networks (CNNs) for preserving modality-specific edge and texture details, and Mamba's selective state-space modeling for efficient long- and short-range feature dependencies. Structurally, our dual-resolution framework leverages SAM2's image encoder to capture organ-scale semantics from high-resolution inputs, while a parallel CNNs branch extracts fine-grained local features. The Robust Feature Fusion Network (RFFN) integrates these epresentations, and the Bidirectional Mamba Residual Network (BMRN) models spatial dependencies using spiral scanning and bidirectional state-space dynamics. A three-stage skip fusion decoder enhances edge and texture fidelity. We employ Efficient Low-Rank Adaptation (LoRA+) fine-tuning to enable precise domain specialization while maintaining the foundational capabilities of the pre-trained components. Extensive experimental validation on the SynthRAD2023 and BraTS2019 datasets demonstrates that ABS-Mamba outperforms state-of-the-art methods, delivering high-fidelity cross-modal synthesis that preserves anatomical semantics and structural details to enhance diagnostic accuracy in clinical applications. The code is available at https://github.com/gatina-yone/ABS-Mamba

Yifan Gao, Haoyue Li, Feng Yuan et al.

We present Camyla, a system for fully autonomous research within the scientific domain of medical image segmentation. Camyla transforms raw datasets into literature-grounded research proposals, executable experiments, and complete manuscripts without human intervention. Autonomous experimentation over long horizons poses three interrelated challenges: search effort drifts toward unpromising directions, knowledge from earlier trials degrades as context accumulates, and recovery from failures collapses into repetitive incremental fixes. To address these challenges, the system combines three coupled mechanisms: Quality-Weighted Branch Exploration for allocating effort across competing proposals, Layered Reflective Memory for retaining and compressing cross-trial knowledge at multiple granularities, and Divergent Diagnostic Feedback for diversifying recovery after underperforming trials. The system is evaluated on CamylaBench, a contamination-free benchmark of 31 datasets constructed exclusively from 2025 publications, under a strict zero-intervention protocol across two independent runs within a total of 28 days on an 8-GPU cluster. Across the two runs, Camyla generates more than 2,700 novel model implementations and 40 complete manuscripts, and surpasses the strongest per-dataset baseline selected from 14 established architectures, including nnU-Net, on 22 and 18 of 31 datasets under identical training budgets, respectively (union: 24/31). Senior human reviewers score the generated manuscripts at the T1/T2 boundary of contemporary medical imaging journals. Relative to automated baselines, Camyla outperforms AutoML and NAS systems on aggregate segmentation performance and exceeds six open-ended research agents on both task completion and baseline-surpassing frequency. These results suggest that domain-scale autonomous research is achievable in medical image segmentation.

Feng Yuan, Yifan Gao, Yuehua Ye et al.

Cross-modal medical image synthesis research focuses on reconstructing missing imaging modalities from available ones to support clinical diagnosis. Driven by clinical necessities for flexible modality reconstruction, we explore K to N medical generation, where three critical challenges emerge: How can we model the heterogeneous contributions of different modalities to various target tasks? How can we ensure fusion quality control to prevent degradation from noisy information? How can we maintain modality identity consistency in multi-output generation? Driven by these clinical necessities, and drawing inspiration from SAM2's sequential frame paradigm and clinicians' progressive workflow of incrementally adding and selectively integrating multi-modal information, we treat multi-modal medical data as sequential frames with quality-driven selection mechanisms. Our key idea is to "learn" adaptive weights for each modality-task pair and "memorize" beneficial fusion patterns through progressive enhancement. To achieve this, we design three collaborative modules: PreWeightNet for global contribution assessment, ThresholdNet for adaptive filtering, and EffiWeightNet for effective weight computation. Meanwhile, to maintain modality identity consistency, we propose the Causal Modality Identity Module (CMIM) that establishes causal constraints between generated images and target modality descriptions using vision-language modeling. Extensive experimental results demonstrate that our proposed Med-K2N outperforms state-of-the-art methods by significant margins on multiple benchmarks. Source code is available.

Hui Wu, Yi Gan, Feng Yuan et al.

Transformer based Large Language Models (LLMs) have been widely used in many fields, and the efficiency of LLM inference becomes hot topic in real applications. However, LLMs are usually complicatedly designed in model structure with massive operations and perform inference in the auto-regressive mode, making it a challenging task to design a system with high efficiency. In this paper, we propose an efficient LLM inference solution with low latency and high throughput. Firstly, we simplify the LLM decoder layer by fusing data movement and element-wise operations to reduce the memory access frequency and lower system latency. We also propose a segment KV cache policy to keep key/value of the request and response tokens in separate physical memory for effective device memory management, helping enlarge the runtime batch size and improve system throughput. A customized Scaled-Dot-Product-Attention kernel is designed to match our fusion policy based on the segment KV cache solution. We implement our LLM inference solution on Intel GPU and publish it publicly. Compared with the standard HuggingFace implementation, the proposed solution achieves up to 7x lower token latency and 27x higher throughput for some popular LLMs on Intel GPU.

Mingming Li, Fuqing Zhu, Feng Yuan et al.

Recently, relational metric learning methods have been received great attention in recommendation community, which is inspired by the translation mechanism in knowledge graph. Different from the knowledge graph where the entity-to-entity relations are given in advance, historical interactions lack explicit relations between users and items in recommender systems. Currently, many researchers have succeeded in constructing the implicit relations to remit this issue. However, in previous work, the learning process of the induction function only depends on a single source of data (i.e., user-item interaction) in a supervised manner, resulting in the co-occurrence relation that is free of any semantic information. In this paper, to tackle the above problem in recommender systems, we propose a joint Semantic-Enhanced Relational Metric Learning (SERML) framework that incorporates the semantic information. Specifically, the semantic signal is first extracted from the target reviews containing abundant item features and personalized user preferences. A novel regression model is then designed via leveraging the extracted semantic signal to improve the discriminative ability of original relation-based training process. On four widely-used public datasets, experimental results demonstrate that SERML produces a competitive performance compared with several state-of-the-art methods in recommender systems.

Manqing Dong, Feng Yuan, Lina Yao et al.

A common challenge for most current recommender systems is the cold-start problem. Due to the lack of user-item interactions, the fine-tuned recommender systems are unable to handle situations with new users or new items. Recently, some works introduce the meta-optimization idea into the recommendation scenarios, i.e. predicting the user preference by only a few of past interacted items. The core idea is learning a global sharing initialization parameter for all users and then learning the local parameters for each user separately. However, most meta-learning based recommendation approaches adopt model-agnostic meta-learning for parameter initialization, where the global sharing parameter may lead the model into local optima for some users. In this paper, we design two memory matrices that can store task-specific memories and feature-specific memories. Specifically, the feature-specific memories are used to guide the model with personalized parameter initialization, while the task-specific memories are used to guide the model fast predicting the user preference. And we adopt a meta-optimization approach for optimizing the proposed method. We test the model on two widely used recommendation datasets and consider four cold-start situations. The experimental results show the effectiveness of the proposed methods.

Manqing Dong, Feng Yuan, Lina Yao et al.

A significant remaining challenge for existing recommender systems is that users may not trust the recommender systems for either lack of explanation or inaccurate recommendation results. Thus, it becomes critical to embrace a trustworthy recommender system. This survey provides a systemic summary of three categories of trust-aware recommender systems: social-aware recommender systems that leverage users' social relationships; robust recommender systems that filter untruthful noises (e.g., spammers and fake information) or enhance attack resistance; explainable recommender systems that provide explanations of recommended items. We focus on the work based on deep learning techniques, an emerging area in the recommendation research.

Feng Yuan, Lina Yao, Boualem Benatallah

Cross-domain recommendation has long been one of the major topics in recommender systems. Recently, various deep models have been proposed to transfer the learned knowledge across domains, but most of them focus on extracting abstract transferable features from auxilliary contents, e.g., images and review texts, and the patterns in the rating matrix itself is rarely touched. In this work, inspired by the concept of domain adaptation, we proposed a deep domain adaptation model (DARec) that is capable of extracting and transferring patterns from rating matrices {\em only} without relying on any auxillary information. We empirically demonstrate on public datasets that our method achieves the best performance among several state-of-the-art alternative cross-domain recommendation models.

Xiang Zhang, Lina Yao, Feng Yuan

Semi-supervised learning is sought for leveraging the unlabelled data when labelled data is difficult or expensive to acquire. Deep generative models (e.g., Variational Autoencoder (VAE)) and semisupervised Generative Adversarial Networks (GANs) have recently shown promising performance in semi-supervised classification for the excellent discriminative representing ability. However, the latent code learned by the traditional VAE is not exclusive (repeatable) for a specific input sample, which prevents it from excellent classification performance. In particular, the learned latent representation depends on a non-exclusive component which is stochastically sampled from the prior distribution. Moreover, the semi-supervised GAN models generate data from pre-defined distribution (e.g., Gaussian noises) which is independent of the input data distribution and may obstruct the convergence and is difficult to control the distribution of the generated data. To address the aforementioned issues, we propose a novel Adversarial Variational Embedding (AVAE) framework for robust and effective semi-supervised learning to leverage both the advantage of GAN as a high quality generative model and VAE as a posterior distribution learner. The proposed approach first produces an exclusive latent code by the model which we call VAE++, and meanwhile, provides a meaningful prior distribution for the generator of GAN. The proposed approach is evaluated over four different real-world applications and we show that our method outperforms the state-of-the-art models, which confirms that the combination of VAE++ and GAN can provide significant improvements in semisupervised classification.

Feng Yuan, Lina Yao, Boualem Benatallah

During the past decade, model-based recommendation methods have evolved from latent factor models to neural network-based models. Most of these techniques mainly focus on improving the overall performance, such as the root mean square error for rating predictions and hit ratio for top-N recommendation, where the users' feedback is considered as the ground-truth. However, in real-world applications, the users' feedback is possibly contaminated by imperfect user behaviours, namely, careless preference selection. Such data contamination poses challenges on the design of robust recommendation methods. In this work, to address the above issue, we propose a general adversial training framework for neural network-based recommendation models, which improves both the model robustness and the overall performance. We point out the tradeoffs between performance and robustness enhancement with detailed instructions on how to strike a balance. Specifically, we implement our approach on the collaborative auto-encoder, followed by experiments on three public available datasets: MovieLens-1M, Ciao, and FilmTrust. We show that our approach outperforms highly competitive state-of-the-art recommendation methods. In addition, we carry out a thorough analysis on the noise impacts, as well as the complex interactions between model nonlinearity and noise levels. Through simple modifications, our adversarial training framework can be applied to a host of neural network-based models whose robustness and performance are expected to be both enhanced.

Soumyabrata Dev, Shilpa Manandhar, Feng Yuan et al.

The analysis of clouds in the earth's atmosphere is important for a variety of applications, viz. weather reporting, climate forecasting, and solar energy generation. In this paper, we focus our attention on the impact of cloud on the total solar irradiance reaching the earth's surface. We use weather station to record the total solar irradiance. Moreover, we employ collocated ground-based sky camera to automatically compute the instantaneous cloud coverage. We analyze the relationship between measured solar irradiance and computed cloud coverage value, and conclude that higher cloud coverage greatly impacts the total solar irradiance. Such studies will immensely help in solar energy generation and forecasting.