Abstract Transforming explicit preference signals into implicit and parameterized behaviors is pivotal for enabling prompt-free, human-aligned generation and improving the usability, efficiency and robustness of large language models. However, existing methods align model preference well but still rely on explicit user instructions to convey specific preferences, leading to cumbersome user experiences and undermining natural, frictionless interaction with the model. To fill the gap between explicit and implicit preference representation, this paper introduces a theoretical framework that establishes both necessary and sufficient conditions for effective preference recognition. Based on this framework, we propose a novel Two-Stage Progressive Preference Transfer (TSPPT) method, which decomposes preference internalization into two manageable stages: preference representation learning and preference internalization transfer. The proposed method fills the gap between explicit and implicit preferences while maintaining the model’s general capabilities. The experiments across multiple models (Qwen2.5, Qwen3, Llama-3.2, DeepSeek-R1-Distill) and datasets (UltraFeedback, HelpSteer) demonstrate superior performance. The proposed method achieves 79.2% win rate on UltraFeedback (vs. 59.2–67.6% for baselines), substantial improvements on MT-Bench (7.86 vs. 7.34 for best baseline), and significant reductions in implicit social bias (0.165 vs. 0.185–0.325 for baselines). Notably, the method maintains comparable performance between implicit and explicit settings, confirming successful preference internalization.1

Differentiable architecture search (DARTS) suffers from over-parameterized supernetworks that make architecture search difficult to perform in complex detection scenarios. In this paper, a shareable attention-mask for object detection architecture search (SAM-Det) is proposed, which is designed to address the high computational cost and poor performance caused by search over-parameterization. To reduce the search cost, a series of convolutional candidate operations is merged by using a shareable attention mask to represent the operations as a single convolution with shared weights. Moreover, to reduce the interference of redundant features on the search, the features are shuffled in groups to learn the attentional importance, thus selecting the important channels to be transferred into the search space. The experimental results show that SAM-Det is able to obtain network architectures with better detection performance and lower search cost in different detection scenarios compared to other Search methods.

Automatic question-answering systems demonstrate valuable utility in the biomedical domain, improving the precision and efficiency of clinical decision-making significantly. Despite large-scale language models achieving notable success in general domains, even outperforming human-level performance in certain aspects, they are still faced with challenges such as data privacy and scarcity in the special domain. This study explores the method for efficient fine-tuning under resource-constrained conditions in the biomedical field. We propose a multi-stage fine-tuning approach that effectively improves the performance of pre-trained language models in biomedical question-answering tasks. Specially, A multi-prompt-based contrastive learning strategy and a multi-prompt self-consistency voting module are introduced, which improve the accuracy of QA tasks. The experiments on the PubMedQA dataset under reasoning-required settings indicate that our approach outperforms domain-specific pre-training models and achieves comparable performance with GPT-4, while the number of fine-tuned parameters is much less than the total parameters of the base model.

Biomedical retrieval-augmented generation (RAG) remains challenged by fragmented context and weak alignment between retrieved evidence and generative reasoning, leading to incoherent or hallucinated outputs. We propose BioCoRE, a continuity-aware retrieval and evidence-aligned generation framework that enables precise, controllable, and parameter-free evidence utilization. By preserving positional continuity, modeling inter-sentence dependencies, and integrating explicit instructions with latent anchor embeddings, BioCoRE bridges the gap between noisy retrieval and coherent generation. Experiments on biomedical question-answering (QA) benchmarks indicate consistent improvements in factual accuracy over domain-specific finetuned models, general-purpose large language models (LLMs), and state-of-the-art RAG methods. Ablation and visualization analyses further demonstrate that BioCoRE more effectively prioritizes and integrates retrieved evidence, achieving substantial gains in evidence-grounded biomedical reasoning.

To address the issue of resource inefficiency caused by traditional triggering methods in fault-tolerant control for nonlinear systems, this paper explores a self-triggered fault-tolerant control strategy for nonlinear systems affected by actuator faults. Firstly, the control gain matrix and self-triggered parameter matrix of the faulty system are obtained by solving the corresponding LMIs. Secondly, system resources are saving by designing a self-triggered mechanism to reduce sampling and computation burden. Thirdly, the Lyapunov function approach is applied to guarantee the stability of the self-triggered system in the presence of actuator faults. Finally, the proposed method’s effectiveness is demonstrated through simulation examples.

Addressing multi-hop reasoning of complex query effectively is a challenging task in information retrieval field. It demands the ability to retrieve and integrate dispersed knowledge across multiple documents dynamically while maintaining coherence in multi-step reasoning process. This study addresses these challenges with three primary contributions. It explores the integrating of large language models with graph-augmented retrieval methods for complex multihop reasoning. Moreover, the Memory-Driven Chain-of-Reasoning strategy is introduced, leveraging the memory of historical queries and results to optimize multi-step reasoning dynamically. Additionally, the Constraint-Aware Filtering in Chunked Window strategy is developed to improve retrieval precision by partitioning and filtering large retrieval windows based on query constraints. The experiments on public benchmarks indicate that our method substantially outperforms competitive approaches, achieving up to 13.8% and 14.0% improvements in EM and F1 on HotpotQA, 9.4% and 12.8% on MuSiQue, and 6.4% and 3.2% on 2WikiMultiHopQA, respectively.

ABSTRACT Large language models (LLMs) have made significant progress in many tasks, but they may also generate biased or misleading outputs. Alignment techniques address this issue by refining models to reflect human values, but high‐quality preference datasets are limited. This study introduces a method to train a high‐performance reward model (RM) by integrating open knowledge with human feedback. We construct the Open Knowledge and Human Feedback (OK‐HF) dataset, comprising 39.8 million open preference data entries and 30,000 human feedback entries. The dual‐stage aligning strategy is proposed to combine preference pre‐training with domain adaptation, leveraging multi‐objective optimization to enhance learning from both preference data and fine‐grained human feedback. The Open Knowledge and Human‐feedback Reward Model (OKH‐RM), designed with the dual‐stage aligning strategy on the OK‐HF dataset, demonstrates exceptional performance in aligning LLMs with human preferences. The experimental results show that OKH‐RM outperforms Llama2‐RM, Qwen‐RM and Ultra‐RM, particularly achieving an accuracy of 85.93% on the Stanford SHP dataset. The model has shown advanced capabilities in detecting low‐quality repetitive responses and mitigating biases related to response length.

Feature pyramids are widely adopted in visual detection models for capturing multiscale features of objects. However, the utilization of feature pyramids in practical object detection tasks is prone to complex background interference, resulting in suboptimal capture of discriminative multiscale foreground semantic features. In this article, a foreground capture feature pyramid network (FCFPN) for multiscale object detection is proposed, to address the problem of inadequate feature learning in complex backgrounds. FCFPN consists of a foreground dual attention (FDA) module and a pathway aggregation (PA) structure. Specifically, the FDA mechanism activates top-down foreground channel responses and lateral spatial foreground location features, so that channel and spatial foreground features are adequately captured. Then, the PA module adaptively learns the fusion weights of multiscale features at different levels of the feature pyramid, which enhances the complementarity of semantic information between different levels of the foreground feature maps. Since the fusion weights are learned adaptively based on different pyramid levels, the detection model accordingly retains the gained information of feature sizes and suppresses the conflicting information. The evaluations on public datasets and the self-built complex background dataset demonstrate that the detection average precision (AP) and the feature learning performance of the proposed method are superior compared with other FPNs, which proves the effectiveness of the proposed FCFPN.

The biomedical question-answering system usually provide accurate and real-time responses, which is crucial for clinical decision-making and scientific research. Although large language models achieve remarkable results in general question-answering tasks, they still face challenges in specialized fields. This paper proposes a novel framework called RAG-Chain, which aims to enhance the performance of general-domain large models on special biomedical reasoning and question-answering tasks. The RAG-Chain framework improves the knowledge retrieval and generation abilities of general models by a multi-stage processing of external knowledge and automatic construction of chain-of-thought templates combined with self-consistency validation process of choice shuffling. The experimental results show that RAG-Chain improves the accuracy of the baseline model by an average of 6.9% on the MedQA dataset without the need for pre-training or fine-tuning in biomedical fields, verifying its strong adaptability and effectiveness in different large language models.

Automatic high-quality medical report generation plays an important role in the clinical diagnosis, which can assist and reduce the burden of doctors. Previous works often adopt the Transformer based structure and focus on the work of report generation process. Few studies consider the importance of image clarity by utilizing image deblurring to improve the quality of generated reports. We propose the multi-scale medical image deblurring network based on U-Net, improving the deblurring performance and recovering clear images gradually. Specially, the encoded representations of medical images and corresponding texts are aligned accurately by cross-modal feature alignment module. The experiments conducted on the IU-Xray and MIMIC-CXR datasets provide evidence that the proposed method outperforms previous studies in terms of both NLG metrics and Clinical Evaluation (CE) metrics, and the generated reports contain the professional medical terminology for higher quality.

Distributed learning-based high-dimensional temporal modeling for the Industrial Internet of Things (IIoT) has become a prevailing trend. However, traditional distributed learning inefficiently extracts information by straightforward architects, resulting in low modeling accuracy and high communication costs. We propose a distributed hierarchical temporal graph learning (DHTGL) approach. In terminal equipment, we construct an adaptive hierarchical dilation convolutional network to dynamically capture spatiotemporal features by adjusting the dilation factor at each layer. Next, we construct adaptive graphs according to the connection similarity between dimensions to capture implicit connections. In the edge device, we design a node-edge graph distance calculation based on Gromov-Wasserstein distance to group feature graphs and construct representative cluster feature graphs. Edge devices upload cluster feature graphs to reduce communication costs while minimizing information loss. In the central server, we incorporate graph attention networks into graph neural networks for edge updating in training models on clustered feature graphs. Experiments using public IIoT datasets and the self-built IIoT platform demonstrate the effectiveness of DHTGL in comparison with common distributed learning approaches. The results confirm that DHTGL consumes fewer communications while achieving higher accuracies.

Ant colony optimization (ACO) algorithm is widely used in the instant delivery order scheduling because of its distributed computing capability. However, the order delivery efficiency decreases when different logistics statuses are faced. In order to improve the performance of ACO, an adaptive ACO algorithm based on real-time logistics features (AACO-RTLFs) is proposed. First, features are extracted from the event dimension, spatial dimension, and time dimension of the instant delivery to describe the real-time logistics status. Five key factors are further selected from the above three features to assist in problem modeling and ACO designing. Second, an adaptive instant delivery model is built considering the customer's acceptable delivery time. The acceptable time is calculated by emergency order mark and weather conditions in the event dimension feature. Third, an adaptive ACO algorithm is proposed to obtain the instant delivery order schedules. The parameters of the probability equation in ACO are adjusted according to the extracted key factors. Finally, the Gurobi solver in Python is used to perform numerical experiments on the classical datasets to verify the effectiveness of the instant delivery model. The proposed AACO-RTLF algorithm shows its advantages in instant delivery order scheduling when compared to the other state-of-the-art algorithms.

The automated question-answering system plays a crucial role in improving the accuracy and efficiency of clinical decision-making. While large-scale language models perform prominently in the general domain, even surpassing human performance in certain aspects, challenges such as data privacy and the massive training costs affect the broader adoption of LLMs in biomedical domain. This study explores the strategies of efficient fine-tuning and optimization methods in biomedical domain. We introduce a multi-stage fine-tuning strategy that improves the accuracy of medical question-answering tasks significantly. Specifically, a contrastive learning technique based on multi-prompts is proposed, and a self-consistency voting approach is used to improve the accuracy of reasoning-required tasks. Experimental results on PubMedQA dataset reveal that even fine-tuning only 0.152% of the baseline’s parameters, our method still improves its performance, making it outperform the domain-specific pre-trained models and achieve performance comparable to GPT-4.

Biomedical Question Answering aims to extract an answer to the given question from a biomedical context. Due to the strong professionalism of specific domain, it's more difficult to build large-scale datasets for specific domain question answering. Existing methods are limited by the lack of training data, and the performance is not as good as in open-domain settings, especially degrading when facing to the adversarial sample. We try to resolve the above issues. First, effective data augmentation strategies are adopted to improve the model training, including slide window, summarization and round-trip translation. Second, we propose a model weighting strategy for the final answer prediction in biomedical domain, which combines the advantage of two models, open-domain model QANet and BioBERT pre-trained in biomedical domain data. Finally, we give adversarial training to reinforce the robustness of the model. The public biomedical dataset collected from PubMed provided by BioASQ challenge is used to evaluate our approach. The results show that the model performance has been improved significantly compared to the single model and other models participated in BioASQ challenge. It can learn richer semantic expression from data augmentation and adversarial samples, which is beneficial to solve more complex question answering problems in biomedical domain.

Biomedical Question Answering aims to extract an answer to the given question from a biomedical context. Due to the strong professionalism of specific domain, it’s more difficult to build large-scale datasets for specific domain question answering. Existing methods are limited by the lack of training data, and the performance is not as good as in open-domain settings. We propose a model weighting strategy for the final answer prediction in biomedical domain, which combines the advantage of two models, open-domain model QANet and BioBERT pretrained in biomedical domain data. Especially, we adopt effective data augmentation strategies to improve the model performance, including round-trip translation and summarization. The public biomedical dataset collected from PubMed provided by BioASQ is used to evaluate our approach. The results show that the model performance has been improved significantly on BioASQ 6B, 7B and 8B datasets compared to the single model.

This paper is concerned with the task of biomedical Question Answering (QA) which refers to extracting an answer to the given question from a biomedical context. Current works have made progress on this task, but they are still severely restricted by the insufficient training data due to the domain-specific nature, which motivates us to further explore a powerful way to solve this problem. We propose a Hierarchical Question-Aware Context Learning (HQACL) model for the biomedical QA task constituted by multi-level attention. The interaction between the question and the context can be captured layer by layer, with multi-grained embeddings to strengthen the ability of the language representation. A special training method called DA, including two parts namely domain adaptation and data augmentation, is also introduced to enhance the model performance. Domain adaptation can be defined as pre-training on a large-scale open-domain dataset and fine-tuning on the small training set of the target domain. As for the data augmentation, the Round-trip translation method is adopted to create new data with various expressions, which almost doubles the training set. The public biomedical dataset collected from PubMed provided by BioASQ is used to evaluate our model. The results show that our approach is superior to the best recent solution and achieves a new state of the art.

BACKGROUND: Biomedical semantic indexing is important for information retrieval and many other research fields in bioinformatics. It annotates biomedical citations with Medical Subject Headings. In face of unbalanced category distribution in the training data, sampling methods are difficult to apply for semantic indexing task. RESULTS: In this paper, we present a novel deep serial multi-task learning model. The primary task treats the biomedical semantic indexing as a multi-label text classification issue that considers the relations of the labels. The auxiliary task is a regression task that predicts the MeSH number of the citation and provides hints for the network to make it converge faster. The experimental results on the BioASQ-Task5A open dataset show that our model outperforms the state-of-the-art solution "MTI", proposed by the US National Library of Medicine. Further, it not only achieves the highest precision among all the solutions in BioASQ-Task5A but also has faster convergence speed compared with some naive deep learning methods. CONCLUSIONS: Rather than parallel in an ordinary multi-task structure, the tasks in our model are serial and tightly coupled. It can achieve satisfied performance without any handcrafted feature.

Biomedical semantic indexing refers to annotating biomedical citations with Medical Subject Headings, which is crucial for texting mining, information retrieval and other researches in the field of bioinformatics. The traditional methods ignore the relations among labels and need complicated feature engineering. In this paper, we present a novel model with a deep serial multi-task learning structure, in which the semantic word embedding and bidirectional Gated Recurrent Unit are integrated in a multi-task learning paradigm. It differs from an ordinary multi-task structure in that the tasks in our model are serial and tightly coupled rather than parallel. The dataset of the 2017 BioASQ-Task5A is used to evaluate the performance. Without any handcrafted feature, our model outperforms MTI, the state-of-the-art solution proposed by the US National Library of Medicine. It also achieves the highest precision among all the solutions in 2017 BioASQ-Task5A, and converges faster than some naive deep learning methods.

The collaborative filtering (CF) technique has been widely used recently in recommendation systems. It needs historical data to give predictions. However, the data sparsity problem still exists. We propose a new item-based restricted Boltzmann machine (RBM) approach for CF and use the deep multilayer RBM network structure, which alleviates the data sparsity problem and has excellent ability to extract features. Each item is treated as a single RBM, and different items share the same weights and biases. The parameters are learned layer by layer in the deep network. The batch gradient descent algorithm with minibatch is used to increase the convergence speed. The new feature vector discovered by the multilayer RBM network structure is very effective in predicting a rating and achieves a better result. Experimental results on the data set of MovieLens show that the item-based multilayer RBM approach achieves the best performance, with a mean absolute error of 0.6424 and a root-mean-square error of 0.7843.

Most classic network entity sorting algorithms are implemented in a homogeneous network, and they are not applicable to a heterogeneous network. Registered patent history data denotes the innovations and the achievements in different research fields. In this paper, we present an iteration algorithm called inventor-ranking, to sort the influences of patent inventors in heterogeneous networks constructed based on their patent data. This approach is a flexible rule-based method, making full use of the features of network topology. We sort the inventors and patents by a set of rules, and the algorithm iterates continuously until it meets a certain convergence condition. We also give a detailed analysis of influential inventor’s interesting topics using a latent Dirichlet allocation (LDA) model. Compared with the traditional methods such as PageRank, our approach takes full advantage of the information in the heterogeneous network, including the relationship between inventors and the relationship between the inventor and the patent. Experimental results show that our method can effectively identify the inventors with high influence in patent data, and that it converges faster than PageRank.