Large language model (LLM)-driven multi-agent systems (MAS) are transforming how humans and AIs collaboratively generate ideas and artifacts. While existing surveys provide comprehensive overviews of MAS infrastructures, they largely overlook the dimension of creativity, including how novel outputs are generated and evaluated, how creativity informs agent personas, and how creative workflows are coordinated. This is the first survey dedicated to creativity in MAS. We focus on text and image generation tasks, and present:(1) a taxonomy of agent proactivity and persona design;(2) an overview of generation techniques, including divergent exploration, iterative refinement, and collaborative synthesis, as well as relevant datasets and evaluation metrics; and(3) a discussion of key challenges, such as inconsistent evaluation standards, insufficient bias mitigation, coordination conflicts, and the lack of unified benchmarks.This survey offers a structured framework and roadmap for advancing the development, evaluation, and standardization of creative MAS.

With the proliferation of digital learning, an increasing number of learners are engaging with audio-visual materials. For preschool and lower elementary students, whose literacy skills are still limited, knowledge acquisition relies more heavily on spoken and visual content. Traditional readability models were primarily developed for written texts, and their applicability to spoken materials remains uncertain. To address this issue, this study investigates the impact of different word segmentation tools and language models on the performance of automatic grade classification models for Chinese spoken materials. Support Vector Machines were employed for grade prediction, aiming to automatically determine the appropriate grade level of learning resources and assist learners in selecting suitable materials. The results show that language models with higher-dimensional word embeddings achieved better classification performance, with an accuracy of up to 61% and an adjacent accuracy of 76%. These findings may contribute to future digital learning platforms or educational resource recommendation systems by automatically providing students with appropriate listening materials to enhance learning outcomes.

This study employs several state-of-the-art techniques, including RoPE and Flash Attention, and leverages large-scale Chinese web corpora and encyclopedic data to pre-train an encoder model specifically designed for long text in Traditional Chinese. We evaluate the model on tasks such as reading comprehension and text classification, and the results show that its overall performance lags behind existing Chinese benchmarks. Through pseudo-perplexity analysis, we infer that the pre-training phase did not sufficiently capture the data distribution, potentially due to factors such as hyperparameters, convergence, and data quality. Although the results are suboptimal, this study still offers valuable experimental insights and directions for improving Chinese language model development.

Multi-behavior recommendation leverages auxiliary behaviors to effectively alleviate the sparsity of target behaviors. Existing approaches can be broadly categorized into two paradigms: sequential models that capture individual temporal dynamics but often omit cross-user information, and graph-based models that mine collaborative patterns yet lack temporal dependency modeling. To address these limitations, this paper proposes an integrated approach that combines sequential and graph modeling: the former focuses on learning temporal dependencies within user behavior sequences, while the latter captures cross-user behavior paths. By fusing the predictions from both components, the method achieves more accurate recommendations. Experiments on two e-commerce datasets, Taobao and RetailRocket, show that the integrated model outperforms the strong baseline MB-STR by about 1% in both HR@10 and NDCG@10. These results indicate that incorporating cross-user collaborative information consistently improves performance, even on top of strong sequential models.

本研究為參加 FSR-2025 客語語音辨識挑戰賽(Hakka ASR II)的技術報告,旨在推進客語自動語音辨識技術的發展。由於客語屬於低資源語言,且存在多種腔調,語音辨識面臨高度挑戰。我們以 Whisperlarge-v2 為骨幹模型,設計兩階段訓練流程:首先利用「Hakka Across Taiwan(HAT)」語料庫進行模型調適,以捕捉客語的一般聲學特徵;其次在賽事方提供的60 小時腔調語料上進行微調,以增強對目標資料的適應性。實驗發現,直接輸出客語漢字可達到良好的字錯率(CER),但由 於腔調差異與拼音規則變化多,拼音任務表現顯著下降。為解決此問題,我們以漢字模型的編碼器初始化拼音模型,並提出結合 RoBERTa 漢字轉拼音、腔調判斷與字典修正的後處理模組,期望可以在比賽中提升辨識的成效。

This paper constructs a Chinese dialogue-based information-seeking question answering dataset CMDQA, which is mainly applied to the scenario of getting Chinese movie related information. It contains 10K QA dialogs (40K turns in total). All questions and background documents are compiled from the Wikipedia via an Internet crawler. The answers to the questions are obtained via extracting the corresponding answer spans within the related text passage. In CMDQA, in addition to searching related documents, pronouns are also added to the question to better mimic the real dialog scenario. This dataset can test the individual performance of the information retrieval, the question answering and the question re-writing modules. This paper also provides a baseline system and shows its performance on this dataset. The experiments elucidate that it still has a big gap to catch the human performance. This dataset thus provides enough challenge for the researcher to conduct related research.

This paper presents a framework to answer the questions that require various kinds of inference mechanisms (such as Extraction, Entailment-Judgement, and Summarization). Most of the previous approaches adopt a rigid framework which handles only one inference mechanism. Only a few of them adopt several answer generation modules for providing different mechanisms; however, they either lack an aggregation mechanism to merge the answers from various modules, or are too complicated to be implemented with neural networks. To alleviate the problems mentioned above, we propose a divide-and-conquer framework, which consists of a set of various answer generation modules, a dispatch module, and an aggregation module. The answer generation modules are designed to provide different inference mechanisms, the dispatch module is used to select a few appropriate answer generation modules to generate answer candidates, and the aggregation module is employed to select the final answer. We test our framework on the 2020 Formosa Grand Challenge Contest dataset. Experiments show that the proposed framework outperforms the state-of-the-art Roberta-large model by about 11.4%.

Due to the development of deep learning, the natural language processing tasks have made great progresses by leveraging the bidirectional encoder representations from Transformers (BERT). The goal of information retrieval is to search the most relevant results for the user’s query from a large set of documents. Although BERT-based retrieval models have shown excellent results in many studies, these models usually suffer from the need for large amounts of computations and/or additional storage spaces. In view of the flaws, a BERT-based Siamese-structured retrieval model (BESS) is proposed in this paper. BESS not only inherits the merits of pre-trained language models, but also can generate extra information to compensate the original query automatically. Besides, the reinforcement learning strategy is introduced to make the model more robust. Accordingly, we evaluate BESS on three public-available corpora, and the experimental results demonstrate the efficiency of the proposed retrieval model.

This technical report aims at the ROCLING 2021 Shared Task: Dimensional Sentiment Analysis for Educational Texts. In order to predict the affective states of Chinese educational texts, we present a practical framework by employing pre-trained language models, such as BERT and MacBERT. Several valuable observations and analyses can be drawn from a series of experiments. From the results, we find that MacBERT-based methods can deliver better results than BERT-based methods on the verification set. Therefore, we average the prediction results of several models obtained using different settings as the final output.

In this paper, we proposed a BERT-based dimensional semantic analyzer, which is designed by incorporating with word-level information. Our model achieved three of the best results in four metrics on “ROCLING 2021 Shared Task: Dimensional Sentiment Analysis for Educational Texts”. We conducted a series of experiments to compare the effectiveness of different pre-trained methods. Besides, the results also proofed that our method can significantly improve the performances than classic methods. Based on the experiments, we also discussed the impact of model architectures and datasets.

In the context of natural language processing, representation learning has emerged as a newly active research subject because of its excellent performance in many applications. Learning representations of words is a pioneering study in this school of research. However, paragraph (or sentence and document) embedding learning is more suitable/reasonable for some tasks, such as sentiment classification and document summarization. Nevertheless, as far as we are aware, there is only a dearth of research focusing on launching unsupervised paragraph embedding methods. Classic paragraph embedding methods infer the representation of a given paragraph by considering all of the words occurring in the paragraph. Consequently, those stop or function words that occur frequently may mislead the embedding learning process to produce a misty paragraph representation. Motivated by these observations, our major contributions are twofold. First, we propose a novel unsupervised paragraph embedding method, named the essence vector (EV) model, which aims at not only distilling the most representative information from a paragraph but also excluding the general background information to produce a more informative low-dimensional vector representation for the paragraph. We evaluate the proposed EV model on benchmark sentiment classification and multi-document summarization tasks. The experimental results demonstrate the effectiveness and applicability of the proposed embedding method. Second, in view of the increasing importance of spoken content processing, an extension of the EV model, named the denoising essence vector (D-EV) model, is proposed. The D-EV model not only inherits the advantages of the EV model but also can infer a more robust representation for a given spoken paragraph against imperfect speech recognition. The utility of the D-EV model is evaluated on a spoken document summarization task, confirming the effectiveness of the proposed embedding method in relation to several well-practiced and state-of-the-art summarization methods.