Yun-Nung Chen

Bo-Han Feng, Yu-Hsuan Li Liang, Chien-Feng Liu et al.

Large Audio Language Models (LALMs) expand jailbreak risks from token-level prompting to the full speech perception-to-reasoning pipeline, where unsafe behavior can be induced through semantics, acoustic style, signal artifacts, or internal representations. Existing work studies these risks under heterogeneous threat models and evaluation protocols, making it difficult to compare attack practicality or defense utility. This paper provides a unified taxonomy and a controlled empirical evaluation of LALM jailbreak attacks and defenses. We organize prior work into semantic, acoustic, signal, and embedding-layer attacks; guard-based, training-free, and training-based defenses; and cross-modal, audio-native, and interactive benchmarks. We then evaluate representative attacks and defenses across ten open-source LALMs, measuring not only attack success rate but also benign refusal and latency. Our results show that Acoustic Best-of-N reveals strong worst-case audio-space vulnerabilities, Narrative Framing is an effective low-latency semantic threat, and current defenses trade robustness against benign usability. These findings support cost- and utility-aware evaluation as a necessary complement to success-rate-only LALM safety benchmarks.

Chao-Han Huck Yang, Taejin Park, Yuan Gong et al. · gatech

Given recent advances in generative AI technology, a key question is how large language models (LLMs) can enhance acoustic modeling tasks using text decoding results from a frozen, pretrained automatic speech recognition (ASR) model. To explore new capabilities in language modeling for speech processing, we introduce the generative speech transcription error correction (GenSEC) challenge. This challenge comprises three post-ASR language modeling tasks: (i) post-ASR transcription correction, (ii) speaker tagging, and (iii) emotion recognition. These tasks aim to emulate future LLM-based agents handling voice-based interfaces while remaining accessible to a broad audience by utilizing open pretrained language models or agent-based APIs. We also discuss insights from baseline evaluations, as well as lessons learned for designing future evaluations.

Chao-Wei Huang, Chen-Yu Hsu, Tsu-Yuan Hsu et al.

Conversational search provides a natural interface for information retrieval (IR). Recent approaches have demonstrated promising results in applying dense retrieval to conversational IR. However, training dense retrievers requires large amounts of in-domain paired data. This hinders the development of conversational dense retrievers, as abundant in-domain conversations are expensive to collect. In this paper, we propose CONVERSER, a framework for training conversational dense retrievers with at most 6 examples of in-domain dialogues. Specifically, we utilize the in-context learning capability of large language models to generate conversational queries given a passage in the retrieval corpus. Experimental results on conversational retrieval benchmarks OR-QuAC and TREC CAsT 19 show that the proposed CONVERSER achieves comparable performance to fully-supervised models, demonstrating the effectiveness of our proposed framework in few-shot conversational dense retrieval. All source code and generated datasets are available at https://github.com/MiuLab/CONVERSER

Yung-Yu Shih, Shang-Yu Su, Tzu-I Ho et al.

E-commerce platforms in emerging markets often operate with underdeveloped product catalogs that contain only category taxonomies but lack structured attribute schemas. This absence of fine-grained product attributes limits search capabilities -- preventing faceted filtering, degrading query understanding, and weakening semantic representations used by search systems. We present BEATS, a human-in-the-loop LLM framework for bootstrapping product attribute taxonomies entirely from scratch. Our approach extends a multi-stage LLM generation pipeline with two critical production stages: (1) proactive quality checking by model developers to filter erroneous outputs, and (2) human annotation by domain-expert local staff to validate generated attributes. The framework operates iteratively -- prompts at each generation stage are refined based on quality check observations and annotator feedback across successive rounds, progressively improving attribute quality. Once the attribute taxonomy is established, we employ LLMs to perform structured attribute tagging on individual product items, enriching their contextual representations. The enriched catalog directly benefits multiple components of the search system: enabling granular attribute-based filtering, providing structured features for ranking models, and improving semantic representations for dense retrieval. We validate the generated taxonomy by training dense retrieval models on attribute-enriched product data, demonstrating consistent improvements over baselines using original catalog information. Our system has been deployed at Rakuten Taiwan, enriching 9 major categories spanning 2,694 sub-categories with 67,277 generated attributes, and over 5.4 million products have been tagged with the generated attributes, with plans to enrich the entire product catalog.

Chao-Wei Huang, Shang-Chi Tsai, Yun-Nung Chen

Automatically classifying electronic health records (EHRs) into diagnostic codes has been challenging to the NLP community. State-of-the-art methods treated this problem as a multilabel classification problem and proposed various architectures to model this problem. However, these systems did not leverage the superb performance of pretrained language models, which achieved superb performance on natural language understanding tasks. Prior work has shown that pretrained language models underperformed on this task with the regular finetuning scheme. Therefore, this paper aims at analyzing the causes of the underperformance and developing a framework for automatic ICD coding with pretrained language models. We spotted three main issues through the experiments: 1) large label space, 2) long input sequences, and 3) domain mismatch between pretraining and fine-tuning. We propose PLMICD, a framework that tackles the challenges with various strategies. The experimental results show that our proposed framework can overcome the challenges and achieves state-of-the-art performance in terms of multiple metrics on the benchmark MIMIC data. The source code is available at https://github.com/MiuLab/PLM-ICD

Yen-Ting Lin, Yun-Nung Chen

In the realm of language models, the nuanced linguistic and cultural intricacies of Traditional Chinese, as spoken in Taiwan, have been largely overlooked. This paper introduces Taiwan LLM, a pioneering Large Language Model that specifically caters to the Traditional Chinese language, with a focus on the variant used in Taiwan. Leveraging a comprehensive pretraining corpus and instruction-finetuning datasets, we have developed a model that not only understands the complexities of Traditional Chinese but also embodies the cultural context of Taiwan. Taiwan LLM represents the first of its kind, a model that is not only linguistically accurate but also culturally resonant with its user base. Our evaluations demonstrate that Taiwan LLM achieves superior performance in understanding and generating Traditional Chinese text, outperforming existing models that are predominantly trained on Simplified Chinese or English. The open-source release of Taiwan LLM invites collaboration and further innovation, ensuring that the linguistic diversity of Chinese speakers is embraced and well-served. The model, datasets, and further resources are made publicly available to foster ongoing research and development in this field.

Ssu Chiu, Maolin Li, Yen-Ting Lin et al.

Dialogue systems are usually categorized into two types, open-domain and task-oriented. The first one focuses on chatting with users and making them engage in the conversations, where selecting a proper topic to fit the dialogue context is essential for a successful dialogue. The other one focuses on a specific task instead of casual talks, e.g., finding a movie on Friday night, or playing a song. These two directions have been studied separately due to their different purposes. However, how smoothly transitioning from social chatting to task-oriented dialogues is important for triggering business opportunities, and there is no public data focusing on such scenarios. Hence, this paper focuses on investigating the conversations starting from open-domain social chatting and then gradually transitioning to task-oriented purposes, and releases a large-scale dataset with detailed annotations for encouraging this research direction. To achieve this goal, this paper proposes a framework to automatically generate many dialogues without human involvement, in which any powerful open-domain dialogue generation model can be easily leveraged. The human evaluation shows that our generated dialogue data has a natural flow at a reasonable quality, showing that our released data has a great potential of guiding future research directions and commercial activities. Furthermore, the released models allow researchers to automatically generate unlimited dialogues in the target scenarios, which can greatly benefit semi-supervised and unsupervised approaches.

Cheng-Kuang Wu, Zhi Rui Tam, Chao-Chung Wu et al.

This study explores the proactive ability of LLMs to seek user support. We propose metrics to evaluate the trade-off between performance improvements and user burden, and investigate whether LLMs can determine when to request help under varying information availability. Our experiments show that without external feedback, many LLMs struggle to recognize their need for user support. The findings highlight the importance of external signals and provide insights for future research on improving support-seeking strategies. Source code: https://github.com/appier-research/i-need-help

Ting-Chun Wang, Shang-Yu Su, Yun-Nung Chen

Recommendation systems focus on helping users find items of interest in the situations of information overload, where users' preferences are typically estimated by the past observed behaviors. In contrast, conversational recommendation systems (CRS) aim to understand users' preferences via interactions in conversation flows. CRS is a complex problem that consists of two main tasks: (1) recommendation and (2) response generation. Previous work often tried to solve the problem in a modular manner, where recommenders and response generators are separate neural models. Such modular architectures often come with a complicated and unintuitive connection between the modules, leading to inefficient learning and other issues. In this work, we propose a unified framework based on BART for conversational recommendation, which tackles two tasks in a single model. Furthermore, we also design and collect a lightweight knowledge graph for CRS in the movie domain. The experimental results show that the proposed methods achieve the state-of-the-art performance in terms of both automatic and human evaluation.

Hung-Chieh Fang, Kuo-Han Hung, Chao-Wei Huang et al.

Open-domain conversational question answering can be viewed as two tasks: passage retrieval and conversational question answering, where the former relies on selecting candidate passages from a large corpus and the latter requires better understanding of a question with contexts to predict the answers. This paper proposes ConvADR-QA that leverages historical answers to boost retrieval performance and further achieves better answering performance. In our proposed framework, the retrievers use a teacher-student framework to reduce noises from previous turns. Our experiments on the benchmark dataset, OR-QuAC, demonstrate that our model outperforms existing baselines in both extractive and generative reader settings, well justifying the effectiveness of historical answers for open-domain conversational question answering.

Hui-Chi Kuo, Yun-Nung Chen

Intelligent virtual assistants are currently designed to perform tasks or services explicitly mentioned by users, so multiple related domains or tasks need to be performed one by one through a long conversation with many explicit intents. Instead, human assistants are capable of reasoning (multiple) implicit intents based on user utterances via commonsense knowledge, reducing complex interactions and improving practicality. Therefore, this paper proposes a framework of multi-domain dialogue systems, which can automatically infer implicit intents based on user utterances and then perform zero-shot prompting using a large pre-trained language model to trigger suitable single task-oriented bots. The proposed framework is demonstrated effective to realize implicit intents and recommend associated bots in a zero-shot manner.

Chun-Mao Lai, Ming-Hao Hsu, Chao-Wei Huang et al.

Prior work has demonstrated that data augmentation is useful for improving dialogue state tracking. However, there are many types of user utterances, while the prior method only considered the simplest one for augmentation, raising the concern about poor generalization capability. In order to better cover diverse dialogue acts and control the generation quality, this paper proposes controllable user dialogue act augmentation (CUDA-DST) to augment user utterances with diverse behaviors. With the augmented data, different state trackers gain improvement and show better robustness, achieving the state-of-the-art performance on MultiWOZ 2.1

Yu-Chen Den, Yung-Yu Shih, Zhi Rui Tam et al.

Generative retrieval (GR) maps queries directly to document identifiers (docids) using parametric knowledge, However, this design makes corpus expansion costly: adding new documents requires updating model parameters to encode new document-docid associations incurs repeated training and catastrophic forgetting of previously indexed documents. In this work, we revisit incremental GR as an in-context retrieval problem, where newly added documents are supplied as inference-time document-docid evidence. We propose ICICLE, an in-context indexing framework that performs source-aware docid generation over both parametric memory and context-provided document-docid pairs. ICICLE combines a `[COPY]`-based routing mechanism, preference-based calibration, and large context adaptation to distinguish context-grounded retrieval from parametric retrieval. Experiments on MS MARCO and NQ320K show that ICICLE improves retrieval of newly introduced documents while preserving seen-document retention without corpus-specific retraining. Our analysis further shows that high-shot degradation is mainly caused by routing failure, highlighting source-selection calibration as a key bottleneck for scaling in-context generative retrieval.

Wen-Yu Chang, Yun-Nung Chen

In recent research on dialogue systems and corpora, there has been a significant focus on two distinct categories: task-oriented (TOD) and open-domain (chit-chat) dialogues. TOD systems aim to satisfy specific user goals, such as finding a movie to watch, whereas open-domain systems primarily focus on generating engaging conversations. A recent study by Chiu et al. (2022) introduced SalesBot, which provides simulators and a dataset with one-turn transition from chit-chat to task-oriented dialogues. However, the previously generated data solely relied on BlenderBot, which raised concerns about its long-turn naturalness and consistency during a conversation. To address this issue, this paper aims to build SalesBot 2.0, a revised version of the published data, by leveraging the commonsense knowledge of large language models (LLMs) through proper prompting. The objective is to gradually bridge the gap between chit-chat and TOD towards better naturalness and consistency. The newly released large-scale dataset with detailed annotations exhibits smoother transitions between topics and is more human-like in terms of naturalness and consistency. It can serve as a valuable resource for both academic research and commercial applications. Furthermore, our proposed framework can be applied to generate numerous dialogues with various target intents.

Zhi Rui Tam, Cheng-Kuang Wu, Yi-Lin Tsai et al.

Structured generation, the process of producing content in standardized formats like JSON and XML, is widely utilized in real-world applications to extract key output information from large language models (LLMs). This study investigates whether such constraints on generation space impact LLMs abilities, including reasoning and domain knowledge comprehension. Specifically, we evaluate LLMs performance when restricted to adhere to structured formats versus generating free-form responses across various common tasks. Surprisingly, we observe a significant decline in LLMs reasoning abilities under format restrictions. Furthermore, we find that stricter format constraints generally lead to greater performance degradation in reasoning tasks.

Yung-Yu Shih, Ziwei Xu, Hiroya Takamura et al.

Question answering (QA) has been a long-standing focus in the NLP field, predominantly addressing reading comprehension and common sense QA. However, scenarios involving the preparation of answers to probable questions during professional oral presentations remain underexplored. In this paper, we pioneer the examination of this crucial yet overlooked topic by utilizing real-world QA conversation transcripts between company managers and professional analysts. We explore the proposed task using three causal knowledge graphs (KGs) and three large language models (LLMs). This work provides foundational insights into the application of LLMs in professional QA scenarios, highlighting the importance of causal KGs and perspective-taking in generating effective responses.

Chao-Wei Huang, Kai-Chou Yang, Zi-Yuan Chen et al.

With thousands of news articles from hundreds of sources distributed and shared every day, news consumption and information acquisition have been increasingly difficult for readers. Additionally, the content of news articles is becoming catchy or even inciting to attract readership, harming the accuracy of news reporting. We present Islander, an online news analyzing system. The system allows users to browse trending topics with articles from multiple sources and perspectives. We define several metrics as proxies for news quality, and develop algorithms for automatic estimation. The quality estimation results are delivered through a web interface to newsreaders for easy access to news and information. The website is publicly available at https://islander.cc/

Ya-Hsin Chang, Yun-Nung Chen

Spoken language understanding (SLU) is an essential task for machines to understand human speech for better interactions. However, errors from the automatic speech recognizer (ASR) usually hurt the understanding performance. In reality, ASR systems may not be easy to adjust for the target scenarios. Therefore, this paper focuses on learning utterance representations that are robust to ASR errors using a contrastive objective, and further strengthens the generalization ability by combining supervised contrastive learning and self-distillation in model fine-tuning. Experiments on three benchmark datasets demonstrate the effectiveness of our proposed approach.

Hsin-Yu Chang, Pei-Yu Chen, Tun-Hsiang Chou et al.

This paper provides a detailed survey of synthetic data techniques. We first discuss the expected goals of using synthetic data in data augmentation, which can be divided into four parts: 1) Improving Diversity, 2) Data Balancing, 3) Addressing Domain Shift, and 4) Resolving Edge Cases. Synthesizing data are closely related to the prevailing machine learning techniques at the time, therefore, we summarize the domain of synthetic data techniques into four categories: 1) Expert-knowledge, 2) Direct Training, 3) Pre-train then Fine-tune, and 4) Foundation Models without Fine-tuning. Next, we categorize the goals of synthetic data filtering into four types for discussion: 1) Basic Quality, 2) Label Consistency, and 3) Data Distribution. In section 5 of this paper, we also discuss the future directions of synthetic data and state three direction that we believe is important: 1) focus more on quality, 2) the evaluation of synthetic data, and 3) multi-model data augmentation.

Yen-Ting Lin, Hui-Chi Kuo, Ze-Song Xu et al.

This paper introduces Miutsu, National Taiwan University's Alexa Prize TaskBot, which is designed to assist users in completing tasks requiring multiple steps and decisions in two different domains -- home improvement and cooking. We overview our system design and architectural goals, and detail the proposed core elements, including question answering, task retrieval, social chatting, and various conversational modules. A dialogue flow is proposed to provide a robust and engaging conversation when handling complex tasks. We discuss the faced challenges during the competition and potential future work.

Tzu-Han Lin, Chen-An Li, Hung-yi Lee et al.

Reinforcement learning from human feedback (RLHF) is a popular strategy for aligning large language models (LLMs) with desired behaviors. Reward modeling is a crucial step in RLHF. However, collecting paired preference data for training reward models is often costly and time-consuming, especially for domain-specific preferences requiring expert annotation. To address this challenge, we propose the \textbf{Do}main knowled\textbf{ge} merged \textbf{R}eward \textbf{M}odel (DogeRM), a novel framework that integrates domain-specific knowledge into a general reward model by model merging. The experiments demonstrate that DogeRM enhances performance across different benchmarks and provide a detailed analysis showcasing the effects of model merging, showing the great potential of facilitating model alignment.

Ze-Song Xu, Yun-Nung Chen

Developing dialogue relation extraction (DRE) systems often requires a large amount of labeled data, which can be costly and time-consuming to annotate. In order to improve scalability and support diverse, unseen relation extraction, this paper proposes a method for leveraging the ability to capture triggers and relate them to previously unseen relation names. Specifically, we introduce a model that enables zero-shot dialogue relation extraction by utilizing trigger-capturing capabilities. Our experiments on a benchmark DialogRE dataset demonstrate that the proposed model achieves significant improvements for both seen and unseen relations. Notably, this is the first attempt at zero-shot dialogue relation extraction using trigger-capturing capabilities, and our results suggest that this approach is effective for inferring previously unseen relation types. Overall, our findings highlight the potential for this method to enhance the scalability and practicality of DRE systems.

Chao-Wei Huang, Chen-An Li, Tsu-Yuan Hsu et al.

Dense retrieval methods have demonstrated promising performance in multilingual information retrieval, where queries and documents can be in different languages. However, dense retrievers typically require a substantial amount of paired data, which poses even greater challenges in multilingual scenarios. This paper introduces UMR, an Unsupervised Multilingual dense Retriever trained without any paired data. Our approach leverages the sequence likelihood estimation capabilities of multilingual language models to acquire pseudo labels for training dense retrievers. We propose a two-stage framework which iteratively improves the performance of multilingual dense retrievers. Experimental results on two benchmark datasets show that UMR outperforms supervised baselines, showcasing the potential of training multilingual retrievers without paired data, thereby enhancing their practicality. Our source code, data, and models are publicly available at https://github.com/MiuLab/UMR

Chao-Wei Huang, Yun-Nung Chen

This paper introduces InstUPR, an unsupervised passage reranking method based on large language models (LLMs). Different from existing approaches that rely on extensive training with query-document pairs or retrieval-specific instructions, our method leverages the instruction-following capabilities of instruction-tuned LLMs for passage reranking without any additional fine-tuning. To achieve this, we introduce a soft score aggregation technique and employ pairwise reranking for unsupervised passage reranking. Experiments on the BEIR benchmark demonstrate that InstUPR outperforms unsupervised baselines as well as an instruction-tuned reranker, highlighting its effectiveness and superiority. Source code to reproduce all experiments is open-sourced at https://github.com/MiuLab/InstUPR

Yen-Shan Chen, Zhi Rui Tam, Cheng-Kuang Wu et al.

Current evaluations of LLM safety predominantly rely on severity-based taxonomies to assess the harmfulness of malicious queries. We argue that this formulation requires re-examination as it assumes uniform risk across all malicious queries, neglecting Execution Likelihood--the conditional probability of a threat being realized given the model's response. In this work, we introduce Expected Harm, a metric that weights the severity of a jailbreak by its execution likelihood, modeled as a function of execution cost. Through empirical analysis of state-of-the-art models, we reveal a systematic Inverse Risk Calibration: models disproportionately exhibit stronger refusal behaviors for low-likelihood (high-cost) threats while remaining vulnerable to high-likelihood (low-cost) queries. We demonstrate that this miscalibration creates a structural vulnerability: by exploiting this property, we increase the attack success rate of existing jailbreaks by up to $2\times$. Finally, we trace the root cause of this failure using linear probing, which reveals that while models encode severity in their latent space to drive refusal decisions, they possess no distinguishable internal representation of execution cost, making them "blind" to this critical dimension of risk.

Zhi Rui Tam, Yun-Nung Chen

As large language models transition from text-based interfaces to audio interactions in clinical settings, they might introduce new vulnerabilities through paralinguistic cues in audio. We evaluated these models on 170 clinical cases, each synthesized into speech from 36 distinct voice profiles spanning variations in age, gender, and emotion. Our findings reveal a severe modality bias: surgical recommendations for audio inputs varied by as much as 35% compared to identical text-based inputs, with one model providing 80% fewer recommendations. Further analysis uncovered age disparities of up to 12% between young and elderly voices, which persisted in most models despite chain-of-thought prompting. While explicit reasoning successfully eliminated gender bias, the impact of emotion was not detected due to poor recognition performance. These results demonstrate that audio LLMs are susceptible to making clinical decisions based on a patient's voice characteristics rather than medical evidence, a flaw that risks perpetuating healthcare disparities. We conclude that bias-aware architectures are essential and urgently needed before the clinical deployment of these models.

Tzu-Han Lin, Wei-Lin Chen, Chen-An Li et al.

Equipping large language models (LLMs) with search engines via reinforcement learning (RL) has emerged as an effective approach for building search agents. However, overreliance on search introduces unnecessary cost and risks exposure to noisy or malicious content, while relying solely on parametric knowledge risks hallucination. The central challenge is to develop agents that adaptively balance parametric knowledge with external search, invoking search only when necessary. Prior work mitigates search overuse by shaping rewards around the number of tool calls. However, these penalties require substantial reward engineering, provide ambiguous credit assignment, and can be exploited by agents that superficially reduce calls. Moreover, evaluating performance solely through call counts conflates necessary and unnecessary search, obscuring the measurement of true adaptive behavior. To address these limitations, we first quantify the self-knowledge awareness of existing search agents via an F1-based decision metric, revealing that methods such as Search-R1 often overlook readily available parametric knowledge. Motivated by these findings, we propose AdaSearch, a simple two-stage, outcome-driven RL framework that disentangles problem solving from the decision of whether to invoke search, and makes this decision process explicit and interpretable. This transparency is crucial for high-stakes domains such as finance and medical question answering, yet is largely neglected by prior approaches. Experiments across multiple model families and sizes demonstrate that AdaSearch substantially improves knowledge-boundary awareness, reduces unnecessary search calls, preserves strong task performance, and offers more transparent, interpretable decision behaviors.

Yen-Ting Piao, Jay Chiehen Liao, Wei-Tang Chien et al.

While Large Audio-Language Models (LALMs) have been shown to exhibit degraded instruction-following capabilities, their ability to infer task patterns from in-context examples under audio conditioning remains unstudied. To address this gap, we present ALICE, a three-stage framework that progressively reduces textual guidance to systematically evaluate LALMs' in-context learning ability under audio conditioning. Evaluating six LALMs across four audio understanding tasks under two output constraint categories, we uncover a consistent asymmetry across all stages and LALMs: in-context demonstrations reliably improve format compliance but fail to improve, and often degrade, the core task performance. This suggests that LALMs can glean surface-level formatting patterns from demonstrations but may struggle to leverage cross-modal semantic grounding to reliably infer task objectives from audio-conditioned examples, highlighting potential limitations in current cross-modal integration.

Yi-Jie Cheng, Oscar Chew, Yun-Nung Chen

Integrating knowledge graphs (KGs) into the reasoning processes of large language models (LLMs) has emerged as a promising approach to mitigate hallucination. However, existing work in this area often relies on proprietary or extremely large models, limiting accessibility and scalability. In this study, we investigate the capabilities of existing integration methods for small language models (SLMs) in KG-based question answering and observe that their performance is often constrained by their limited ability to traverse and reason over knowledge graphs. To address this limitation, we propose leveraging simple and efficient exploration modules to handle knowledge graph traversal in place of the language model itself. Experiment results demonstrate that these lightweight modules effectively improve the performance of small language models on knowledge graph question answering tasks. Source code: https://github.com/yijie-cheng/SLM-ToG/.

Yu-Ang Lee, Guan-Ting Yi, Mei-Yi Liu et al.

Recent advancements in large language models (LLMs) and AI systems have led to a paradigm shift in the design and optimization of complex AI workflows. By integrating multiple components, compound AI systems have become increasingly adept at performing sophisticated tasks. However, as these systems grow in complexity, new challenges arise in optimizing not only individual components but also their interactions. While traditional optimization methods such as supervised fine-tuning (SFT) and reinforcement learning (RL) remain foundational, the rise of natural language feedback introduces promising new approaches, especially for optimizing non-differentiable systems. This paper provides a systematic review of recent progress in optimizing compound AI systems, encompassing both numerical and language-based techniques. We formalize the notion of compound AI system optimization, classify existing methods along several key dimensions, and highlight open research challenges and future directions in this rapidly evolving field. A list of surveyed papers is publicly available at https://github.com/MiuLab/AISysOpt-Survey.

Cheng-Kuang Wu, Zhi Rui Tam, Chieh-Yen Lin et al.

Recent works have shown that large language model (LLM) agents are able to improve themselves from experience, which is an important ability for continuous enhancement post-deployment. However, existing benchmarks primarily evaluate their innate capabilities and do not assess their ability to improve over time. To address this gap, we introduce StreamBench, a pioneering benchmark designed to evaluate the continuous improvement of LLM agents over an input-feedback sequence. StreamBench simulates an online learning environment where LLMs receive a continuous flow of feedback stream and iteratively enhance their performance. In addition, we propose several simple yet effective baselines for improving LLMs on StreamBench, and provide a comprehensive analysis to identify critical components that contribute to successful streaming strategies. Our work serves as a stepping stone towards developing effective online learning strategies for LLMs, paving the way for more adaptive AI systems in streaming scenarios. Source code: https://github.com/stream-bench/stream-bench. Benchmark website: https://stream-bench.github.io.

Cheng-Hsun Hsueh, Paul Kuo-Ming Huang, Tzu-Han Lin et al.

Knowledge editing is a rising technique for efficiently updating factual knowledge in large language models (LLMs) with minimal alteration of parameters. However, recent studies have identified side effects, such as knowledge distortion and the deterioration of general abilities, that have emerged after editing. Despite these findings, evaluating the pitfalls of knowledge editing often relies on inconsistent metrics and benchmarks, lacking a uniform standard. In response, this survey presents a comprehensive study of these side effects, providing a unified perspective on the challenges of knowledge editing in LLMs by conducting experiments with consistent metrics and benchmarks. Additionally, we review related works and outline potential research directions to address these limitations. Our survey highlights the limitations of current knowledge editing methods, emphasizing the need for a deeper understanding of the inner knowledge structures of LLMs and improved knowledge editing methods. To foster future research, we have released the complementary materials publicly in https://github.com/MiuLab/EditLLM-Survey.

Tzu-Lin Kuo, Tzu-Wei Chiu, Tzung-Sheng Lin et al.

Generative Retrieval (GR) is an emerging paradigm in information retrieval that leverages generative models to directly map queries to relevant document identifiers (DocIDs) without the need for traditional query processing or document reranking. This survey provides a comprehensive overview of GR, highlighting key developments, indexing and retrieval strategies, and challenges. We discuss various document identifier strategies, including numerical and string-based identifiers, and explore different document representation methods. Our primary contribution lies in outlining future research directions that could profoundly impact the field: improving the quality of query generation, exploring learnable document identifiers, enhancing scalability, and integrating GR with multi-task learning frameworks. By examining state-of-the-art GR techniques and their applications, this survey aims to provide a foundational understanding of GR and inspire further innovations in this transformative approach to information retrieval. We also make the complementary materials such as paper collection publicly available at https://github.com/MiuLab/GenIR-Survey/

Yu-Min Tseng, Yu-Chao Huang, Teng-Yun Hsiao et al.

The concept of persona, originally adopted in dialogue literature, has re-surged as a promising framework for tailoring large language models (LLMs) to specific context (e.g., personalized search, LLM-as-a-judge). However, the growing research on leveraging persona in LLMs is relatively disorganized and lacks a systematic taxonomy. To close the gap, we present a comprehensive survey to categorize the current state of the field. We identify two lines of research, namely (1) LLM Role-Playing, where personas are assigned to LLMs, and (2) LLM Personalization, where LLMs take care of user personas. Additionally, we introduce existing methods for LLM personality evaluation. To the best of our knowledge, we present the first survey for role-playing and personalization in LLMs under the unified view of persona. We continuously maintain a paper collection to foster future endeavors: https://github.com/MiuLab/PersonaLLM-Survey

Shang-Chi Tsai, Chao-Wei Huang, Yun-Nung Chen

Given the clinical notes written in electronic health records (EHRs), it is challenging to predict the diagnostic codes which is formulated as a multi-label classification task. The large set of labels, the hierarchical dependency, and the imbalanced data make this prediction task extremely hard. Most existing work built a binary prediction for each label independently, ignoring the dependencies between labels. To address this problem, we propose a two-stage framework to improve automatic ICD coding by capturing the label correlation. Specifically, we train a label set distribution estimator to rescore the probability of each label set candidate generated by a base predictor. This paper is the first attempt at learning the label set distribution as a reranking module for medical code prediction. In the experiments, our proposed framework is able to improve upon best-performing predictors on the benchmark MIMIC datasets. The source code of this project is available at https://github.com/MiuLab/ICD-Correlation.

Chun-Hsing Lin, Siang-Ruei Wu, Hung-Yi Lee et al.

Score function-based natural language generation (NLG) approaches such as REINFORCE, in general, suffer from low sample efficiency and training instability problems. This is mainly due to the non-differentiable nature of the discrete space sampling and thus these methods have to treat the discriminator as a black box and ignore the gradient information. To improve the sample efficiency and reduce the variance of REINFORCE, we propose a novel approach, TaylorGAN, which augments the gradient estimation by off-policy update and the first-order Taylor expansion. This approach enables us to train NLG models from scratch with smaller batch size -- without maximum likelihood pre-training, and outperforms existing GAN-based methods on multiple metrics of quality and diversity. The source code and data are available at https://github.com/MiuLab/TaylorGAN

Chao-Wei Huang, Yun-Nung Chen

Lattices are compact representations that encode multiple hypotheses, such as speech recognition results or different word segmentations. It is shown that encoding lattices as opposed to 1-best results generated by automatic speech recognizer (ASR) boosts the performance of spoken language understanding (SLU). Recently, pretrained language models with the transformer architecture have achieved the state-of-the-art results on natural language understanding, but their ability of encoding lattices has not been explored. Therefore, this paper aims at adapting pretrained transformers to lattice inputs in order to perform understanding tasks specifically for spoken language. Our experiments on the benchmark ATIS dataset show that fine-tuning pretrained transformers with lattice inputs yields clear improvement over fine-tuning with 1-best results. Further evaluation demonstrates the effectiveness of our methods under different acoustic conditions. Our code is available at https://github.com/MiuLab/Lattice-SLU

Chao-Wei Huang, Yun-Nung Chen

Pre-trained language models have achieved huge improvement on many NLP tasks. However, these methods are usually designed for written text, so they do not consider the properties of spoken language. Therefore, this paper aims at generalizing the idea of language model pre-training to lattices generated by recognition systems. We propose a framework that trains neural lattice language models to provide contextualized representations for spoken language understanding tasks. The proposed two-stage pre-training approach reduces the demands of speech data and has better efficiency. Experiments on intent detection and dialogue act recognition datasets demonstrate that our proposed method consistently outperforms strong baselines when evaluated on spoken inputs. The code is available at https://github.com/MiuLab/Lattice-ELMo.

Chao-Wei Huang, Yun-Nung Chen

Employing pre-trained language models (LM) to extract contextualized word representations has achieved state-of-the-art performance on various NLP tasks. However, applying this technique to noisy transcripts generated by automatic speech recognizer (ASR) is concerned. Therefore, this paper focuses on making contextualized representations more ASR-robust. We propose a novel confusion-aware fine-tuning method to mitigate the impact of ASR errors to pre-trained LMs. Specifically, we fine-tune LMs to produce similar representations for acoustically confusable words that are obtained from word confusion networks (WCNs) produced by ASR. Experiments on the benchmark ATIS dataset show that the proposed method significantly improves the performance of spoken language understanding when performing on ASR transcripts. Our source code is available at https://github.com/MiuLab/SpokenVec

Ting-Rui Chiang, Hao-Tong Ye, Yun-Nung Chen

With a lot of work about context-free question answering systems, there is an emerging trend of conversational question answering models in the natural language processing field. Thanks to the recently collected datasets, including QuAC and CoQA, there has been more work on conversational question answering, and recent work has achieved competitive performance on both datasets. However, to best of our knowledge, two important questions for conversational comprehension research have not been well studied: 1) How well can the benchmark dataset reflect models' content understanding? 2) Do the models well utilize the conversation content when answering questions? To investigate these questions, we design different training settings, testing settings, as well as an attack to verify the models' capability of content understanding on QuAC and CoQA. The experimental results indicate some potential hazards in the benchmark datasets, QuAC and CoQA, for conversational comprehension research. Our analysis also sheds light on both what models may learn and how datasets may bias the models. With deep investigation of the task, it is believed that this work can benefit the future progress of conversation comprehension. The source code is available at https://github.com/MiuLab/CQA-Study.

Ta-Chung Chi, Ching-Yen Shih, Yun-Nung Chen

This paper introduces the first dataset for evaluating English-Chinese Bilingual Contextual Word Similarity, namely BCWS (https://github.com/MiuLab/BCWS). The dataset consists of 2,091 English-Chinese word pairs with the corresponding sentential contexts and their similarity scores annotated by the human. Our annotated dataset has higher consistency compared to other similar datasets. We establish several baselines for the bilingual embedding task to benchmark the experiments. Modeling cross-lingual sense representations as provided in this dataset has the potential of moving artificial intelligence from monolingual understanding towards multilingual understanding.

Bhuwan Dhingra, Lihong Li, Xiujun Li et al.

This paper proposes KB-InfoBot -- a multi-turn dialogue agent which helps users search Knowledge Bases (KBs) without composing complicated queries. Such goal-oriented dialogue agents typically need to interact with an external database to access real-world knowledge. Previous systems achieved this by issuing a symbolic query to the KB to retrieve entries based on their attributes. However, such symbolic operations break the differentiability of the system and prevent end-to-end training of neural dialogue agents. In this paper, we address this limitation by replacing symbolic queries with an induced "soft" posterior distribution over the KB that indicates which entities the user is interested in. Integrating the soft retrieval process with a reinforcement learner leads to higher task success rate and reward in both simulations and against real users. We also present a fully neural end-to-end agent, trained entirely from user feedback, and discuss its application towards personalized dialogue agents. The source code is available at https://github.com/MiuLab/KB-InfoBot.

Song-Duo Ma, Yi-Hung Liu, Hsin-Yu Lin et al.

To efficiently combat the spread of LLM-generated misinformation, we present RADAR, a retrieval-augmented detector with adversarial refinement for robust fake news detection. Our approach employs a generator that rewrites real articles with factual perturbations, paired with a lightweight detector that verifies claims using dense passage retrieval. To enable effective co-evolution, we introduce verbal adversarial feedback (VAF). Rather than relying on scalar rewards, VAF issues structured natural-language critiques; these guide the generator toward more sophisticated evasion attempts, compelling the detector to adapt and improve. On a fake news detection benchmark, RADAR achieves 86.98% ROC-AUC, significantly outperforming general-purpose LLMs with retrieval. Ablation studies confirm that detector-side retrieval yields the largest gains, while VAF and few-shot demonstrations provide critical signals for robust training.

Wen-Yu Chang, Yun-Nung Chen

Recent research in dialogue systems and corpora has focused on two main categories: task-oriented (TOD) and open-domain (chit-chat) dialogues. TOD systems help users accomplish specific tasks, while open-domain systems aim to create engaging conversations. However, in real-world scenarios, user intents are often revealed during interactions. A recent study introduced SalesBot, which simulates dialogues transitioning from chit-chat to task-oriented scenarios to train sales agents. Unfortunately, the initial data lacked smooth transitions and coherent long-turn dialogues, resulting in poor naturalness in sales-customer interactions. To address these issues, this paper presents SalesBot 2.0, an improved dataset. It leverages commonsense knowledge from large language models (LLMs) through strategic prompting. Additionally, we introduce a novel model called SalesAgent, trained on salesperson's interactions, using chain-of-thought (CoT) reasoning. This model excels in transitioning topics, understanding user intents, and selecting appropriate strategies. Experiments using diverse user simulations validate the effectiveness of our method in controlling dialogue strategies in LLMs. Furthermore, SalesBot 2.0 enhances coherence and reduces aggression, facilitating better model learning for sales-customer interactions.

Yen-Ting Lin, Di Jin, Tengyu Xu et al.

Large language models (LLMs) have recently demonstrated remarkable success in mathematical reasoning. Despite progress in methods like chain-of-thought prompting and self-consistency sampling, these advances often focus on final correctness without ensuring that the underlying reasoning process is coherent and reliable. This paper introduces Step-KTO, a training framework that combines process-level and outcome-level binary feedback to guide LLMs toward more trustworthy reasoning trajectories. By providing binary evaluations for both the intermediate reasoning steps and the final answer, Step-KTO encourages the model to adhere to logical progressions rather than relying on superficial shortcuts. Our experiments on challenging mathematical benchmarks show that Step-KTO significantly improves both final answer accuracy and the quality of intermediate reasoning steps. For example, on the MATH-500 dataset, Step-KTO achieves a notable improvement in Pass@1 accuracy over strong baselines. These results highlight the promise of integrating stepwise process feedback into LLM training, paving the way toward more interpretable and dependable reasoning capabilities.

Yen-Shan Chen, Sian-Yao Huang, Cheng-Lin Yang et al.

As large language models (LLMs) evolve from static chatbots into autonomous agents, the primary vulnerability surface shifts from final outputs to intermediate execution traces. While safety guardrails are well-benchmarked for natural language responses, their efficacy remains largely unexplored within multi-step tool-use trajectories. To address this gap, we introduce TraceSafe-Bench, the first comprehensive benchmark specifically designed to assess mid-trajectory safety. It encompasses 12 risk categories, ranging from security threats (e.g., prompt injection, privacy leaks) to operational failures (e.g., hallucinations, interface inconsistencies), featuring over 1,000 unique execution instances. Our evaluation of 13 LLM-as-a-guard models and 7 specialized guardrails yields three critical findings: 1) Structural Bottleneck: Guardrail efficacy is driven more by structural data competence (e.g., JSON parsing) than semantic safety alignment. Performance correlates strongly with structured-to-text benchmarks ($ρ=0.79$) but shows near-zero correlation with standard jailbreak robustness. 2) Architecture over Scale: Model architecture influences risk detection performance more significantly than model size, with general-purpose LLMs consistently outperforming specialized safety guardrails in trajectory analysis. 3) Temporal Stability: Accuracy remains resilient across extended trajectories. Increased execution steps allow models to pivot from static tool definitions to dynamic execution behaviors, actually improving risk detection performance in later stages. Our findings suggest that securing agentic workflows requires jointly optimizing for structural reasoning and safety alignment to effectively mitigate mid-trajectory risks.

Chang-Sheng Kao, Yun-Nung Chen

Recent advancements in dialogue systems have highlighted the significance of integrating multimodal responses, which enable conveying ideas through diverse modalities rather than solely relying on text-based interactions. This enrichment not only improves overall communicative efficacy but also enhances the quality of conversational experiences. However, existing methods for dialogue-to-image retrieval face limitations due to the constraints of pre-trained vision language models (VLMs) in comprehending complex dialogues accurately. To address this, we present a novel approach leveraging the robust reasoning capabilities of large language models (LLMs) to generate precise dialogue-associated visual descriptors, facilitating seamless connection with images. Extensive experiments conducted on benchmark data validate the effectiveness of our proposed approach in deriving concise and accurate visual descriptors, leading to significant enhancements in dialogue-to-image retrieval performance. Furthermore, our findings demonstrate the method's generalizability across diverse visual cues, various LLMs, and different datasets, underscoring its practicality and potential impact in real-world applications.

Ho-Lam Chung, Teng-Yun Hsiao, Hsiao-Ying Huang et al.

Test-Time Scaling (TTS) improves the reasoning performance of Large Language Models (LLMs) by allocating additional compute during inference. We conduct a structured survey of TTS methods and categorize them into sampling-based, search-based, and trajectory optimization strategies. We observe that reasoning-optimized models often produce less diverse outputs, which limits TTS effectiveness. To address this, we propose ADAPT (A Diversity Aware Prefix fine-Tuning), a lightweight method that applies prefix tuning with a diversity-focused data strategy. Experiments on mathematical reasoning tasks show that ADAPT reaches 80% accuracy using eight times less compute than strong baselines. Our findings highlight the essential role of generative diversity in maximizing TTS effectiveness.

Yi-Cheng Lin, Kang-Chieh Chen, Zhe-Yan Li et al.

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 \emph{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.

Zhi Rui Tam, Cheng-Kuang Wu, Yu Ying Chiu et al.

Large reasoning models (LRMs) have demonstrated impressive performance across a range of reasoning tasks, yet little is known about their internal reasoning processes in multilingual settings. We begin with a critical question: {\it In which language do these models reason when solving problems presented in different languages?} Our findings reveal that, despite multilingual training, LRMs tend to default to reasoning in high-resource languages (e.g., English) at test time, regardless of the input language. When constrained to reason in the same language as the input, model performance declines, especially for low-resource languages. In contrast, reasoning in high-resource languages generally preserves performance. We conduct extensive evaluations across reasoning-intensive tasks (MMMLU, MATH-500) and non-reasoning benchmarks (CulturalBench, LMSYS-toxic), showing that the effect of language choice varies by task type: input-language reasoning degrades performance on reasoning tasks but benefits cultural tasks, while safety evaluations exhibit language-specific behavior. By exposing these linguistic biases in LRMs, our work highlights a critical step toward developing more equitable models that serve users across diverse linguistic backgrounds.