Zhengyuan Liu

Jiangwei Chen, Xinyuan Niu, Rachael Hwee Ling Sim et al.

Machine unlearning aims to remove the influence of specific forget training data due to privacy, copyright or bias concerns while maintaining the model performance on the remaining retain data. Existing unlearning algorithms, such as optimizing a weighted combination of losses, have tried to achieve these objectives of improving forget quality and maintaining retain utility. However, they do not guarantee that these objectives can be improved by a specified extent for all forget and retain data. In this work, we address this limitation with a novel and theoretically-grounded approach from a constrained optimization perspective. Firstly, we identify that the hardness of reconciling both objectives can be quantified by the similarity between the forget data and the retain data. Next, we derive an unlearning algorithm (HAMU) with the overall goal of guaranteeing a specified improvement in forget quality while minimizing the retain utility cost/degradation by updating the model weights based on our hardness measure. Our hardness measure also informs users when retain utility degradation is unavoidable, i.e., both objectives cannot be improved simultaneously, and stopping should be considered. Our algorithm is applicable to non-convex models and is easily parallelizable, making it readily deployable in real-world scenarios. We empirically demonstrate HAMU's superior performance over baselines on both image and text datasets using large models. Our code is available at https://github.com/aoi3142/HAMU.

Bin Wang, Zhengyuan Liu, Xin Huang et al.

We present SeaEval, a benchmark for multilingual foundation models. In addition to characterizing how these models understand and reason with natural language, we also investigate how well they comprehend cultural practices, nuances, and values. Alongside standard accuracy metrics, we investigate the brittleness of foundation models in the dimensions of semantics and multilinguality. Our analyses span both open-sourced and closed models, leading to empirical results across classic NLP tasks, reasoning, and cultural comprehension. Key findings indicate (1) Most models exhibit varied behavior when given paraphrased instructions. (2) Many models still suffer from exposure bias (e.g., positional bias, majority label bias). (3) For questions rooted in factual, scientific, and commonsense knowledge, consistent responses are expected across multilingual queries that are semantically equivalent. Yet, most models surprisingly demonstrate inconsistent performance on these queries. (4) Multilingually-trained models have not attained "balanced multilingual" capabilities. Our endeavors underscore the need for more generalizable semantic representations and enhanced multilingual contextualization. SeaEval can serve as a launchpad for more thorough investigations and evaluations for multilingual and multicultural scenarios.

Minzhi Li, Taiwei Shi, Caleb Ziems et al.

Annotated data plays a critical role in Natural Language Processing (NLP) in training models and evaluating their performance. Given recent developments in Large Language Models (LLMs), models such as ChatGPT demonstrate zero-shot capability on many text-annotation tasks, comparable with or even exceeding human annotators. Such LLMs can serve as alternatives for manual annotation, due to lower costs and higher scalability. However, limited work has leveraged LLMs as complementary annotators, nor explored how annotation work is best allocated among humans and LLMs to achieve both quality and cost objectives. We propose CoAnnotating, a novel paradigm for Human-LLM co-annotation of unstructured texts at scale. Under this framework, we utilize uncertainty to estimate LLMs' annotation capability. Our empirical study shows CoAnnotating to be an effective means to allocate work from results on different datasets, with up to 21% performance improvement over random baseline. For code implementation, see https://github.com/SALT-NLP/CoAnnotating.

Galann Pennec, Zhengyuan Liu, Nicholas Asher et al.

We introduce SagaQA, a long-form video benchmark for multi-hop reasoning over full-length TV series. Existing video reasoning benchmarks often emphasize local understanding of adjacent frames or clips. SagaQA addresses this gap by requiring high-level comprehension of extended multimodal narratives in entire TV shows. A distinguishing feature of SagaQA is the granularity of its reasoning steps. Our dataset necessitates long-range reasoning hops to connect information across completely different episodes. This requires models to reason over entire events and actions, demanding a deep understanding of the show's narration and progression at a multimodal level. Motivated by recent progress in agentic methods, we further study how different planning strategies handle such complex reasoning. We categorize these approaches into three classes-Parallel, Sequential, and Hybrid planners-and evaluate their ability to generate coherent and complete reasoning plans. Our results on SagaQA suggest that hybrid planners consistently produce higher-quality plans and exhibit stronger capabilities for complex, high-level narrative understanding in TV shows.

Ayrton San Joaquin, Bin Wang, Zhengyuan Liu et al.

Despite advancements, fine-tuning Large Language Models (LLMs) remains costly due to the extensive parameter count and substantial data requirements for model generalization. Accessibility to computing resources remains a barrier for the open-source community. To address this challenge, we propose the In2Core algorithm, which selects a coreset by analyzing the correlation between training and evaluation samples with a trained model. Notably, we assess the model's internal gradients to estimate this relationship, aiming to rank the contribution of each training point. To enhance efficiency, we propose an optimization to compute influence functions with a reduced number of layers while achieving similar accuracy. By applying our algorithm to instruction fine-tuning data of LLMs, we can achieve similar performance with just 50% of the training data. Meantime, using influence functions to analyze model coverage to certain testing samples could provide a reliable and interpretable signal on the training set's coverage of those test points.

Derrick Goh Xin Deik, Quanyu Long, Zhengyuan Liu et al.

Multi-constraint planning involves identifying, evaluating, and refining candidate plans while satisfying multiple, potentially conflicting constraints. Existing large language model (LLM) approaches face fundamental limitations in this domain. Pure reasoning paradigms, which rely on long natural language chains, are prone to inconsistency, error accumulation, and prohibitive cost as constraints compound. Conversely, LLMs combined with coding- or solver-based strategies lack flexibility: they often generate problem-specific code from scratch or depend on fixed solvers, failing to capture generalizable logic across diverse problems. To address these challenges, we introduce the Scalable COde Planning Engine (SCOPE), a framework that disentangles query-specific reasoning from generic code execution. By separating reasoning from execution, SCOPE produces solver functions that are consistent, deterministic, and reusable across queries while requiring only minimal changes to input parameters. SCOPE achieves state-of-the-art performance while lowering cost and latency. For example, with GPT-4o, it reaches 93.1% success on TravelPlanner, a 61.6% gain over the best baseline (CoT) while cutting inference cost by 1.4x and time by ~4.67x. Code is available at https://github.com/DerrickGXD/SCOPE.

Bryan Chen Zhengyu Tan, Shaun Khoo, Bich Ngoc Doan et al.

Large Language Models (LLMs) are increasingly deployed in resume screening pipelines. Although explicit PII (e.g., names) is commonly redacted, resumes typically retain subtle sociocultural markers (languages, co-curricular activities, volunteering, hobbies) that can act as demographic proxies. We introduce a generalisable stress-test framework for hiring fairness instantiated in the Singapore context: 100 neutral job-aligned resumes are augmented into 4100 variants spanning four ethnicities and two genders, differing only in job-irrelevant markers. We evaluate 18 LLMs in two settings: (i) Direct Comparison (1v1) and (ii) Score & Shortlist (Top-Score Rates), each with and without rationale prompting. We find that even without explicit identifiers, models recover demographic attributes with high F1 and exhibit systematic disparities, with models favouring markers associated with Chinese and Caucasian males. Ablations show language markers suffice for inferring ethnicity, while hobbies and activities are utilised for gender. Furthermore, prompting for explanations may paradoxically amplify bias. Our findings suggest that seemingly innocuous markers surviving anonymisation can materially skew automated hiring outcomes.

Zhengyuan Liu, Stella Xin Yin, Bryan Chen Zhengyu Tan et al.

User simulation is important for developing and evaluating human-centered AI, yet current student simulation in educational applications has significant limitations. Existing approaches focus on single learning experiences and do not account for students' gradual knowledge construction and evolving skill sets. Moreover, large language models are optimized to produce direct and accurate responses, making it challenging to represent the incomplete understanding and developmental constraints that characterize real learners. In this paper, we introduce a novel framework for memory-based student simulation that incorporates developmental trajectories through a hierarchical memory mechanism with structured knowledge representation. The framework also integrates metacognitive processes and personality traits to enrich the individual learner profiling, through dynamical consolidation of both cognitive development and personal learning characteristics. In practice, we implement a curriculum-aligned simulator grounded on the Next Generation Science Standards. Experimental results show that our approach can effectively reflect the gradual nature of knowledge development and the characteristic difficulties students face, providing a more accurate representation of learning processes.

Bin Wang, Zhengyuan Liu, Nancy F. Chen

Conventional dialogue summarization methods directly generate summaries and do not consider user's specific interests. This poses challenges in cases where the users are more focused on particular topics or aspects. With the advancement of instruction-finetuned language models, we introduce instruction-tuning to dialogues to expand the capability set of dialogue summarization models. To overcome the scarcity of instructive dialogue summarization data, we propose a three-step approach to synthesize high-quality query-based summarization triples. This process involves summary-anchored query generation, query filtering, and query-based summary generation. By training a unified model called InstructDS (Instructive Dialogue Summarization) on three summarization datasets with multi-purpose instructive triples, we expand the capability of dialogue summarization models. We evaluate our method on four datasets, including dialogue summarization and dialogue reading comprehension. Experimental results show that our approach outperforms the state-of-the-art models and even models with larger sizes. Additionally, our model exhibits higher generalizability and faithfulness, as confirmed by human subjective evaluations.

Zifu Zhang, Shengxi Li, Xiancheng Sun et al.

Human-machine collaborative compression has been receiving increasing research efforts for reducing image/video data, serving as the basis for both human perception and machine intelligence. Existing collaborative methods are dominantly built upon the de facto human-vision compression pipeline, witnessing deficiency on complexity and bit-rates when aggregating the machine-vision compression. Indeed, machine vision solely focuses on the core regions within the image/video, requiring much less information compared with the compressed information for human vision. In this paper, we thus set out the first successful attempt by a novel collaborative compression method based on the machine-vision-oriented compression, instead of human-vision pipeline. In other words, machine vision serves as the basis for human vision within collaborative compression. A plug-and-play variable bit-rate strategy is also developed for machine vision tasks. Then, we propose to progressively aggregate the semantics from the machine-vision compression, whilst seamlessly tailing the diffusion prior to restore high-fidelity details for human vision, thus named as diffusion-prior based feature compression for human and machine visions (Diff-FCHM). Experimental results verify the consistently superior performances of our Diff-FCHM, on both machine-vision and human-vision compression with remarkable margins. Our code will be released upon acceptance.

Zhengyuan Liu, Hai Leong Chieu, Nancy F. Chen

Data collection from manual labeling provides domain-specific and task-aligned supervision for data-driven approaches, and a critical mass of well-annotated resources is required to achieve reasonable performance in natural language processing tasks. However, manual annotations are often challenging to scale up in terms of time and budget, especially when domain knowledge, capturing subtle semantic features, and reasoning steps are needed. In this paper, we investigate the efficacy of leveraging large language models on automated labeling for computational stance detection. We empirically observe that while large language models show strong potential as an alternative to human annotators, their sensitivity to task-specific instructions and their intrinsic biases pose intriguing yet unique challenges in machine annotation. We introduce a multi-label and multi-target sampling strategy to optimize the annotation quality. Experimental results on the benchmark stance detection corpora show that our method can significantly improve performance and learning efficacy.

Zhengyuan Liu, Nancy F. Chen

Text style transfer is an important task in controllable language generation. Supervised approaches have pushed performance improvement on style-oriented rewriting such as formality conversion. However, challenges remain due to the scarcity of large-scale parallel data in many domains. While unsupervised approaches do not rely on annotated sentence pairs for each style, they are often plagued with instability issues such as mode collapse or quality degradation. To take advantage of both supervised and unsupervised paradigms and tackle the challenges, in this work, we propose a semi-supervised framework for text style transfer. First, the learning process is bootstrapped with supervision guided by automatically constructed pseudo-parallel pairs using lexical and semantic-based methods. Then the model learns from unlabeled data via reinforcement rewards. Specifically, we propose to improve the sequence-to-sequence policy gradient via stepwise reward optimization, providing fine-grained learning signals and stabilizing the reinforced learning process. Experimental results show that the proposed approach achieves state-of-the-art performance on multiple datasets, and produces effective generation with as minimal as 10\% of training data.

Bin Wang, Chengwei Wei, Zhengyuan Liu et al.

As the rapidly advancing domain of natural language processing (NLP), large language models (LLMs) have emerged as powerful tools for interpreting human commands and generating text across various tasks. Nonetheless, the resilience of LLMs to handle text containing inherent errors, stemming from human interactions and collaborative systems, has not been thoroughly explored. Our study investigates the resilience of LLMs against five common types of disruptions including 1) ASR (Automatic Speech Recognition) errors, 2) OCR (Optical Character Recognition) errors, 3) grammatical mistakes, 4) typographical errors, and 5) distractive content. We aim to investigate how these models react by deliberately embedding these errors into instructions. Our findings reveal that while some LLMs show a degree of resistance to certain types of noise, their overall performance significantly suffers. This emphasizes the importance of further investigation into enhancing model resilience. In response to the observed decline in performance, our study also evaluates a "re-pass" strategy, designed to purify the instructions of noise before the LLMs process them. Our analysis indicates that correcting noisy instructions, particularly for open-source LLMs, presents significant challenges.

Zhengyuan Liu, Pavitra Krishnaswamy, Nancy F. Chen

There is growing interest in the automated extraction of relevant information from clinical dialogues. However, it is difficult to collect and construct large annotated resources for clinical dialogue tasks. Recent developments in natural language processing suggest that large-scale pre-trained language backbones could be leveraged for such machine comprehension and information extraction tasks. Yet, due to the gap between pre-training and downstream clinical domains, it remains challenging to exploit the generic backbones for domain-specific applications. Therefore, in this work, we propose a domain-specific language pre-training, to improve performance on downstream tasks like dialogue comprehension. Aside from the common token-level masking pre-training method, according to the nature of human conversations and interactive flow of multi-topic inquiry-answering dialogues, we further propose sample generation strategies with speaker and utterance manipulation. The conversational pre-training guides the language backbone to reconstruct the utterances coherently based on the remaining context, thus bridging the gap between general and specific domains. Experiments are conducted on a clinical conversation dataset for symptom checking, where nurses inquire and discuss symptom information with patients. We empirically show that the neural model with our proposed approach brings improvement in the dialogue comprehension task, and can achieve favorable results in the low resource training scenario.

Zhiliang Chen, Alfred Wei Lun Leong, Shao Yong Ong et al.

Co-optimizing data and model configurations for training LLMs presents a classic chicken-and-egg dilemma: The best training data configuration (e.g., data mixture) for a downstream task depends on the chosen model configuration (e.g., model architecture), and vice versa. However, jointly optimizing both data and model configurations is often deemed intractable, and existing methods focus on either data or model optimization without considering their interaction. We introduce JoBS, an approach that uses a scaling-law-inspired performance predictor to aid Bayesian optimization (BO) in jointly optimizing LLM training data and model configurations efficiently. JoBS allocates a portion of the optimization budget to learn an LLM performance predictor that predicts how promising a training configuration is from a small number of training steps. The remaining budget is used to perform BO entirely with the predictor, effectively amortizing the cost of running full-training runs. We study JoBS's average regret and devise the optimal budget allocation to minimize regret. JoBS outperforms existing multi-fidelity BO baselines, as well as data and model optimization approaches across diverse LLM tasks under the same optimization budget.

Bryan Chen Zhengyu Tan, Zhengyuan Liu, Xiaoyuan Yi et al.

Despite their global prevalence, many Large Language Models (LLMs) are aligned to a monolithic, often Western-centric set of values. This paper investigates the more challenging task of fine-grained value alignment: examining whether LLMs can emulate the distinct cultural values of demographic subgroups. Using Singapore as a case study and the World Values Survey (WVS), we examine the value landscape and show that even state-of-the-art models like GPT-4.1 achieve only 57.4% accuracy in predicting subgroup modal preferences. We construct a dataset of over 20,000 samples to train and evaluate a range of models. We demonstrate that simple fine-tuning on structured numerical preferences yields substantial gains, improving accuracy on unseen, out-of-distribution subgroups by an average of 17.4%. These gains partially transfer to open-ended generation. However, we find significant pre-existing performance biases, where models better emulate young, male, Chinese, and Christian personas. Furthermore, while fine-tuning improves average performance, it widens the disparity between subgroups when measured by distance-aware metrics. Our work offers insights into the limits and fairness implications of subgroup-level cultural alignment.

Galann Pennec, Zhengyuan Liu, Nicholas Asher et al.

Vision-Language Models (VLMs) are able to process increasingly longer videos. Yet, important visual information is easily lost throughout the entire context and missed by VLMs. Also, it is important to design tools that enable cost-effective analysis of lengthy video content. In this paper, we propose a clip selection method that targets key video moments to be included in a multimodal summary. We divide the video into short clips and generate compact visual descriptions of each using a lightweight video captioning model. These are then passed to a large language model (LLM), which selects the K clips containing the most relevant visual information for a multimodal summary. We evaluate our approach on reference clips for the task, automatically derived from full human-annotated screenplays and summaries in the MovieSum dataset. We further show that these reference clips (less than 6% of the movie) are sufficient to build a complete multimodal summary of the movies in MovieSum. Using our clip selection method, we achieve a summarization performance close to that of these reference clips while capturing substantially more relevant video information than random clip selection. Importantly, we maintain low computational cost by relying on a lightweight captioning model.

Bryan Chen Zhengyu Tan, Daniel Wai Kit Chin, Zhengyuan Liu et al.

Large Language Models (LLMs) can struggle to balance gullibility to misinformation and resistance to valid corrections in persuasive dialogues, a critical challenge for reliable deployment. We introduce DuET-PD (Dual Evaluation for Trust in Persuasive Dialogues), a framework evaluating multi-turn stance-change dynamics across dual dimensions: persuasion type (corrective/misleading) and domain (knowledge via MMLU-Pro, and safety via SALAD-Bench). We find that even a state-of-the-art model like GPT-4o achieves only 27.32% accuracy in MMLU-Pro under sustained misleading persuasions. Moreover, results reveal a concerning trend of increasing sycophancy in newer open-source models. To address this, we introduce Holistic DPO, a training approach balancing positive and negative persuasion examples. Unlike prompting or resist-only training, Holistic DPO enhances both robustness to misinformation and receptiveness to corrections, improving Llama-3.1-8B-Instruct's accuracy under misleading persuasion in safety contexts from 4.21% to 76.54%. These contributions offer a pathway to developing more reliable and adaptable LLMs for multi-turn dialogue. Code is available at https://github.com/Social-AI-Studio/DuET-PD.

Xin Huang, Tarun Kumar Vangani, Minh Duc Pham et al.

Multilingual large language models (MLLMs) have shown impressive capabilities across a variety of languages. However, efficacy can differ greatly between different language families, especially for those with limited linguistic resources. This report presents MERaLiON-TextLLM, a series of open-source language models specifically tailored to improve understanding and generation in Chinese, Indonesian, Malay, and Singlish. The initial released model is built on Llama-3-8B-Base and refined through a meticulously crafted process of continued pre-training and weight merging. Our approach achieves performance improvements across benchmarks in these languages, exceeding the capabilities of the official Llama-3 models. We provide the model checkpoints as a resource to support further research and development in cross-lingual language understanding.

Weihua Zheng, Chang Liu, Zhengyuan Liu et al.

Multilingual Large Language Models (LLMs) struggle with cross-lingual tasks due to data imbalances between high-resource and low-resource languages, as well as monolingual bias in pre-training. Existing methods, such as bilingual fine-tuning and contrastive alignment, can improve cross-lingual performance, but they often require extensive parallel data or suffer from instability. To address these challenges, we introduce a Cross-Lingual Mapping Task during the pre-training phase, which enhances cross-lingual alignment without compromising monolingual fluency. Our approach bi-directionally maps languages within the LLM embedding space, improving both language generation and comprehension. We further propose a Language Alignment Coefficient to robustly quantify cross-lingual consistency, even in limited-data scenarios. Experimental results on machine translation (MT), cross-lingual natural language understanding (CLNLU), and cross-lingual question answering (CLQA) show that our model achieves gains of up to 11.9 BLEU points in MT, 6.72 points in CLQA BERTScore-Precision, and more than 5% in CLNLU accuracy over strong multilingual baselines. These findings highlight the potential of incorporating cross-lingual objectives into pre-training to improve multilingual LLMs.

Bin Wang, Xunlong Zou, Geyu Lin et al.

We introduce AudioBench, a universal benchmark designed to evaluate Audio Large Language Models (AudioLLMs). It encompasses 8 distinct tasks and 26 datasets, among which, 7 are newly proposed datasets. The evaluation targets three main aspects: speech understanding, audio scene understanding, and voice understanding (paralinguistic). Despite recent advancements, there lacks a comprehensive benchmark for AudioLLMs on instruction following capabilities conditioned on audio signals. AudioBench addresses this gap by setting up datasets as well as desired evaluation metrics. Besides, we also evaluated the capabilities of five popular models and found that no single model excels consistently across all tasks. We outline the research outlook for AudioLLMs and anticipate that our open-sourced evaluation toolkit, data, and leaderboard will offer a robust testbed for future model developments.

Fangkai Jiao, Chengwei Qin, Zhengyuan Liu et al.

Large Language Models (LLMs) have demonstrated significant potential in handling complex reasoning tasks through step-by-step rationale generation. However, recent studies have raised concerns regarding the hallucination and flaws in their reasoning process. Substantial efforts are being made to improve the reliability and faithfulness of the generated rationales. Some approaches model reasoning as planning, while others focus on annotating for process supervision. Nevertheless, the planning-based search process often results in high latency due to the frequent assessment of intermediate reasoning states and the extensive exploration space. Additionally, supervising the reasoning process with human annotation is costly and challenging to scale for LLM training. To address these issues, in this paper, we propose a framework to learn planning-based reasoning through Direct Preference Optimization (DPO) on collected trajectories, which are ranked according to synthesized process rewards. Our results on challenging logical reasoning benchmarks demonstrate the effectiveness of our learning framework, showing that our 7B model can surpass the strong counterparts like GPT-3.5-Turbo.

Nedjma Ousidhoum, Junho Myung, Carla Perez-Almendros et al.

We present our shared task on evaluating the adaptability of LLMs and NLP systems across multiple languages and cultures. The task data consist of an extended version of our manually constructed BLEnD benchmark (Myung et al. 2024), covering more than 30 language-culture pairs, predominantly representing low-resource languages spoken across multiple continents. As the task is designed strictly for evaluation, participants were not permitted to use the data for training, fine-tuning, few-shot learning, or any other form of model modification. Our task includes two tracks: (a) Short-Answer Questions (SAQ) and (b) Multiple-Choice Questions (MCQ). Participants were required to predict labels and were allowed to submit any NLP system and adopt diverse modelling strategies, provided that the benchmark was used solely for evaluation. The task attracted more than 140 registered participants, and we received final submissions from 62 teams, along with 19 system description papers. We report the results and present an analysis of the best-performing systems and the most commonly adopted approaches. Furthermore, we discuss shared insights into open questions and challenges related to evaluation, misalignment, and methodological perspectives on model behaviour in low-resource languages and for under-represented cultures.

Zhengyuan Liu, Stella Xin Yin, Carolyn Lee et al.

Intelligent tutoring systems (ITSs) that imitate human tutors and aim to provide immediate and customized instructions or feedback to learners have shown their effectiveness in education. With the emergence of generative artificial intelligence, large language models (LLMs) further entitle the systems to complex and coherent conversational interactions. These systems would be of great help in language education as it involves developing skills in communication, which, however, drew relatively less attention. Additionally, due to the complicated cognitive development at younger ages, more endeavors are needed for practical uses. Scaffolding refers to a teaching technique where teachers provide support and guidance to students for learning and developing new concepts or skills. It is an effective way to support diverse learning needs, goals, processes, and outcomes. In this work, we investigate how pedagogical instructions facilitate the scaffolding in ITSs, by conducting a case study on guiding children to describe images for language learning. We construct different types of scaffolding tutoring systems grounded in four fundamental learning theories: knowledge construction, inquiry-based learning, dialogic teaching, and zone of proximal development. For qualitative and quantitative analyses, we build and refine a seven-dimension rubric to evaluate the scaffolding process. In our experiment on GPT-4V, we observe that LLMs demonstrate strong potential to follow pedagogical instructions and achieve self-paced learning in different student groups. Moreover, we extend our evaluation framework from a manual to an automated approach, paving the way to benchmark various conversational tutoring systems.

Geyu Lin, Bin Wang, Zhengyuan Liu et al.

Multilingual proficiency presents a significant challenge for large language models (LLMs). English-centric models are usually suboptimal in other languages, particularly those that are linguistically distant from English. This performance discrepancy mainly stems from the imbalanced distribution of training data across languages during pre-training and instruction tuning stages. To address this problem, we propose a novel approach called CrossIn, which utilizes a mixed composition of cross-lingual instruction tuning data. Our method leverages the compressed representation shared by various languages to efficiently enhance the model's task-solving capabilities and multilingual proficiency within a single process. In addition, we introduce a multi-task and multi-faceted benchmark to evaluate the effectiveness of CrossIn. Experimental results demonstrate that our method substantially improves performance across tasks and languages, and we provide extensive insights into the impact of cross-lingual data volume and the integration of translation data on enhancing multilingual consistency and accuracy.

Zhengyuan Liu, Stella Xin Yin, Geyu Lin et al.

Intelligent Tutoring Systems (ITSs) can provide personalized and self-paced learning experience. The emergence of large language models (LLMs) further enables better human-machine interaction, and facilitates the development of conversational ITSs in various disciplines such as math and language learning. In dialogic teaching, recognizing and adapting to individual characteristics can significantly enhance student engagement and learning efficiency. However, characterizing and simulating student's persona remain challenging in training and evaluating conversational ITSs. In this work, we propose a framework to construct profiles of different student groups by refining and integrating both cognitive and noncognitive aspects, and leverage LLMs for personality-aware student simulation in a language learning scenario. We further enhance the framework with multi-aspect validation, and conduct extensive analysis from both teacher and student perspectives. Our experimental results show that state-of-the-art LLMs can produce diverse student responses according to the given language ability and personality traits, and trigger teacher's adaptive scaffolding strategies.

Minzhi Li, Zhengyuan Liu, Shumin Deng et al.

The acceleration of Large Language Models (LLMs) research has opened up new possibilities for evaluating generated texts. They serve as scalable and economical evaluators, but the question of how reliable these evaluators are has emerged as a crucial research question. Prior research efforts in the meta-evaluation of LLMs as judges limit the prompting of an LLM to a single use to obtain a final evaluation decision. They then compute the agreement between LLMs' outputs and human labels. This lacks interpretability in understanding the evaluation capability of LLMs. In light of this challenge, we propose Decompose and Aggregate, which breaks down the evaluation process into different stages based on pedagogical practices. Our experiments illustrate that it not only provides a more interpretable window for how well LLMs evaluate, but also leads to improvements up to 39.6% for different LLMs on a variety of meta-evaluation benchmarks.

Zhengyuan Liu, Geyu Lin, Hui Li Tan et al.

The integration of generative artificial intelligence into educational applications has enhanced personalized and interactive learning experiences, and it shows strong potential to promote young learners language acquisition. However, it is still challenging to ensure consistent and robust performance across different languages and cultural contexts, and kids-friendly design requires simplified instructions, engaging interactions, and age-appropriate scaffolding to maintain motivation and optimize learning outcomes. In this work, we introduce SingaKids, a dialogic tutor designed to facilitate language learning through picture description tasks. Our system integrates dense image captioning, multilingual dialogic interaction, speech understanding, and engaging speech generation to create an immersive learning environment in four languages: English, Mandarin, Malay, and Tamil. We further improve the system through multilingual pre-training, task-specific tuning, and scaffolding optimization. Empirical studies with elementary school students demonstrate that SingaKids provides effective dialogic teaching, benefiting learners at different performance levels.

Weihua Zheng, Roy Ka-Wei Lee, Zhengyuan Liu et al.

Multilingual Large Language Models(MLLMs) demonstrate strong generalization across languages, yet they remain prone to hallucinations, especially in low-resource languages, due to training data imbalances. These hallucinations, which include inaccurate or fabricated outputs, are particularly problematic in domain-specific generation tasks (Chataigner et al., 2024). To address this challenge, we propose CCL-XCoT(Curriculum-based Contrastive Learning-based Cross-lingual Chain-of-Thought), a two-stage fine-tuning framework for mitigating hallucination in MLLMs. Our approach first enhances cross-lingual semantic alignment through curriculum-based contrastive learning combined with next-token prediction during continued pre-training. Building on this foundation, we then introduce a cross-lingual Chain-of-Thought (XCoT) prompting strategy during instruction fine-tuning, which guides the model to reason in a high-resource language before generating answers in the target low-resource language. Experimental results show that CCL-XCoT reduces hallucination rates by up to 62% and substantially improves factual knowledge transfer across language pairs, without relying on external retrieval or multi-model ensembles.

Zhengyuan Liu, Stella Xin Yin, Dion Hoe-Lian Goh et al.

While Generative AI has demonstrated strong potential and versatility in content generation, its application to educational contexts presents several challenges. Models often fail to align with curriculum standards and maintain grade-appropriate reading levels consistently. Furthermore, STEM education poses additional challenges in balancing scientific explanations with everyday language when introducing complex and abstract ideas and phenomena to younger students. In this work, we propose COGENT, a curriculum-oriented framework for generating grade-appropriate educational content. We incorporate three curriculum components (science concepts, core ideas, and learning objectives), control readability through length, vocabulary, and sentence complexity, and adopt a ``wonder-based'' approach to increase student engagement and interest. We conduct a multi-dimensional evaluation via both LLM-as-a-judge and human expert analysis. Experimental results show that COGENT consistently produces grade-appropriate passages that are comparable or superior to human references. Our work establishes a viable approach for scaling adaptive and high-quality learning resources.

Quanyu Long, Jianda Chen, Zhengyuan Liu et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities across numerous tasks, yet they often rely on external context to handle complex tasks. While retrieval-augmented frameworks traditionally focus on selecting top-ranked documents in a single pass, many real-world scenarios demand compositional retrieval, where multiple sources must be combined in a coordinated manner. In this work, we propose a tri-encoder sequential retriever that models this process as a Markov Decision Process (MDP), decomposing the probability of retrieving a set of elements into a sequence of conditional probabilities and allowing each retrieval step to be conditioned on previously selected examples. We train the retriever in two stages: first, we efficiently construct supervised sequential data for initial policy training; we then refine the policy to align with the LLM's preferences using a reward grounded in the structural correspondence of generated programs. Experimental results show that our method consistently and significantly outperforms baselines, underscoring the importance of explicitly modeling inter-example dependencies. These findings highlight the potential of compositional retrieval for tasks requiring multiple pieces of evidence or examples.

Weihua Zheng, Xin Huang, Zhengyuan Liu et al.

Large language models (LLMs) have shown impressive multilingual capabilities through pretraining on diverse corpora. Although these models show strong reasoning abilities, their performance varies significantly between languages due to the imbalanced distribution of training data. Existing approaches using sample-level translation for extensive multilingual pretraining and cross-lingual tuning face scalability challenges and often fail to capture nuanced reasoning processes across languages. In this paper, we introduce AdaMCOT (Adaptive Multilingual Chain-of-Thought), a framework that enhances multilingual factual reasoning by dynamically routing thought processes in intermediary "thinking languages" before generating target-language responses. AdaMCOT leverages a language-agnostic core and incorporates an adaptive, reward-based mechanism for selecting optimal reasoning pathways without requiring additional pretraining. Our comprehensive evaluation across multiple benchmarks demonstrates substantial improvements in both factual reasoning quality and cross-lingual consistency, with particularly strong performance gains in low-resource language settings. An in-depth analysis of the model's hidden states and semantic space further elucidates the underlying mechanism of our method. The results suggest that adaptive reasoning paths can effectively bridge the performance gap between high and low-resource languages while maintaining cultural and linguistic nuances.

Zhengyuan Liu, Nancy F. Chen

The Transformer-based models with the multi-head self-attention mechanism are widely used in natural language processing, and provide state-of-the-art results. While the pre-trained language backbones are shown to implicitly capture certain linguistic knowledge, explicitly incorporating structure-aware features can bring about further improvement on the downstream tasks. However, such enhancement often requires additional neural components and increases training parameter size. In this work, we investigate the attention head selection and manipulation strategy for feature injection from a network pruning perspective, and conduct a case study on dialogue summarization. We first rank attention heads in a Transformer-based summarizer with layer-wise importance. We then select the underused heads through extensive analysis, and inject structure-aware features by manipulating the selected heads. Experimental results show that the importance-based head selection is effective for feature injection, and dialogue summarization can be improved by incorporating coreference information via head manipulation.

Bryan Chen Zhengyu Tan, Zheng Weihua, Zhengyuan Liu et al.

As vision-language models (VLMs) are deployed globally, their ability to understand culturally situated knowledge becomes essential. Yet, existing evaluations largely assess static recall or isolated visual grounding, leaving unanswered whether VLMs possess robust and transferable cultural understanding. We introduce BLEnD-Vis, a multimodal, multicultural benchmark designed to evaluate the robustness of everyday cultural knowledge in VLMs across linguistic rephrasings and visual modalities. Building on the BLEnD dataset, BLEnD-Vis constructs 313 culturally grounded question templates spanning 16 regions and generates three aligned multiple-choice formats: (i) a text-only baseline querying from Region $\to$ Entity, (ii) an inverted text-only variant (Entity $\to$ Region), and (iii) a VQA-style version of (ii) with generated images. The resulting benchmark comprises 4,916 images and over 21,000 multiple-choice question (MCQ) instances, validated through human annotation. BLEnD-Vis reveals significant fragility in current VLM cultural knowledge; models exhibit performance drops under linguistic rephrasing and, whilst visual cues often aid performance, low cross-modal consistency highlights challenges in robustly integrating textual and visual understanding, particularly for lower-resource regions. BLEnD-Vis thus provides a crucial testbed for systematically analysing cultural robustness and multimodal grounding, exposing limitations and guiding the development of more culturally competent VLMs.

Weihua Zheng, Zhengyuan Liu, Tanmoy Chakraborty et al.

Large language models (LLMs) are now used worldwide, yet their multimodal understanding and reasoning often degrade outside Western, high-resource settings. We propose MMA-ASIA, a comprehensive framework to evaluate LLMs' cultural awareness with a focus on Asian contexts. MMA-ASIA centers on a human-curated, multilingual, and multimodally aligned multiple-choice benchmark covering 8 Asian countries and 10 languages, comprising 27,000 questions; over 79 percent require multi-step reasoning grounded in cultural context, moving beyond simple memorization. To our knowledge, this is the first dataset aligned at the input level across three modalities: text, image (visual question answering), and speech. This enables direct tests of cross-modal transfer. Building on this benchmark, we propose a five-dimensional evaluation protocol that measures: (i) cultural-awareness disparities across countries, (ii) cross-lingual consistency, (iii) cross-modal consistency, (iv) cultural knowledge generalization, and (v) grounding validity. To ensure rigorous assessment, a Cultural Awareness Grounding Validation Module detects "shortcut learning" by checking whether the requisite cultural knowledge supports correct answers. Finally, through comparative model analysis, attention tracing, and an innovative Vision-ablated Prefix Replay (VPR) method, we probe why models diverge across languages and modalities, offering actionable insights for building culturally reliable multimodal LLMs.

Galann Pennec, Zhengyuan Liu, Nicholas Asher et al.

Vision-Language Models (VLMs) often struggle to balance visual and textual information when summarizing complex multimodal inputs, such as entire TV show episodes. In this paper, we propose a zero-shot video-to-text summarization approach that builds its own screenplay representation of an episode, effectively integrating key video moments, dialogue, and character information into a unified document. Unlike previous approaches, we simultaneously generate screenplays and name the characters in zero-shot, using only the audio, video, and transcripts as input. Additionally, we highlight that existing summarization metrics can fail to assess the multimodal content in summaries. To address this, we introduce MFactSum, a multimodal metric that evaluates summaries with respect to both vision and text modalities. Using MFactSum, we evaluate our screenplay summaries on the SummScreen3D dataset, demonstrating superiority against state-of-the-art VLMs such as Gemini 1.5 by generating summaries containing 20% more relevant visual information while requiring 75% less of the video as input.

Bin Wang, Geyu Lin, Zhengyuan Liu et al.

Large language models (LLMs) have rapidly evolved as the foundation of various natural language processing (NLP) applications. Despite their wide use cases, their understanding of culturally-related concepts and reasoning remains limited. Meantime, there is a significant need to enhance these models' cultural reasoning capabilities, especially concerning underrepresented regions. This paper introduces a novel pipeline for extracting high-quality, culturally-related instruction tuning datasets from vast unstructured corpora. We utilize a self-instruction generation pipeline to identify cultural concepts and trigger instruction. By integrating with a general-purpose instruction tuning dataset, our model demonstrates enhanced capabilities in recognizing and understanding regional cultural nuances, thereby enhancing its reasoning capabilities. We conduct experiments across three regions: Singapore, the Philippines, and the United States, achieving performance improvement of up to 6%. Our research opens new avenues for extracting cultural instruction tuning sets directly from unstructured data, setting a precedent for future innovations in the field.

Zhengyuan Liu, Yong Keong Yap, Hai Leong Chieu et al.

Stance detection determines whether the author of a piece of text is in favor of, against, or neutral towards a specified target, and can be used to gain valuable insights into social media. The ubiquitous indirect referral of targets makes this task challenging, as it requires computational solutions to model semantic features and infer the corresponding implications from a literal statement. Moreover, the limited amount of available training data leads to subpar performance in out-of-domain and cross-target scenarios, as data-driven approaches are prone to rely on superficial and domain-specific features. In this work, we decompose the stance detection task from a linguistic perspective, and investigate key components and inference paths in this task. The stance triangle is a generic linguistic framework previously proposed to describe the fundamental ways people express their stance. We further expand it by characterizing the relationship between explicit and implicit objects. We then use the framework to extend one single training corpus with additional annotation. Experimental results show that strategically-enriched data can significantly improve the performance on out-of-domain and cross-target evaluation.

Yifu Zhang, Hongru Li, Tao Yang et al.

Lesion segmentation of ultrasound medical images based on deep learning techniques is a widely used method for diagnosing diseases. Although there is a large amount of ultrasound image data in medical centers and other places, labeled ultrasound datasets are a scarce resource, and it is likely that no datasets are available for new tissues/organs. Transfer learning provides the possibility to solve this problem, but there are too many features in natural images that are not related to the target domain. As a source domain, redundant features that are not conducive to the task will be extracted. Migration between ultrasound images can avoid this problem, but there are few types of public datasets, and it is difficult to find sufficiently similar source domains. Compared with natural images, ultrasound images have less information, and there are fewer transferable features between different ultrasound images, which may cause negative transfer. To this end, a multi-source adversarial transfer learning network for ultrasound image segmentation is proposed. Specifically, to address the lack of annotations, the idea of adversarial transfer learning is used to adaptively extract common features between a certain pair of source and target domains, which provides the possibility to utilize unlabeled ultrasound data. To alleviate the lack of knowledge in a single source domain, multi-source transfer learning is adopted to fuse knowledge from multiple source domains. In order to ensure the effectiveness of the fusion and maximize the use of precious data, a multi-source domain independent strategy is also proposed to improve the estimation of the target domain data distribution, which further increases the learning ability of the multi-source adversarial migration learning network in multiple domains.

Fangkai Jiao, Zhiyang Teng, Bosheng Ding et al.

Existing efforts to improve logical reasoning ability of language models have predominantly relied on supervised fine-tuning, hindering generalization to new domains and/or tasks. The development of Large Langauge Models (LLMs) has demonstrated the capacity of compressing abundant knowledge into a single proxy, enabling them to tackle multiple tasks effectively. Our preliminary experiments, nevertheless, show that LLMs do not show capability on logical reasoning. The performance of LLMs on logical reasoning benchmarks is far behind the existing state-of-the-art baselines. In this paper, we make the first attempt to investigate the feasibility of incorporating logical knowledge through self-supervised post-training, and activating it via in-context learning, which we termed as LogicLLM. Specifically, we devise an auto-regressive objective variant of MERIt and integrate it with two LLM series, i.e., FLAN-T5 and LLaMA, with parameter size ranging from 3 billion to 13 billion. The results on two challenging logical reasoning benchmarks demonstrate the effectiveness of LogicLLM. Besides, we conduct extensive ablation studies to analyze the key factors in designing logic-oriented proxy tasks.

Zhengyuan Liu, Nancy F. Chen

While multi-party conversations are often less structured than monologues and documents, they are implicitly organized by semantic level correlations across the interactive turns, and dialogue discourse analysis can be applied to predict the dependency structure and relations between the elementary discourse units, and provide feature-rich structural information for downstream tasks. However, the existing corpora with dialogue discourse annotation are collected from specific domains with limited sample sizes, rendering the performance of data-driven approaches poor on incoming dialogues without any domain adaptation. In this paper, we first introduce a Transformer-based parser, and assess its cross-domain performance. We next adopt three methods to gain domain integration from both data and language modeling perspectives to improve the generalization capability. Empirical results show that the neural parser can benefit from our proposed methods, and performs better on cross-domain dialogue samples.

Zhengyuan Liu, Ke Shi, Nancy F. Chen

Text discourse parsing weighs importantly in understanding information flow and argumentative structure in natural language, making it beneficial for downstream tasks. While previous work significantly improves the performance of RST discourse parsing, they are not readily applicable to practical use cases: (1) EDU segmentation is not integrated into most existing tree parsing frameworks, thus it is not straightforward to apply such models on newly-coming data. (2) Most parsers cannot be used in multilingual scenarios, because they are developed only in English. (3) Parsers trained from single-domain treebanks do not generalize well on out-of-domain inputs. In this work, we propose a document-level multilingual RST discourse parsing framework, which conducts EDU segmentation and discourse tree parsing jointly. Moreover, we propose a cross-translation augmentation strategy to enable the framework to support multilingual parsing and improve its domain generality. Experimental results show that our model achieves state-of-the-art performance on document-level multilingual RST parsing in all sub-tasks.

Zhengyuan Liu, Nancy F. Chen

In this paper, we propose a controllable neural generation framework that can flexibly guide dialogue summarization with personal named entity planning. The conditional sequences are modulated to decide what types of information or what perspective to focus on when forming summaries to tackle the under-constrained problem in summarization tasks. This framework supports two types of use cases: (1) Comprehensive Perspective, which is a general-purpose case with no user-preference specified, considering summary points from all conversational interlocutors and all mentioned persons; (2) Focus Perspective, positioning the summary based on a user-specified personal named entity, which could be one of the interlocutors or one of the persons mentioned in the conversation. During training, we exploit occurrence planning of personal named entities and coreference information to improve temporal coherence and to minimize hallucination in neural generation. Experimental results show that our proposed framework generates fluent and factually consistent summaries under various planning controls using both objective metrics and human evaluations.

Zhengyuan Liu, Nancy F. Chen

Recently abstractive spoken language summarization raises emerging research interest, and neural sequence-to-sequence approaches have brought significant performance improvement. However, summarizing long meeting transcripts remains challenging. Due to the large length of source contents and targeted summaries, neural models are prone to be distracted on the context, and produce summaries with degraded quality. Moreover, pre-trained language models with input length limitations cannot be readily applied to long sequences. In this work, we first analyze the linguistic characteristics of meeting transcripts on a representative corpus, and find that the sentences comprising the summary correlate with the meeting agenda. Based on this observation, we propose a dynamic sliding window strategy for meeting summarization. Experimental results show that performance benefit from the proposed method, and outputs obtain higher factual consistency than the base model.

Zhengyuan Liu, Ke Shi, Nancy F. Chen

Summarizing conversations via neural approaches has been gaining research traction lately, yet it is still challenging to obtain practical solutions. Examples of such challenges include unstructured information exchange in dialogues, informal interactions between speakers, and dynamic role changes of speakers as the dialogue evolves. Many of such challenges result in complex coreference links. Therefore, in this work, we investigate different approaches to explicitly incorporate coreference information in neural abstractive dialogue summarization models to tackle the aforementioned challenges. Experimental results show that the proposed approaches achieve state-of-the-art performance, implying it is useful to utilize coreference information in dialogue summarization. Evaluation results on factual correctness suggest such coreference-aware models are better at tracing the information flow among interlocutors and associating accurate status/actions with the corresponding interlocutors and person mentions.

Taha Aksu, Zhengyuan Liu, Min-Yen Kan et al.

Augmentation of task-oriented dialogues has followed standard methods used for plain-text such as back-translation, word-level manipulation, and paraphrasing despite its richly annotated structure. In this work, we introduce an augmentation framework that utilizes belief state annotations to match turns from various dialogues and form new synthetic dialogues in a bottom-up manner. Unlike other augmentation strategies, it operates with as few as five examples. Our augmentation strategy yields significant improvements when both adapting a DST model to a new domain, and when adapting a language model to the DST task, on evaluations with TRADE and TOD-BERT models. Further analysis shows that our model performs better on seen values during training, and it is also more robust to unseen values. We conclude that exploiting belief state annotations enhances dialogue augmentation and results in improved models in n-shot training scenarios.

Ke Shi, Zhengyuan Liu, Nancy F. Chen

Document-level discourse parsing, in accordance with the Rhetorical Structure Theory (RST), remains notoriously challenging. Challenges include the deep structure of document-level discourse trees, the requirement of subtle semantic judgments, and the lack of large-scale training corpora. To address such challenges, we propose to exploit robust representations derived from multiple levels of granularity across syntax and semantics, and in turn incorporate such representations in an end-to-end encoder-decoder neural architecture for more resourceful discourse processing. In particular, we first use a pre-trained contextual language model that embodies high-order and long-range dependency to enable finer-grain semantic, syntactic, and organizational representations. We further encode such representations with boundary and hierarchical information to obtain more refined modeling for document-level discourse processing. Experimental results show that our parser achieves the state-of-the-art performance, approaching human-level performance on the benchmarked RST dataset.

Zhengyuan Liu, Ke Shi, Nancy F. Chen

Text discourse parsing plays an important role in understanding information flow and argumentative structure in natural language. Previous research under the Rhetorical Structure Theory (RST) has mostly focused on inducing and evaluating models from the English treebank. However, the parsing tasks for other languages such as German, Dutch, and Portuguese are still challenging due to the shortage of annotated data. In this work, we investigate two approaches to establish a neural, cross-lingual discourse parser via: (1) utilizing multilingual vector representations; and (2) adopting segment-level translation of the source content. Experiment results show that both methods are effective even with limited training data, and achieve state-of-the-art performance on cross-lingual, document-level discourse parsing on all sub-tasks.

Zhengyuan Liu, Ke Shi, Nancy F. Chen

Much progress has been made in text summarization, fueled by neural architectures using large-scale training corpora. However, in the news domain, neural models easily overfit by leveraging position-related features due to the prevalence of the inverted pyramid writing style. In addition, there is an unmet need to generate a variety of summaries for different users. In this paper, we propose a neural framework that can flexibly control summary generation by introducing a set of sub-aspect functions (i.e. importance, diversity, position). These sub-aspect functions are regulated by a set of control codes to decide which sub-aspect to focus on during summary generation. We demonstrate that extracted summaries with minimal position bias is comparable with those generated by standard models that take advantage of position preference. We also show that news summaries generated with a focus on diversity can be more preferred by human raters. These results suggest that a more flexible neural summarization framework providing more control options could be desirable in tailoring to different user preferences, which is useful since it is often impractical to articulate such preferences for different applications a priori.

Zhengyuan Liu

The morphological attributes of retinal vessels, such as length, width, tortuosity and branching pattern and angles, play an important role in diagnosis, screening, treatment, and evaluation of various cardiovascular and ophthalmologic diseases such as diabetes, hypertension and arteriosclerosis. The crucial step before extracting these morphological characteristics of retinal vessels from retinal fundus images is vessel segmentation. In this work, we propose a method for retinal vessel segmentation based on fully convolutional networks. Thousands of patches are extracted from each retinal image and then fed into the network, and data argumentation is applied by rotating extracted patches. Two architectures of fully convolutional networks, U-Net and LadderNet, are used for vessel segmentation. The performance of our method is evaluated on three public datasets: DRIVE, STARE, and CHASE\_DB1. Experimental results of our method show superior performance compared to recent state-of-the-art methods.