Roy Ka-Wei Lee

Deepanway Ghosal, Navonil Majumder, Roy Ka-Wei Lee et al. · deepmind

Visual question answering (VQA) is the task of answering questions about an image. The task assumes an understanding of both the image and the question to provide a natural language answer. VQA has gained popularity in recent years due to its potential applications in a wide range of fields, including robotics, education, and healthcare. In this paper, we focus on knowledge-augmented VQA, where answering the question requires commonsense knowledge, world knowledge, and reasoning about ideas and concepts not present in the image. We propose a multimodal framework that uses language guidance (LG) in the form of rationales, image captions, scene graphs, etc to answer questions more accurately. We benchmark our method on the multi-choice question-answering task of the A-OKVQA, Science-QA, VSR, and IconQA datasets using CLIP and BLIP models. We show that the use of language guidance is a simple but powerful and effective strategy for visual question answering. Our language guidance improves the performance of CLIP by 7.6% and BLIP-2 by 4.8% in the challenging A-OKVQA dataset. We also observe consistent improvement in performance on the Science-QA, VSR, and IconQA datasets when using the proposed language guidances. The implementation of LG-VQA is publicly available at https:// github.com/declare-lab/LG-VQA.

Zhiqiang Hu, Lei Wang, Yihuai Lan et al.

The success of large language models (LLMs), like GPT-4 and ChatGPT, has led to the development of numerous cost-effective and accessible alternatives that are created by finetuning open-access LLMs with task-specific data (e.g., ChatDoctor) or instruction data (e.g., Alpaca). Among the various fine-tuning methods, adapter-based parameter-efficient fine-tuning (PEFT) is undoubtedly one of the most attractive topics, as it only requires fine-tuning a few external parameters instead of the entire LLMs while achieving comparable or even better performance. To enable further research on PEFT methods of LLMs, this paper presents LLM-Adapters, an easy-to-use framework that integrates various adapters into LLMs and can execute these adapter-based PEFT methods of LLMs for different tasks. The framework includes state-of-the-art open-access LLMs such as LLaMA, BLOOM, and GPT-J, as well as widely used adapters such as Series adapters, Parallel adapter, Prompt-based learning and Reparametrization-based methods. Moreover, we conduct extensive empirical studies on the impact of adapter types, placement locations, and hyper-parameters to the best design for each adapter-based methods. We evaluate the effectiveness of the adapters on fourteen datasets from two different reasoning tasks, Arithmetic Reasoning and Commonsense Reasoning. The results demonstrate that using adapter-based PEFT in smaller-scale LLMs (7B) with few extra trainable parameters yields comparable, and in some cases superior, performance to powerful LLMs (175B) in zero-shot inference on both reasoning tasks.

Md Rabiul Awal, Roy Ka-Wei Lee, Eshaan Tanwar et al. · mila

Hate speech in social media is a growing phenomenon, and detecting such toxic content has recently gained significant traction in the research community. Existing studies have explored fine-tuning language models (LMs) to perform hate speech detection, and these solutions have yielded significant performance. However, most of these studies are limited to detecting hate speech only in English, neglecting the bulk of hateful content that is generated in other languages, particularly in low-resource languages. Developing a classifier that captures hate speech and nuances in a low-resource language with limited data is extremely challenging. To fill the research gap, we propose HateMAML, a model-agnostic meta-learning-based framework that effectively performs hate speech detection in low-resource languages. HateMAML utilizes a self-supervision strategy to overcome the limitation of data scarcity and produces better LM initialization for fast adaptation to an unseen target language (i.e., cross-lingual transfer) or other hate speech datasets (i.e., domain generalization). Extensive experiments are conducted on five datasets across eight different low-resource languages. The results show that HateMAML outperforms the state-of-the-art baselines by more than 3% in the cross-domain multilingual transfer setting. We also conduct ablation studies to analyze the characteristics of HateMAML.

Rui Cao, Roy Ka-Wei Lee, Wen-Haw Chong et al.

Hateful meme classification is a challenging multimodal task that requires complex reasoning and contextual background knowledge. Ideally, we could leverage an explicit external knowledge base to supplement contextual and cultural information in hateful memes. However, there is no known explicit external knowledge base that could provide such hate speech contextual information. To address this gap, we propose PromptHate, a simple yet effective prompt-based model that prompts pre-trained language models (PLMs) for hateful meme classification. Specifically, we construct simple prompts and provide a few in-context examples to exploit the implicit knowledge in the pre-trained RoBERTa language model for hateful meme classification. We conduct extensive experiments on two publicly available hateful and offensive meme datasets. Our experimental results show that PromptHate is able to achieve a high AUC of 90.96, outperforming state-of-the-art baselines on the hateful meme classification task. We also perform fine-grained analyses and case studies on various prompt settings and demonstrate the effectiveness of the prompts on hateful meme classification.

Qing Meng, Tharun Suresh, Roy Ka-Wei Lee et al.

Tweets are the most concise form of communication in online social media, wherein a single tweet has the potential to make or break the discourse of the conversation. Online hate speech is more accessible than ever, and stifling its propagation is of utmost importance for social media companies and users for congenial communication. Most of the research barring a recent few has focused on classifying an individual tweet regardless of the tweet thread/context leading up to that point. One of the classical approaches to curb hate speech is to adopt a reactive strategy after the hate speech postage. The ex-post facto strategy results in neglecting subtle posts that do not show the potential to instigate hate speech on their own but may portend in the subsequent discussion ensuing in the post's replies. In this paper, we propose DRAGNET++, which aims to predict the intensity of hatred that a tweet can bring in through its reply chain in the future. It uses the semantic and propagating structure of the tweet threads to maximize the contextual information leading up to and the fall of hate intensity at each subsequent tweet. We explore three publicly available Twitter datasets -- Anti-Racism contains the reply tweets of a collection of social media discourse on racist remarks during US political and Covid-19 background; Anti-Social presents a dataset of 40 million tweets amidst the COVID-19 pandemic on anti-social behaviours; and Anti-Asian presents Twitter datasets collated based on anti-Asian behaviours during COVID-19 pandemic. All the curated datasets consist of structural graph information of the Tweet threads. We show that DRAGNET++ outperforms all the state-of-the-art baselines significantly. It beats the best baseline by an 11% margin on the Person correlation coefficient and a decrease of 25% on RMSE for the Anti-Racism dataset with a similar performance on the other two datasets.

Rui Cao, Ming Shan Hee, Adriel Kuek et al.

Hateful meme detection is a challenging multimodal task that requires comprehension of both vision and language, as well as cross-modal interactions. Recent studies have tried to fine-tune pre-trained vision-language models (PVLMs) for this task. However, with increasing model sizes, it becomes important to leverage powerful PVLMs more efficiently, rather than simply fine-tuning them. Recently, researchers have attempted to convert meme images into textual captions and prompt language models for predictions. This approach has shown good performance but suffers from non-informative image captions. Considering the two factors mentioned above, we propose a probing-based captioning approach to leverage PVLMs in a zero-shot visual question answering (VQA) manner. Specifically, we prompt a frozen PVLM by asking hateful content-related questions and use the answers as image captions (which we call Pro-Cap), so that the captions contain information critical for hateful content detection. The good performance of models with Pro-Cap on three benchmarks validates the effectiveness and generalization of the proposed method.

Ming Shan Hee, Roy Ka-Wei Lee, Wen-Haw Chong

Hateful meme detection is a new multimodal task that has gained significant traction in academic and industry research communities. Recently, researchers have applied pre-trained visual-linguistic models to perform the multimodal classification task, and some of these solutions have yielded promising results. However, what these visual-linguistic models learn for the hateful meme classification task remains unclear. For instance, it is unclear if these models are able to capture the derogatory or slurs references in multimodality (i.e., image and text) of the hateful memes. To fill this research gap, this paper propose three research questions to improve our understanding of these visual-linguistic models performing the hateful meme classification task. We found that the image modality contributes more to the hateful meme classification task, and the visual-linguistic models are able to perform visual-text slurs grounding to a certain extent. Our error analysis also shows that the visual-linguistic models have acquired biases, which resulted in false-positive predictions.

Yuhao Wu, Maojia Song, Yihuai Lan et al.

Understanding the physical structure is essential for real-world applications such as embodied agents, interactive design, and long-horizon manipulation. Yet, prevailing Vision-Language Model (VLM) evaluations still center on structure-agnostic, single-turn setups (e.g., VQA), which fail to assess agents' ability to reason about how geometry, contact, and support relations jointly constrain what actions are possible in a dynamic environment. To address this gap, we introduce the Causal Hierarchy of Actions and Interactions (CHAIN) benchmark, an interactive 3D, physics-driven testbed designed to evaluate whether models can understand, plan, and execute structured action sequences grounded in physical constraints. CHAIN shifts evaluation from passive perception to active problem solving, spanning tasks such as interlocking mechanical puzzles and 3D stacking and packing. We conduct a comprehensive study of state-of-the-art VLMs and diffusion-based models under unified interactive settings. Our results show that top-performing models still struggle to internalize physical structure and causal constraints, often failing to produce reliable long-horizon plans and cannot robustly translate perceived structure into effective actions. The project is available at https://social-ai-studio.github.io/CHAIN/.

Adam Valen Levinson, Abhay Goyal, Roger Ho Chun Man et al.

Telehealth is a valuable tool for primary health care (PHC), where depression is a common condition. PHC is the first point of contact for most people with depression, but about 25% of diagnoses made by PHC physicians are inaccurate. Many other barriers also hinder depression detection and treatment in PHC. Artificial intelligence (AI) may help reduce depression misdiagnosis in PHC and improve overall diagnosis and treatment outcomes. Telehealth consultations often have video issues, such as poor connectivity or dropped calls. Audio-only telehealth is often more practical for lower-income patients who may lack stable internet connections. Thus, our study focused on using audio data to predict depression risk. The objectives were to: 1) Collect audio data from 24 people (12 with depression and 12 without mental health or major health condition diagnoses); 2) Build a machine learning model to predict depression risk. TPOT, an autoML tool, was used to select the best machine learning algorithm, which was the K-nearest neighbors classifier. The selected model had high performance in classifying depression risk (Precision: 0.98, Recall: 0.93, F1-Score: 0.96). These findings may lead to a range of tools to help screen for and treat depression. By developing tools to detect depression risk, patients can be routed to AI-driven chatbots for initial screenings. Partnerships with a range of stakeholders are crucial to implementing these solutions. Moreover, ethical considerations, especially around data privacy and potential biases in AI models, need to be at the forefront of any AI-driven intervention in mental health care.

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.

Abhay Goyal, Muhammad Siddique, Nimay Parekh et al.

Recent developments in natural language processing have demonstrated the potential of large language models (LLMs) to improve a range of educational and learning outcomes. Of recent chatbots based on LLMs, ChatGPT and Bard have made it clear that artificial intelligence (AI) technology will have significant implications on the way we obtain and search for information. However, these tools sometimes produce text that is convincing, but often incorrect, known as hallucinations. As such, their use can distort scientific facts and spread misinformation. To counter polarizing responses on these tools, it is critical to provide an overview of such responses so stakeholders can determine which topics tend to produce more contentious responses -- key to developing targeted regulatory policy and interventions. In addition, there currently exists no annotated dataset of ChatGPT and Bard responses around possibly polarizing topics, central to the above aims. We address the indicated issues through the following contribution: Focusing on highly polarizing topics in the US, we created and described a dataset of ChatGPT and Bard responses. Broadly, our results indicated a left-leaning bias for both ChatGPT and Bard, with Bard more likely to provide responses around polarizing topics. Bard seemed to have fewer guardrails around controversial topics, and appeared more willing to provide comprehensive, and somewhat human-like responses. Bard may thus be more likely abused by malicious actors. Stakeholders may utilize our findings to mitigate misinformative and/or polarizing responses from LLMs

Yuhao Wu, Ming Shan Hee, Zhiqing Hu et al.

Current benchmarks like Needle-in-a-Haystack (NIAH), Ruler, and Needlebench focus on models' ability to understand long-context input sequences but fail to capture a critical dimension: the generation of high-quality long-form text. Applications such as design proposals, technical documentation, and creative writing rely on coherent, instruction-following outputs over extended sequences - a challenge that existing benchmarks do not adequately address. To fill this gap, we introduce LongGenBench, a novel benchmark designed to rigorously evaluate large language models' (LLMs) ability to generate long text while adhering to complex instructions. Through tasks requiring specific events or constraints within generated text, LongGenBench evaluates model performance across four distinct scenarios, three instruction types, and two generation-lengths (16K and 32K tokens). Our evaluation of ten state-of-the-art LLMs reveals that, despite strong results on Ruler, all models struggled with long text generation on LongGenBench, particularly as text length increased. This suggests that current LLMs are not yet equipped to meet the demands of real-world, long-form text generation.

Nirmalendu Prakash, Narmeen Oozeer, Michael Lan et al.

Despite the rapid adoption of text-to-image (T2I) diffusion models, causal and representation-level analysis remains fragmented and largely limited to isolated probing techniques. To address this gap, we introduce DreamReader: a unified framework that formalizes diffusion interpretability as composable representation operators spanning activation extraction, causal patching, structured ablations, and activation steering across modules and timesteps. DreamReader provides a model-agnostic abstraction layer enabling systematic analysis and intervention across diffusion architectures. Beyond consolidating existing methods, DreamReader introduces three novel intervention primitives for diffusion models: (1) representation fine-tuning (LoReFT) for subspace-constrained internal adaptation; (2) classifier-guided gradient steering using MLP probes trained on activations; and (3) component-level cross-model mapping for systematic study of transferability of representations across modalities. These mechanisms allows us to do lightweight white-box interventions on T2I models by drawing inspiration from interpretability techniques on LLMs. We demonstrate DreamReader through controlled experiments that (i) perform activation stitching between two models, and (ii) apply LoReFT to steer multiple activation units, reliably injecting a target concept into the generated images. Experiments are specified declaratively and executed in controlled batched pipelines to enable reproducible large-scale analysis. Across multiple case studies, we show that techniques adapted from language model interpretability yield promising and controllable interventions in diffusion models. DreamReader is released as an open source toolkit for advancing research on T2I interpretability.

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.

Kaidong Feng, Zhu Sun, Roy Ka-Wei Lee et al.

Smartphone sensing offers an unobtrusive and scalable way to track daily behaviors linked to mental health, capturing changes in sleep, mobility, and phone use that often precede symptoms of stress, anxiety, or depression. While most prior studies focus on detection that responds to existing conditions, forecasting mental health enables proactive support through Just-in-Time Adaptive Interventions. In this paper, we present the first comprehensive benchmarking study comparing traditional machine learning (ML), deep learning (DL), and large language model (LLM) approaches for mental health forecasting using the College Experience Sensing (CES) dataset, the most extensive longitudinal dataset of college student mental health to date. We systematically evaluate models across temporal windows, feature granularities, personalization strategies, and class imbalance handling. Our results show that DL models, particularly Transformer (Macro-F1 = 0.58), achieve the best overall performance, while LLMs show strength in contextual reasoning but weaker temporal modeling. Personalization substantially improves forecasts of severe mental health states. By revealing how different modeling approaches interpret phone sensing behavioral data over time, this work lays the groundwork for next-generation, adaptive, and human-centered mental health technologies that can advance both research and real-world well-being.

Lynnette Hui Xian Ng, Iain Cruickshank, Roy Ka-Wei Lee

Large Language Models (LLMs) have demonstrated remarkable capabilities in executing tasks based on natural language queries. However, these models, trained on curated datasets, inherently embody biases ranging from racial to national and gender biases. It remains uncertain whether these biases impact the performance of LLMs for certain tasks. In this study, we investigate the political biases of LLMs within the stance classification task, specifically examining whether these models exhibit a tendency to more accurately classify politically-charged stances. Utilizing three datasets, seven LLMs, and four distinct prompting schemes, we analyze the performance of LLMs on politically oriented statements and targets. Our findings reveal a statistically significant difference in the performance of LLMs across various politically oriented stance classification tasks. Furthermore, we observe that this difference primarily manifests at the dataset level, with models and prompting schemes showing statistically similar performances across different stance classification datasets. Lastly, we observe that when there is greater ambiguity in the target the statement is directed towards, LLMs have poorer stance classification accuracy. Code & Dataset: http://doi.org/10.5281/zenodo.12938478

Chia-Yu Hung, Zhiqiang Hu, Yujia Hu et al.

Authorship verification (AV) is a fundamental task in natural language processing (NLP) and computational linguistics, with applications in forensic analysis, plagiarism detection, and identification of deceptive content. Existing AV techniques, including traditional stylometric and deep learning approaches, face limitations in terms of data requirements and lack of explainability. To address these limitations, this paper proposes PromptAV, a novel technique that leverages Large-Language Models (LLMs) for AV by providing step-by-step stylometric explanation prompts. PromptAV outperforms state-of-the-art baselines, operates effectively with limited training data, and enhances interpretability through intuitive explanations, showcasing its potential as an effective and interpretable solution for the AV task.

Jingyi Kang, Junyu Lu, Bo Xu et al.

Large language models (LLMs) require robust toxicity evaluation beyond explicit wording. This setting remains underexplored in Chinese, where toxicity may combine semantic indirectness with surface obfuscation. We introduce Chinese Implicit Toxicity Attack (CITA), a controlled red-team evaluation and defense-data generation framework, not a deployable evasion tool. CITA uses three stages: (i) Harmful Intent Learning, (ii) Implicit Toxicity Enhancement, and (iii) Obfuscation Variant Rewriting, to preserve harmful intent, increase implicitness, and add controlled surface variants. On CITA-generated evaluation samples, the seven tested detectors exhibit substantial missed-detection risks, reaching an average ASR of 69.48%; human evaluation further confirms preserved harmfulness and increased implicitness/evasiveness. As a downstream defense application, we fine-tune a Chinese Implicit Toxicity Defense model (CITD) with CITA-generated red-team data, showing that such data can improve robustness through additional training.

Han Wang, Tan Rui Yang, Usman Naseem et al.

Hate speech is a pressing issue in modern society, with significant effects both online and offline. Recent research in hate speech detection has primarily centered on text-based media, largely overlooking multimodal content such as videos. Existing studies on hateful video datasets have predominantly focused on English content within a Western context and have been limited to binary labels (hateful or non-hateful), lacking detailed contextual information. This study presents MultiHateClip1 , an novel multilingual dataset created through hate lexicons and human annotation. It aims to enhance the detection of hateful videos on platforms such as YouTube and Bilibili, including content in both English and Chinese languages. Comprising 2,000 videos annotated for hatefulness, offensiveness, and normalcy, this dataset provides a cross-cultural perspective on gender-based hate speech. Through a detailed examination of human annotation results, we discuss the differences between Chinese and English hateful videos and underscore the importance of different modalities in hateful and offensive video analysis. Evaluations of state-of-the-art video classification models, such as VLM, GPT-4V and Qwen-VL, on MultiHateClip highlight the existing challenges in accurately distinguishing between hateful and offensive content and the urgent need for models that are both multimodally and culturally nuanced. MultiHateClip represents a foundational advance in enhancing hateful video detection by underscoring the necessity of a multimodal and culturally sensitive approach in combating online hate speech.

Jaehyeok Lee, Xiaoyuan Yi, Jing Yao et al.

As LLMs are globally deployed, aligning their cultural value orientations is critical for safety and user engagement. However, existing benchmarks face the Construct-Composition-Context ($C^3$) challenge: relying on discriminative, multiple-choice formats that probe value knowledge rather than true orientations, overlook subcultural heterogeneity, and mismatch with real-world open-ended generation. We introduce DOVE, a distributional evaluation framework that directly compares human-written text distributions with LLM-generated outputs. DOVE utilizes a rate-distortion variational optimization objective to construct a compact value-codebook from 10K documents, mapping text into a structured value space to filter semantic noise. Alignment is measured using unbalanced optimal transport, capturing intra-cultural distributional structures and sub-group diversity. Experiments across 12 LLMs show that DOVE achieves superior predictive validity, attaining a 31.56% correlation with downstream tasks, while maintaining high reliability with as few as 500 samples per culture.

Han Wang, Deyi Ji, Lanyun Zhu et al.

Live streaming platforms require real-time monitoring and reaction to social signals, utilizing partial and asynchronous evidence from video, text, and audio. We propose StreamSense, a streaming detector that couples a lightweight streaming encoder with selective routing to a Vision-Language Model (VLM) expert. StreamSense handles most timestamps with the lightweight streaming encoder, escalates hard/ambiguous cases to the VLM, and defers decisions when context is insufficient. The encoder is trained using (i) a cross-modal contrastive term to align visual/audio cues with textual signals, and (ii) an IoU-weighted loss that down-weights poorly overlapping target segments, mitigating label interference across segment boundaries. We evaluate StreamSense on multiple social streaming detection tasks (e.g., sentiment classification and hate content moderation), and the results show that StreamSense achieves higher accuracy than VLM-only streaming while only occasionally invoking the VLM, thereby reducing average latency and compute. Our results indicate that selective escalation and deferral are effective primitives for understanding streaming social tasks. Code is publicly available on GitHub.

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.

Han Wang, Deyi Ji, Junyu Lu et al.

Accurate detection of offensive content on social media demands high-quality labeled data; however, such data is often scarce due to the low prevalence of offensive instances and the high cost of manual annotation. To address this low-resource challenge, we propose a self-training framework that leverages abundant unlabeled data through collaborative pseudo-labeling. Starting with a lightweight classifier trained on limited labeled data, our method iteratively assigns pseudo-labels to unlabeled instances with the support of Multi-Agent Vision-Language Models (MA-VLMs). Un-labeled data on which the classifier and MA-VLMs agree are designated as the Agreed-Unknown set, while conflicting samples form the Disagreed-Unknown set. To enhance label reliability, MA-VLMs simulate dual perspectives, moderator and user, capturing both regulatory and subjective viewpoints. The classifier is optimized using a novel Positive-Negative-Unlabeled (PNU) loss, which jointly exploits labeled, Agreed-Unknown, and Disagreed-Unknown data while mitigating pseudo-label noise. Experiments on benchmark datasets demonstrate that our framework substantially outperforms baselines under limited supervision and approaches the performance of large-scale models

Ngai Lam Ho, Roy Ka-Wei Lee, Kwan Hui Lim

An essential task for tourists having a pleasant holiday is to have a well-planned itinerary with relevant recommendations, especially when visiting unfamiliar cities. Many tour recommendation tools only take into account a limited number of factors, such as popular Points of Interest (POIs) and routing constraints. Consequently, the solutions they provide may not always align with the individual users of the system. We propose an iterative algorithm in this paper, namely: BTREC (BERT-based Trajectory Recommendation), that extends from the POIBERT embedding algorithm to recommend personalized itineraries on POIs using the BERT framework. Our BTREC algorithm incorporates users' demographic information alongside past POI visits into a modified BERT language model to recommend a personalized POI itinerary prediction given a pair of source and destination POIs. Our recommendation system can create a travel itinerary that maximizes POIs visited, while also taking into account user preferences for categories of POIs and time availability. Our recommendation algorithm is largely inspired by the problem of sentence completion in natural language processing (NLP). Using a dataset of eight cities of different sizes, our experimental results demonstrate that our proposed algorithm is stable and outperforms many other sequence prediction algorithms, measured by recall, precision, and F1-scores.

Nirmalendu Prakash, Narmeen Fatimah Oozeer, Xin Su et al.

CLIP retrieval is typically framed as a pointwise similarity problem in a shared embedding space. While CLIP achieves strong global cross-modal alignment, many retrieval failures arise from local geometric inconsistencies: nearby items are incorrectly ordered, leading to systematic confusions (e.g., pentagon vs. hexagon) and produces diffuse, weakly controlled result sets. Prior work largely optimizes for point wise relevance or finetuning to mitigate these problems. We instead view retrieval as a problem of neighborhood alignment. Our work introduces (1) neighborhood-level re-ranking via Hungarian matching, which rewards structural consistency; (2) query-conditioned local steering, where directions derived from contrastive neighborhoods around the query reshape retrieval. We show that these techniques improve retrieval performance on attribute-binding and compositional retrieval tasks. Together, these methods operate on local neighborhoods but serve different roles: re-ranking rewards alignment whereas local steering controls neighborhood structure. This shows that retrieval quality and controllability depend critically on local structure, which can be exploited at inference time without retraining.

Rui Yang Tan, Yujia Hu, Roy Ka-Wei Lee

Multimodal Large Language Models (MLLMs) extend text-only LLMs with visual reasoning, but also introduce new safety failure modes under visually grounded instructions. We study comic-template jailbreaks that embed harmful goals inside simple three-panel visual narratives and prompt the model to role-play and "complete the comic." Building on JailbreakBench and JailbreakV, we introduce ComicJailbreak, a comic-based jailbreak benchmark with 1,167 attack instances spanning 10 harm categories and 5 task setups. Across 15 state-of-the-art MLLMs (six commercial and nine open-source), comic-based attacks achieve success rates comparable to strong rule-based jailbreaks and substantially outperform plain-text and random-image baselines, with ensemble success rates exceeding 90% on several commercial models. Then, with the existing defense methodologies, we show that these methods are effective against the harmful comics, they will induce a high refusal rate when prompted with benign prompts. Finally, using automatic judging and targeted human evaluation, we show that current safety evaluators can be unreliable on sensitive but non-harmful content. Our findings highlight the need for safety alignment robust to narrative-driven multimodal jailbreaks.

Yujia Hu, Zhiqiang Hu, Chun-Wei Seah et al.

Large Language Models (LLMs) have demonstrated remarkable proficiency in a wide range of NLP tasks. However, when it comes to authorship verification (AV) tasks, which involve determining whether two given texts share the same authorship, even advanced models like ChatGPT exhibit notable limitations. This paper introduces a novel approach, termed InstructAV, for authorship verification. This approach utilizes LLMs in conjunction with a parameter-efficient fine-tuning (PEFT) method to simultaneously improve accuracy and explainability. The distinctiveness of InstructAV lies in its ability to align classification decisions with transparent and understandable explanations, representing a significant progression in the field of authorship verification. Through comprehensive experiments conducted across various datasets, InstructAV demonstrates its state-of-the-art performance on the AV task, offering high classification accuracy coupled with enhanced explanation reliability.

Han Wang, Rui Yang Tan, Roy Ka-Wei Lee

Detecting hate speech in online content is essential to ensuring safer digital spaces. While significant progress has been made in text and meme modalities, video-based hate speech detection remains under-explored, hindered by a lack of annotated datasets and the high cost of video annotation. This gap is particularly problematic given the growing reliance on large models, which demand substantial amounts of training data. To address this challenge, we leverage meme datasets as both a substitution and an augmentation strategy for training hateful video detection models. Our approach introduces a human-assisted reannotation pipeline to align meme dataset labels with video datasets, ensuring consistency with minimal labeling effort. Using two state-of-the-art vision-language models, we demonstrate that meme data can substitute for video data in resource-scarce scenarios and augment video datasets to achieve further performance gains. Our results consistently outperform state-of-the-art benchmarks, showcasing the potential of cross-modal transfer learning for advancing hateful video detection. Dataset and code are available at https://github.com/Social-AI-Studio/CrossModalTransferLearning.

Ngai Lam Ho, Roy Ka-Wei Lee, Kwan Hui Lim

When traveling to an unfamiliar city for holidays, tourists often rely on guidebooks, travel websites, or recommendation systems to plan their daily itineraries and explore popular points of interest (POIs). However, these approaches may lack optimization in terms of time feasibility, localities, and user preferences. In this paper, we propose the SBTRec algorithm: a BERT-based Trajectory Recommendation with sentiment analysis, for recommending personalized sequences of POIs as itineraries. The key contributions of this work include analyzing users' check-ins and uploaded photos to understand the relationship between POI visits and distance. We introduce SBTRec, which encompasses sentiment analysis to improve recommendation accuracy by understanding users' preferences and satisfaction levels from reviews and comments about different POIs. Our proposed algorithms are evaluated against other sequence prediction methods using datasets from 8 cities. The results demonstrate that SBTRec achieves an average F1 score of 61.45%, outperforming baseline algorithms. The paper further discusses the flexibility of the SBTRec algorithm, its ability to adapt to different scenarios and cities without modification, and its potential for extension by incorporating additional information for more reliable predictions. Overall, SBTRec provides personalized and relevant POI recommendations, enhancing tourists' overall trip experiences. Future work includes fine-tuning personalized embeddings for users, with evaluation of users' comments on POIs,~to further enhance prediction accuracy.

Yuhao Wu, Yushi Bai, Zhiqiang Hu et al. · tsinghua

Ultra-long generation by large language models (LLMs) is a widely demanded scenario, yet it remains a significant challenge due to their maximum generation length limit and overall quality degradation as sequence length increases. Previous approaches, exemplified by LongWriter, typically rely on ''teaching'', which involves supervised fine-tuning (SFT) on synthetic long-form outputs. However, this strategy heavily depends on synthetic SFT data, which is difficult and costly to construct, often lacks coherence and consistency, and tends to be overly artificial and structurally monotonous. In this work, we propose an incentivization-based approach that, starting entirely from scratch and without relying on any annotated or synthetic data, leverages reinforcement learning (RL) to foster the emergence of ultra-long, high-quality text generation capabilities in LLMs. We perform RL training starting from a base model, similar to R1-Zero, guiding it to engage in reasoning that facilitates planning and refinement during the writing process. To support this, we employ specialized reward models that steer the LLM towards improved length control, writing quality, and structural formatting. Experimental evaluations show that our LongWriter-Zero model, trained from Qwen2.5-32B, consistently outperforms traditional SFT methods on long-form writing tasks, achieving state-of-the-art results across all metrics on WritingBench and Arena-Write, and even surpassing 100B+ models such as DeepSeek R1 and Qwen3-235B. We open-source our data and model checkpoints under https://huggingface.co/THU-KEG/LongWriter-Zero-32B

Ziyu Ge, Yuhao Wu, Daniel Wai Kit Chin et al.

Large Language Models (LLMs) augmented with retrieval mechanisms have demonstrated significant potential in fact-checking tasks by integrating external knowledge. However, their reliability decreases when confronted with conflicting evidence from sources of varying credibility. This paper presents the first systematic evaluation of Retrieval-Augmented Generation (RAG) models for fact-checking in the presence of conflicting evidence. To support this study, we introduce \textbf{CONFACT} (\textbf{Con}flicting Evidence for \textbf{Fact}-Checking) (Dataset available at https://github.com/zoeyyes/CONFACT), a novel dataset comprising questions paired with conflicting information from various sources. Extensive experiments reveal critical vulnerabilities in state-of-the-art RAG methods, particularly in resolving conflicts stemming from differences in media source credibility. To address these challenges, we investigate strategies to integrate media background information into both the retrieval and generation stages. Our results show that effectively incorporating source credibility significantly enhances the ability of RAG models to resolve conflicting evidence and improve fact-checking performance.

Wei Jie Yeo, Nirmalendu Prakash, Clement Neo et al.

Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and addresses key research questions such as investigating upstream-downstream latent relationship and understanding the mechanisms of adversarial jailbreaking techniques. We also establish the usefulness of refusal features in enhancing generalization for linear probes to out-of-distribution adversarial samples in classification tasks. We open source our code in https://github.com/wj210/refusal_sae.

Soorya Ram Shimgekar, Abhay Goyal, Lam Yin Cheung et al.

Conspiratorial discourse is increasingly embedded within digital communication ecosystems, yet its structure and spread remain difficult to study. This work analyzes conspiratorial narratives in Singapore-based Telegram groups, showing that such content is woven into everyday discussions rather than confined to isolated echo chambers. We propose a two-stage computational framework. First, we fine-tune RoBERTa-large to classify messages as conspiratorial or not, achieving an F1-score of 0.866 on 2,000 expert-labeled messages. Second, we build a signed belief graph in which nodes represent messages and edge signs reflect alignment in belief labels, weighted by textual similarity. We introduce a Signed Belief Graph Neural Network (SiBeGNN) that uses a Sign Disentanglement Loss to learn embeddings that separate ideological alignment from stylistic features. Using hierarchical clustering on these embeddings, we identify seven narrative archetypes across 553,648 messages: legal topics, medical concerns, media discussions, finance, contradictions in authority, group moderation, and general chat. SiBeGNN yields stronger clustering quality (cDBI = 8.38) than baseline methods (13.60 to 67.27), supported by 88 percent inter-rater agreement in expert evaluations. Our analysis shows that conspiratorial messages appear not only in clusters focused on skepticism or distrust, but also within routine discussions of finance, law, and everyday matters. These findings challenge common assumptions about online radicalization by demonstrating that conspiratorial discourse operates within ordinary social interaction. The proposed framework advances computational methods for belief-driven discourse analysis and offers applications for stance detection, political communication studies, and content moderation policy.

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.

Lei Wang, Wanyu Xu, Zhiqiang Hu et al.

This paper introduces a new in-context learning (ICL) mechanism called In-Image Learning (I$^2$L) that combines demonstration examples, visual cues, and chain-of-thought reasoning into an aggregated image to enhance the capabilities of Large Multimodal Models (e.g., GPT-4V) in multimodal reasoning tasks. Unlike previous approaches that rely on converting images to text or incorporating visual input into language models, I$^2$L consolidates all information into an aggregated image and leverages image processing, understanding, and reasoning abilities. This has several advantages: it reduces inaccurate textual descriptions of complex images, provides flexibility in positioning demonstration examples, and avoids multiple input images and lengthy prompts. We also introduce I$^2$L-Hybrid, a method that combines the strengths of I$^2$L with other ICL methods. Specifically, it uses an automatic strategy to select the most suitable method (I$^2$L or another certain ICL method) for a specific task instance. We conduct extensive experiments to assess the effectiveness of I$^2$L and I$^2$L-Hybrid on MathVista, which covers a variety of complex multimodal reasoning tasks. Additionally, we investigate the influence of image resolution, the number of demonstration examples in a single image, and the positions of these demonstrations in the aggregated image on the effectiveness of I$^2$L. Our code is publicly available at https://github.com/AGI-Edgerunners/IIL.

Ming Shan Hee, Aditi Kumaresan, Nguyen Khoi Hoang et al.

The rise of social media platforms has brought about a new digital culture called memes. Memes, which combine visuals and text, can strongly influence public opinions on social and cultural issues. As a result, people have become interested in categorizing memes, leading to the development of various datasets and multimodal models that show promising results in this field. However, there is currently a lack of a single library that allows for the reproduction, evaluation, and comparison of these models using fair benchmarks and settings. To fill this gap, we introduce the Meme Analytical Tool Kit (MATK), an open-source toolkit specifically designed to support existing memes datasets and cutting-edge multimodal models. MATK aims to assist researchers and engineers in training and reproducing these multimodal models for meme classification tasks, while also providing analysis techniques to gain insights into their strengths and weaknesses. To access MATK, please visit \url{https://github.com/Social-AI-Studio/MATK}.

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.

Leanne Tan, Gabriel Chua, Ziyu Ge et al.

Modern moderation systems increasingly support multiple languages, but often fail to address localisation and low-resource variants - creating safety gaps in real-world deployments. Small models offer a potential alternative to large LLMs, yet still demand considerable data and compute. We present LionGuard 2, a lightweight, multilingual moderation classifier tailored to the Singapore context, supporting English, Chinese, Malay, and partial Tamil. Built on pre-trained OpenAI embeddings and a multi-head ordinal classifier, LionGuard 2 outperforms several commercial and open-source systems across 17 benchmarks, including both Singapore-specific and public English datasets. The system is actively deployed within the Singapore Government, demonstrating practical efficacy at scale. Our findings show that high-quality local data and robust multilingual embeddings can achieve strong moderation performance, without fine-tuning large models. We release our model weights and part of our training data to support future work on LLM safety.

Yao Xiao, Lei Wang, Yue Deng et al.

Reinforcement Learning with Verifiable Rewards~(RLVR) has become a prominent paradigm to enhance the capabilities (i.e.\ long-context) of Large Language Models~(LLMs). However, it often relies on gold-standard answers or explicit evaluation rubrics provided by powerful teacher models or human experts, which are costly and time-consuming. In this work, we investigate unsupervised approaches to enhance the long-context capabilities of LLMs, eliminating the need for heavy human annotations or teacher models' supervision. Specifically, we first replace a few paragraphs with special placeholders in a long document. LLMs are trained through reinforcement learning to reconstruct the document by correctly identifying and sequencing missing paragraphs from a set of candidate options. This training paradigm enables the model to capture global narrative coherence, significantly boosting long-context performance. We validate the effectiveness of our method on two widely used benchmarks, RULER and LongBench~v2. While acquiring noticeable gains on RULER, it can also achieve a reasonable improvement on LongBench~v2 without any manually curated long-context QA data. Furthermore, we conduct extensive ablation studies to analyze the impact of reward design, data curation strategies, training schemes, and data scaling effects on model performance. We publicly release our code, data, and models.

Yujia Hu, Ming Shan Hee, Preslav Nakov et al.

The advancement of Large Language Models (LLMs) has transformed natural language processing; however, their safety mechanisms remain under-explored in low-resource, multilingual settings. Here, we aim to bridge this gap. In particular, we introduce \textsf{SGToxicGuard}, a novel dataset and evaluation framework for benchmarking LLM safety in Singapore's diverse linguistic context, including Singlish, Chinese, Malay, and Tamil. SGToxicGuard adopts a red-teaming approach to systematically probe LLM vulnerabilities in three real-world scenarios: \textit{conversation}, \textit{question-answering}, and \textit{content composition}. We conduct extensive experiments with state-of-the-art multilingual LLMs, and the results uncover critical gaps in their safety guardrails. By offering actionable insights into cultural sensitivity and toxicity mitigation, we lay the foundation for safer and more inclusive AI systems in linguistically diverse environments.\footnote{Link to the dataset: https://github.com/Social-AI-Studio/SGToxicGuard.} \textcolor{red}{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}

Han Wang, Zhuoran Wang, Roy Ka-Wei Lee

Detecting hate speech in videos remains challenging due to the complexity of multimodal content and the lack of fine-grained annotations in existing datasets. We present HateClipSeg, a large-scale multimodal dataset with both video-level and segment-level annotations, comprising over 11,714 segments labeled as Normal or across five Offensive categories: Hateful, Insulting, Sexual, Violence, Self-Harm, along with explicit target victim labels. Our three-stage annotation process yields high inter-annotator agreement (Krippendorff's alpha = 0.817). We propose three tasks to benchmark performance: (1) Trimmed Hateful Video Classification, (2) Temporal Hateful Video Localization, and (3) Online Hateful Video Classification. Results highlight substantial gaps in current models, emphasizing the need for more sophisticated multimodal and temporally aware approaches. The HateClipSeg dataset are publicly available at https://github.com/Social-AI-Studio/HateClipSeg.git.

Junhua Liu, Roy Ka-Wei Lee, Kwan Hui Lim

Human decision-making in high-stakes domains often relies on expertise and heuristics, but is vulnerable to hard-to-detect cognitive biases that threaten fairness and long-term outcomes. This work presents a novel approach to enhancing complex decision-making workflows through the integration of hierarchical learning alongside various enhancements. Focusing on university admissions as a representative high-stakes domain, we propose BGM-HAN, an enhanced Byte-Pair Encoded, Gated Multi-head Hierarchical Attention Network, designed to effectively model semi-structured applicant data. BGM-HAN captures multi-level representations that are crucial for nuanced assessment, improving both interpretability and predictive performance. Experimental results on real admissions data demonstrate that our proposed model significantly outperforms both state-of-the-art baselines from traditional machine learning to large language models, offering a promising framework for augmenting decision-making in domains where structure, context, and fairness matter. Source code is available at: https://github.com/junhua/bgm-han.

Gabriel Chua, Leanne Tan, Ziyu Ge et al.

Large language models (LLMs) and their safety classifiers often perform poorly on low-resource languages due to limited training data and evaluation benchmarks. This paper introduces RabakBench, a new multilingual safety benchmark localized to Singapore's unique linguistic context, covering Singlish, Chinese, Malay, and Tamil. RabakBench is constructed through a scalable three-stage pipeline: (i) Generate - adversarial example generation by augmenting real Singlish web content with LLM-driven red teaming; (ii) Label - semi-automated multi-label safety annotation using majority-voted LLM labelers aligned with human judgments; and (iii) Translate - high-fidelity translation preserving linguistic nuance and toxicity across languages. The final dataset comprises over 5,000 safety-labeled examples across four languages and six fine-grained safety categories with severity levels. Evaluations of 11 popular open-source and closed-source guardrail classifiers reveal significant performance degradation. RabakBench not only enables robust safety evaluation in Southeast Asian multilingual settings but also offers a reproducible framework for building localized safety datasets in low-resource environments. The benchmark dataset, including the human-verified translations, and evaluation code are publicly available.

Wenhao Shi, Zhiqiang Hu, Yi Bin et al.

Large language models (LLMs) have demonstrated impressive reasoning capabilities, particularly in textual mathematical problem-solving. However, existing open-source image instruction fine-tuning datasets, containing limited question-answer pairs per image, do not fully exploit visual information to enhance the multimodal mathematical reasoning capabilities of Multimodal LLMs (MLLMs). To bridge this gap, we address the lack of high-quality, diverse multimodal mathematical datasets by collecting 40K high-quality images with question-answer pairs from 24 existing datasets and synthesizing 320K new pairs, creating the MathV360K dataset, which enhances both the breadth and depth of multimodal mathematical questions. We introduce Math-LLaVA, a LLaVA-1.5-based model fine-tuned with MathV360K. This novel approach significantly improves the multimodal mathematical reasoning capabilities of LLaVA-1.5, achieving a 19-point increase and comparable performance to GPT-4V on MathVista's minitest split, and yielding leading performance on Math-V and MathVerse. Furthermore, Math-LLaVA demonstrates enhanced generalizability, showing substantial improvements on the MMMU benchmark. Our research highlights the importance of dataset diversity and synthesis in advancing MLLMs' mathematical reasoning abilities. The code and data are available at: \url{https://github.com/HZQ950419/Math-LLaVA}.

Han Wang, Ming Shan Hee, Md Rabiul Awal et al.

Recent research has focused on using large language models (LLMs) to generate explanations for hate speech through fine-tuning or prompting. Despite the growing interest in this area, these generated explanations' effectiveness and potential limitations remain poorly understood. A key concern is that these explanations, generated by LLMs, may lead to erroneous judgments about the nature of flagged content by both users and content moderators. For instance, an LLM-generated explanation might inaccurately convince a content moderator that a benign piece of content is hateful. In light of this, we propose an analytical framework for examining hate speech explanations and conducted an extensive survey on evaluating such explanations. Specifically, we prompted GPT-3 to generate explanations for both hateful and non-hateful content, and a survey was conducted with 2,400 unique respondents to evaluate the generated explanations. Our findings reveal that (1) human evaluators rated the GPT-generated explanations as high quality in terms of linguistic fluency, informativeness, persuasiveness, and logical soundness, (2) the persuasive nature of these explanations, however, varied depending on the prompting strategy employed, and (3) this persuasiveness may result in incorrect judgments about the hatefulness of the content. Our study underscores the need for caution in applying LLM-generated explanations for content moderation. Code and results are available at https://github.com/Social-AI-Studio/GPT3-HateEval.

Ming Shan Hee, Wen-Haw Chong, Roy Ka-Wei Lee

Recent studies have proposed models that yielded promising performance for the hateful meme classification task. Nevertheless, these proposed models do not generate interpretable explanations that uncover the underlying meaning and support the classification output. A major reason for the lack of explainable hateful meme methods is the absence of a hateful meme dataset that contains ground truth explanations for benchmarking or training. Intuitively, having such explanations can educate and assist content moderators in interpreting and removing flagged hateful memes. This paper address this research gap by introducing Hateful meme with Reasons Dataset (HatReD), which is a new multimodal hateful meme dataset annotated with the underlying hateful contextual reasons. We also define a new conditional generation task that aims to automatically generate underlying reasons to explain hateful memes and establish the baseline performance of state-of-the-art pre-trained language models on this task. We further demonstrate the usefulness of HatReD by analyzing the challenges of the new conditional generation task in explaining memes in seen and unseen domains. The dataset and benchmark models are made available here: https://github.com/Social-AI-Studio/HatRed

Lei Wang, Wanyu Xu, Yihuai Lan et al.

Large language models (LLMs) have recently been shown to deliver impressive performance in various NLP tasks. To tackle multi-step reasoning tasks, few-shot chain-of-thought (CoT) prompting includes a few manually crafted step-by-step reasoning demonstrations which enable LLMs to explicitly generate reasoning steps and improve their reasoning task accuracy. To eliminate the manual effort, Zero-shot-CoT concatenates the target problem statement with "Let's think step by step" as an input prompt to LLMs. Despite the success of Zero-shot-CoT, it still suffers from three pitfalls: calculation errors, missing-step errors, and semantic misunderstanding errors. To address the missing-step errors, we propose Plan-and-Solve (PS) Prompting. It consists of two components: first, devising a plan to divide the entire task into smaller subtasks, and then carrying out the subtasks according to the plan. To address the calculation errors and improve the quality of generated reasoning steps, we extend PS prompting with more detailed instructions and derive PS+ prompting. We evaluate our proposed prompting strategy on ten datasets across three reasoning problems. The experimental results over GPT-3 show that our proposed zero-shot prompting consistently outperforms Zero-shot-CoT across all datasets by a large margin, is comparable to or exceeds Zero-shot-Program-of-Thought Prompting, and has comparable performance with 8-shot CoT prompting on the math reasoning problem. The code can be found at https://github.com/AGI-Edgerunners/Plan-and-Solve-Prompting.

Han Wang, Nirmalendu Prakash, Nguyen Khoi Hoang et al.

Topic modeling is a widely used technique for revealing underlying thematic structures within textual data. However, existing models have certain limitations, particularly when dealing with short text datasets that lack co-occurring words. Moreover, these models often neglect sentence-level semantics, focusing primarily on token-level semantics. In this paper, we propose PromptTopic, a novel topic modeling approach that harnesses the advanced language understanding of large language models (LLMs) to address these challenges. It involves extracting topics at the sentence level from individual documents, then aggregating and condensing these topics into a predefined quantity, ultimately providing coherent topics for texts of varying lengths. This approach eliminates the need for manual parameter tuning and improves the quality of extracted topics. We benchmark PromptTopic against the state-of-the-art baselines on three vastly diverse datasets, establishing its proficiency in discovering meaningful topics. Furthermore, qualitative analysis showcases PromptTopic's ability to uncover relevant topics in multiple datasets.

Yujia Hu, Roy Ka-Wei Lee

Hateful speech detection is a key component of content moderation, yet current evaluation frameworks rarely assess why a text is deemed hateful. We introduce \textsf{HateXScore}, a four-component metric suite designed to evaluate the reasoning quality of model explanations. It assesses (i) conclusion explicitness, (ii) faithfulness and causal grounding of quoted spans, (iii) protected group identification (policy-configurable), and (iv) logical consistency among these elements. Evaluated on six diverse hate speech datasets, \textsf{HateXScore} is intended as a diagnostic complement to reveal interpretability failures and annotation inconsistencies that are invisible to standard metrics like Accuracy or F1. Moreover, human evaluation shows strong agreement with \textsf{HateXScore}, validating it as a practical tool for trustworthy and transparent moderation. \textcolor{red}{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}

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.