Authors: Yeyubei Zhang, Zhongyan Wang, Zhanyi Ding, Yexin Tian, Jianglai Dai, Xiaorui Shen, Yunchong Liu, Yuchen Cao
Abstract: Social media has become an important source for understanding mental health, providing researchers with a way to detect conditions like depression from user-generated posts. This tutorial provides practical guidance to address common challenges in applying machine learning and deep learning methods for mental health detection on these platforms. It focuses on strategies for working with diverse datasets, improving text preprocessing, and addressing issues such as imbalanced data and model evaluation. Real-world examples and step-by-step instructions demonstrate how to apply these techniques effectively, with an emphasis on transparency, reproducibility, and ethical considerations. By sharing these approaches, this tutorial aims to help researchers build more reliable and widely applicable models for mental health research, contributing to better tools for early detection and intervention.
Authors: Yehan Yan, Tianhao Ma, Ruotai Li, Xinhan Zheng, Guodong Shan, Chisheng Li
Abstract: Traditional Chinese medicine (TCM) plays a vital role in health protection and disease treatment, but its practical application requires extensive medical knowledge and clinical experience. Existing TCM Large Language Models (LLMs) exhibit critical limitations of uncomprehensive medical consultation and diagnoses, and inaccurate syndrome differentiation-based treatment. To address these issues, this study establishes JingFang (JF): a novel TCM Large Language Model that demonstrates the expert-level capability of medical diagnosis and syndrome differentiation-based treatment. We innovate a Multi-agent Dynamic Collaborative Chain-of-Thought Mechanism (MDCCTM) for medical consultation, enabling JF with effective and accurate diagnostic ability. In addition, a Syndrome Agent and a Dual-Stage Retrieval Scheme (DSRS) are developed to significantly enhance the capacity of JF for disease treatment based on syndrome differentiation. JingFang not only facilitates the application of LLMs but also promotes the effective practice of TCM in human health protection and disease treatment.
Authors: Saydul Akbar Murad, Ashim Dahal, Nick Rahimi
Abstract: Cyber threat detection has become an important area of focus in today's digital age due to the growing spread of fake information and harmful content on social media platforms such as Twitter (now 'X'). These cyber threats, often disguised within tweets, pose significant risks to individuals, communities, and even nations, emphasizing the need for effective detection systems. While previous research has explored tweet-based threats, much of the work is limited to specific languages, domains, or locations, or relies on single-model approaches, reducing their applicability to diverse real-world scenarios. To address these gaps, our study focuses on multi-lingual tweet cyber threat detection using a variety of advanced models. The research was conducted in three stages: (1) We collected and labeled tweet datasets in four languages English, Chinese, Russian, and Arabic employing both manual and polarity-based labeling methods to ensure high-quality annotations. (2) Each dataset was analyzed individually using machine learning (ML) and deep learning (DL) models to assess their performance on distinct languages. (3) Finally, we combined all four datasets into a single multi-lingual dataset and applied DL and large language model (LLM) architectures to evaluate their efficacy in identifying cyber threats across various languages. Our results show that among machine learning models, Random Forest (RF) attained the highest performance; however, the Bi-LSTM architecture consistently surpassed other DL and LLM architectures across all datasets. These findings underline the effectiveness of Bi-LSTM in multilingual cyber threat detection. The code for this paper can be found at this link: https://github.com/Mmurrad/Tweet-Data-Classification.git.
URLs: https://github.com/Mmurrad/Tweet-Data-Classification.git.
Authors: Zongwei Li, Xiaoqi Li, Wenkai Li, Xin Wang
Abstract: As the Ethereum platform continues to mature and gain widespread usage, it is crucial to maintain high standards of smart contract writing practices. While bad practices in smart contracts may not directly lead to security issues, they do elevate the risk of encountering problems. Therefore, to understand and avoid these bad practices, this paper introduces the first systematic study of bad practices in smart contracts, delving into over 35 specific issues. Specifically, we propose a large language models (LLMs)-based framework, SCALM. It combines Step-Back Prompting and Retrieval-Augmented Generation (RAG) to identify and address various bad practices effectively. Our extensive experiments using multiple LLMs and datasets have shown that SCALM outperforms existing tools in detecting bad practices in smart contracts.
Authors: Juyun Wee, Minjae Park, Jaeho Lee
Abstract: Depth pruning aims to reduce the inference cost of a large language model without any hardware-specific complications, by simply removing several less important transformer blocks. However, our empirical findings suggest that the importance of a transformer block may be highly task-dependent -- a block that is crucial for a task can be removed without degrading the accuracy on another task. Based on this observation, we develop a dynamic depth pruning algorithm, coined PuDDing (Prompt-routed Dynamic Depth Pruning), which determines which blocks to omit from the model based on the input prompt. PuDDing operates by training a lightweight router to predict the best omission set among a set of options, where this option set has also been constructed in a data-driven manner. Empirical results on commonsense reasoning benchmarks demonstrate that PuDDing effectively accelerates the inference language models, and achieves better on-task performance than static depth pruning baselines.
Authors: Pietro Alessandro Aluffi, Patrick Zietkiewicz, Marya Bazzi, Matt Arderne, Vladimirs Murevics
Abstract: The rapid evolution of large language models (LLMs) has transformed conversational agents, enabling complex human-machine interactions. However, evaluation frameworks often focus on single tasks, failing to capture the dynamic nature of multi-turn dialogues. This paper introduces a dynamic benchmarking framework to assess LLM-based conversational agents through interactions with synthetic users. The framework integrates generative agent simulation to evaluate performance on key dimensions: information extraction, context awareness, and adaptive engagement. By simulating various aspects of user behavior, our work provides a scalable, automated, and flexible benchmarking approach. Experimental evaluation - within a loan application use case - demonstrates the framework's effectiveness under one-shot and few-shot extraction conditions. Results show that adaptive strategies improve data extraction accuracy, especially when handling ambiguous responses. Future work will extend its applicability to broader domains and incorporate additional metrics (e.g., conversational coherence, user engagement). This study contributes a structured, scalable approach to evaluating LLM-based conversational agents, facilitating real-world deployment.
Authors: Yongchao Chen, Yilun Hao, Yueying Liu, Yang Zhang, Chuchu Fan
Abstract: Existing methods fail to effectively steer Large Language Models (LLMs) between textual reasoning and code generation, leaving symbolic computing capabilities underutilized. We introduce CodeSteer, an effective method for guiding LLM code/text generation. We construct a comprehensive benchmark SymBench comprising 37 symbolic tasks with adjustable complexity and also synthesize datasets of 12k multi-round guidance/generation trajectories and 5.5k guidance comparison pairs. We fine-tune the Llama-3-8B model with a newly designed multi-round supervised fine-tuning (SFT) and direct preference optimization (DPO). The resulting model, CodeSteerLLM, augmented with the proposed symbolic and self-answer checkers, effectively guides the code/text generation of larger models. Augmenting GPT-4o with CodeSteer raises its average performance score from 53.3 to 86.4, even outperforming the existing best LLM OpenAI o1 (82.7), o1-preview (74.8), and DeepSeek R1 (76.8) across all 37 tasks (28 seen, 9 unseen). Trained for GPT-4o, CodeSteer demonstrates superior generalizability, providing an average 41.8 performance boost on Claude, Mistral, and GPT-3.5. CodeSteer-guided LLMs fully harness symbolic computing to maintain strong performance on highly complex tasks. Models, Datasets, and Codes are available at https://github.com/yongchao98/CodeSteer-v1.0.
Authors: Torsten Hiltmann, Martin Dr\"oge, Nicole Dresselhaus, Till Grallert, Melanie Althage, Paul Bayer, Sophie Eckenstaler, Koray Mendi, Jascha Marijn Schmitz, Philipp Schneider, Wiebke Sczeponik, Anica Skibba
Abstract: Named entity recognition (NER) is a core task for historical research in automatically establishing all references to people, places, events and the like. Yet, do to the high linguistic and genre diversity of sources, only limited canonisation of spellings, the level of required historical domain knowledge, and the scarcity of annotated training data, established approaches to natural language processing (NLP) have been both extremely expensive and yielded only unsatisfactory results in terms of recall and precision. Our paper introduces a new approach. We demonstrate how readily-available, state-of-the-art LLMs significantly outperform two leading NLP frameworks, spaCy and flair, for NER in historical documents by seven to twentytwo percent higher F1-Scores. Our ablation study shows how providing historical context to the task and a bit of persona modelling that turns focus away from a purely linguistic approach are core to a successful prompting strategy. We also demonstrate that, contrary to our expectations, providing increasing numbers of examples in few-shot approaches does not improve recall or precision below a threshold of 16-shot. In consequence, our approach democratises access to NER for all historians by removing the barrier of scripting languages and computational skills required for established NLP tools and instead leveraging natural language prompts and consumer-grade tools and frontends.
Authors: Fabian Hoppe, Filip Ilievski, Jan-Christoph Kalo
Abstract: Large Language Models (LLMs) have been shown to achieve impressive results for many reasoning-based Natural Language Processing (NLP) tasks, suggesting a degree of deductive reasoning capability. However, it remains unclear to which extent LLMs, in both informal and autoformalisation methods, are robust on logical deduction tasks. Moreover, while many LLM-based deduction methods have been proposed, there is a lack of a systematic study that analyses the impact of their design components. Addressing these two challenges, we propose the first study of the robustness of LLM-based deductive reasoning methods. We devise a framework with two families of perturbations: adversarial noise and counterfactual statements, which jointly generate seven perturbed datasets. We organize the landscape of LLM reasoners according to their reasoning format, formalisation syntax, and feedback for error recovery. The results show that adversarial noise affects autoformalisation, while counterfactual statements influence all approaches. Detailed feedback does not improve overall accuracy despite reducing syntax errors, pointing to the challenge of LLM-based methods to self-correct effectively.
Authors: Afshin Khadangi, Amir Sartipi, Igor Tchappi, Gilbert Fridgen
Abstract: Art, as a universal language, can be interpreted in diverse ways, with artworks embodying profound meanings and nuances. The advent of Large Language Models (LLMs) and the availability of Multimodal Large Language Models (MLLMs) raise the question of how these transformative models can be used to assess and interpret the artistic elements of artworks. While research has been conducted in this domain, to the best of our knowledge, a deep and detailed understanding of the technical and expressive features of artworks using LLMs has not been explored. In this study, we investigate the automation of a formal art analysis framework to analyze a high-throughput number of artworks rapidly and examine how their patterns evolve over time. We explore how LLMs can decode artistic expressions, visual elements, composition, and techniques, revealing emerging patterns that develop across periods. Finally, we discuss the strengths and limitations of LLMs in this context, emphasizing their ability to process vast quantities of art-related data and generate insightful interpretations. Due to the exhaustive and granular nature of the results, we have developed interactive data visualizations, available online https://cognartive.github.io/, to enhance understanding and accessibility.
Authors: Yunyi Shen, Hao Sun, Jean-Fran\c{c}ois Ton
Abstract: Building neural reward models from human preferences is a pivotal component in reinforcement learning from human feedback (RLHF) and large language model alignment research. Given the scarcity and high cost of human annotation, how to select the most informative pairs to annotate is an essential yet challenging open problem. In this work, we highlight the insight that an ideal comparison dataset for reward modeling should balance exploration of the representation space and make informative comparisons between pairs with moderate reward differences. Technically, challenges arise in quantifying the two objectives and efficiently prioritizing the comparisons to be annotated. To address this, we propose the Fisher information-based selection strategies, adapt theories from the classical experimental design literature, and apply them to the final linear layer of the deep neural network-based reward modeling tasks. Empirically, our method demonstrates remarkable performance, high computational efficiency, and stability compared to other selection methods from deep learning and classical statistical literature across multiple open-source LLMs and datasets. Further ablation studies reveal that incorporating cross-prompt comparisons in active reward modeling significantly enhances labeling efficiency, shedding light on the potential for improved annotation strategies in RLHF.
Authors: Md Sifat Hossain, Anika Tabassum, Md. Fahim Arefin, Tarannum Shaila Zaman
Abstract: The rapid advancement of large language models has opened new avenues for automating complex problem-solving tasks such as algorithmic coding and competitive programming. This paper introduces a novel evaluation technique, LLM-ProS, to assess the performance of state-of-the-art LLMs on International Collegiate Programming Contest (ICPC) problems. Using a curated dataset of 166 World Finals problems from 2011 to 2024, we benchmark the models' reasoning, accuracy, and efficiency. We evaluate the five models-GPT-4o, Mistral Large, Llama-3.1-405B, and the o1 family, consisting of o1-mini and o1-preview, across critical metrics like correctness, resource utilization, and response calibration. Our results reveal significant differences in the models' abilities to generalize, adapt, and solve novel problems. We also investigated the impact of training methodologies, dataset contamination, and chain-of-thought reasoning on model performance. The findings provide new insights into optimizing LLMs for algorithmic tasks, highlighting both strengths and limitations of current models.
Authors: Mahdi Alkaeed, Sofiat Abioye, Adnan Qayyum, Yosra Magdi Mekki, Ilhem Berrou, Mohamad Abdallah, Ala Al-Fuqaha, Muhammad Bilal, Junaid Qadir
Abstract: In response to the success of proprietary Large Language Models (LLMs) such as OpenAI's GPT-4, there is a growing interest in developing open, non-proprietary LLMs and AI foundation models (AIFMs) for transparent use in academic, scientific, and non-commercial applications. Despite their inability to match the refined functionalities of their proprietary counterparts, open models hold immense potential to revolutionize healthcare applications. In this paper, we examine the prospects of open-source LLMs and AIFMs for developing healthcare applications and make two key contributions. Firstly, we present a comprehensive survey of the current state-of-the-art open-source healthcare LLMs and AIFMs and introduce a taxonomy of these open AIFMs, categorizing their utility across various healthcare tasks. Secondly, to evaluate the general-purpose applications of open LLMs in healthcare, we present a case study on personalized prescriptions. This task is particularly significant due to its critical role in delivering tailored, patient-specific medications that can greatly improve treatment outcomes. In addition, we compare the performance of open-source models with proprietary models in settings with and without Retrieval-Augmented Generation (RAG). Our findings suggest that, although less refined, open LLMs can achieve performance comparable to proprietary models when paired with grounding techniques such as RAG. Furthermore, to highlight the clinical significance of LLMs-empowered personalized prescriptions, we perform subjective assessment through an expert clinician. We also elaborate on ethical considerations and potential risks associated with the misuse of powerful LLMs and AIFMs, highlighting the need for a cautious and responsible implementation in healthcare.
Authors: Hao Sun, Yunyi Shen, Jean-Francois Ton, Mihaela van der Schaar
Abstract: Large Language Models (LLMs) have made substantial strides in structured tasks through Reinforcement Learning (RL), demonstrating proficiency in mathematical reasoning and code generation. However, applying RL in broader domains like chatbots and content generation -- through the process known as Reinforcement Learning from Human Feedback (RLHF) -- presents unique challenges. Reward models in RLHF are critical, acting as proxies that evaluate the alignment of LLM outputs with human intent. Despite advancements, the development of reward models is hindered by challenges such as computational heavy training, costly evaluation, and therefore poor reproducibility. We advocate for using embedding-based input in reward model research as an accelerated solution to those challenges. By leveraging embeddings for reward modeling, we can enhance reproducibility, reduce computational demands on hardware, improve training stability, and significantly reduce training and evaluation costs, hence facilitating fair and efficient comparisons in this active research area. We then show a case study of reproducing existing reward model ensemble research using embedding-based reward models. We discussed future avenues for research, aiming to contribute to safer and more effective LLM deployments.
Authors: Elliot Meyerson, Xin Qiu
Abstract: Decomposing hard problems into subproblems often makes them easier and more efficient to solve. With large language models (LLMs) crossing critical reliability thresholds for a growing slate of capabilities, there is an increasing effort to decompose systems into sets of LLM-based agents, each of whom can be delegated sub-tasks. However, this decomposition (even when automated) is often intuitive, e.g., based on how a human might assign roles to members of a human team. How close are these role decompositions to optimal? This position paper argues that asymptotic analysis with LLM primitives is needed to reason about the efficiency of such decomposed systems, and that insights from such analysis will unlock opportunities for scaling them. By treating the LLM forward pass as the atomic unit of computational cost, one can separate out the (often opaque) inner workings of a particular LLM from the inherent efficiency of how a set of LLMs are orchestrated to solve hard problems. In other words, if we want to scale the deployment of LLMs to the limit, instead of anthropomorphizing LLMs, asymptotic analysis with LLM primitives should be used to reason about and develop more powerful decompositions of large problems into LLM agents.
Authors: Akshar Tumu, Parisa Kordjamshidi
Abstract: Spatial Reasoning is an important component of human cognition and is an area in which the latest Vision-language models (VLMs) show signs of difficulty. The current analysis works use image captioning tasks and visual question answering. In this work, we propose using the Referring Expression Comprehension task instead as a platform for the evaluation of spatial reasoning by VLMs. This platform provides the opportunity for a deeper analysis of spatial comprehension and grounding abilities when there is 1) ambiguity in object detection, 2) complex spatial expressions with a longer sentence structure and multiple spatial relations, and 3) expressions with negation ('not'). In our analysis, we use task-specific architectures as well as large VLMs and highlight their strengths and weaknesses in dealing with these specific situations. While all these models face challenges with the task at hand, the relative behaviors depend on the underlying models and the specific categories of spatial semantics (topological, directional, proximal, etc.). Our results highlight these challenges and behaviors and provide insight into research gaps and future directions.
Authors: Xiao Hu, Eric Liu, Weizhou Wang, Xiangyu Guo, David Lie
Abstract: Retrieval-Augmented Generation (RAG) offers a solution to mitigate hallucinations in Large Language Models (LLMs) by grounding their outputs to knowledge retrieved from external sources. The use of private resources and data in constructing these external data stores can expose them to risks of extraction attacks, in which attackers attempt to steal data from these private databases. Existing RAG extraction attacks often rely on manually crafted prompts, which limit their effectiveness. In this paper, we introduce a framework called MARAGE for optimizing an adversarial string that, when appended to user queries submitted to a target RAG system, causes outputs containing the retrieved RAG data verbatim. MARAGE leverages a continuous optimization scheme that integrates gradients from multiple models with different architectures simultaneously to enhance the transferability of the optimized string to unseen models. Additionally, we propose a strategy that emphasizes the initial tokens in the target RAG data, further improving the attack's generalizability. Evaluations show that MARAGE consistently outperforms both manual and optimization-based baselines across multiple LLMs and RAG datasets, while maintaining robust transferability to previously unseen models. Moreover, we conduct probing tasks to shed light on the reasons why MARAGE is more effective compared to the baselines and to analyze the impact of our approach on the model's internal state.
Authors: Yerin Hwang, Yongil Kim, Jahyun Koo, Taegwan Kang, Hyunkyung Bae, Kyomin Jung
Abstract: Despite the fact that large language models (LLMs) show exceptional skill in instruction following tasks, this strength can turn into a vulnerability when the models are required to disregard certain instructions. Instruction-following tasks typically involve a clear task description and input text containing the target data to be processed. However, when the input itself resembles an instruction, confusion may arise, even if there is explicit prompting to distinguish between the task instruction and the input. We refer to this phenomenon as instructional distraction. In this paper, we introduce a novel benchmark, named DIM-Bench, specifically designed to assess LLMs' performance under instructional distraction. The benchmark categorizes real-world instances of instructional distraction and evaluates LLMs across four instruction tasks: rewriting, proofreading, translation, and style transfer -- alongside five input tasks: reasoning, code generation, mathematical reasoning, bias detection, and question answering. Our experimental results reveal that even the most advanced LLMs are susceptible to instructional distraction, often failing to accurately follow user intent in such cases.
Authors: Daniel Wai Kit Chin, Roy Ka-Wei Lee
Abstract: The widespread use of social media has accelerated the dissemination of information, but it has also facilitated the spread of harmful rumours, which can disrupt economies, influence political outcomes, and exacerbate public health crises, such as the COVID-19 pandemic. While Graph Neural Network (GNN)-based approaches have shown significant promise in automated rumour detection, they often lack transparency, making their predictions difficult to interpret. Existing graph explainability techniques fall short in addressing the unique challenges posed by the dependencies among feature dimensions in high-dimensional text embeddings used in GNN-based models. In this paper, we introduce Contrastive Token Layerwise Relevance Propagation (CT-LRP), a novel framework designed to enhance the explainability of GNN-based rumour detection. CT-LRP extends current graph explainability methods by providing token-level explanations that offer greater granularity and interpretability. We evaluate the effectiveness of CT-LRP across multiple GNN models trained on three publicly available rumour detection datasets, demonstrating that it consistently produces high-fidelity, meaningful explanations, paving the way for more robust and trustworthy rumour detection systems.
Authors: Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, Tat-Seng Chua
Abstract: Text-to-3D generation automates 3D content creation from textual descriptions, which offers transformative potential across various fields. However, existing methods often struggle to align generated content with human preferences, limiting their applicability and flexibility. To address these limitations, in this paper, we propose DreamDPO, an optimization-based framework that integrates human preferences into the 3D generation process, through direct preference optimization. Practically, DreamDPO first constructs pairwise examples, then compare their alignment with human preferences using reward or large multimodal models, and lastly optimizes the 3D representation with a preference-driven loss function. By leveraging pairwise comparison to reflect preferences, DreamDPO reduces reliance on precise pointwise quality evaluations while enabling fine-grained controllability through preference-guided optimization. Experiments demonstrate that DreamDPO achieves competitive results, and provides higher-quality and more controllable 3D content compared to existing methods. The code and models will be open-sourced.
Authors: Juliano Genari, Guilherme Tegoni Goedert
Abstract: The extraction of relevant data from Electronic Health Records (EHRs) is crucial to identifying symptoms and automating epidemiological surveillance processes. By harnessing the vast amount of unstructured text in EHRs, we can detect patterns that indicate the onset of disease outbreaks, enabling faster, more targeted public health responses. Our proposed framework provides a flexible and efficient solution for mining data from unstructured texts, significantly reducing the need for extensive manual labeling by specialists. Experiments show that our framework achieving strong performance with as few as 200 manually labeled texts, even for complex classification problems. Additionally, our approach can function with simple lightweight models, achieving competitive and occasionally even better results compared to more resource-intensive deep learning models. This capability not only accelerates processing times but also preserves patient privacy, as the data can be processed on weaker on-site hardware rather than being transferred to external systems. Our methodology, therefore, offers a practical, scalable, and privacy-conscious approach to real-time epidemiological monitoring, equipping health institutions to respond rapidly and effectively to emerging health threats.
Authors: Junjie Yao, Zhongwang Zhang, Zhi-Qin John Xu
Abstract: Transformer-based Large Language Models (LLMs) have revolutionized Natural Language Processing by demonstrating exceptional performance across diverse tasks. This study investigates the impact of the parameter initialization scale on the training behavior and task preferences of LLMs. We discover that smaller initialization scales encourage models to favor reasoning tasks, whereas larger initialization scales lead to a preference for memorization tasks. We validate this reasoning bias via real datasets and meticulously designed anchor functions. Further analysis of initial training dynamics suggests that specific model components, particularly the embedding space and self-attention mechanisms, play pivotal roles in shaping these learning biases. We provide a theoretical framework from the perspective of model training dynamics to explain these phenomena. Additionally, experiments on real-world language tasks corroborate our theoretical insights. This work enhances our understanding of how initialization strategies influence LLM performance on reasoning tasks and offers valuable guidelines for training models.
Authors: Lingxiang Hu, Shurun Yuan, Xiaoting Qin, Jue Zhang, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, Qi Zhang
Abstract: In contemporary workplaces, meetings are essential for exchanging ideas and ensuring team alignment but often face challenges such as time consumption, scheduling conflicts, and inefficient participation. Recent advancements in Large Language Models (LLMs) have demonstrated their strong capabilities in natural language generation and reasoning, prompting the question: can LLMs effectively delegate participants in meetings? To explore this, we develop a prototype LLM-powered meeting delegate system and create a comprehensive benchmark using real meeting transcripts. Our evaluation reveals that GPT-4/4o maintain balanced performance between active and cautious engagement strategies. In contrast, Gemini 1.5 Pro tends to be more cautious, while Gemini 1.5 Flash and Llama3-8B/70B display more active tendencies. Overall, about 60\% of responses address at least one key point from the ground-truth. However, improvements are needed to reduce irrelevant or repetitive content and enhance tolerance for transcription errors commonly found in real-world settings. Additionally, we implement the system in practical settings and collect real-world feedback from demos. Our findings underscore the potential and challenges of utilizing LLMs as meeting delegates, offering valuable insights into their practical application for alleviating the burden of meetings.
Authors: Zhenqing Ling, Daoyuan Chen, Liuyi Yao, Yaliang Li, Ying Shen
Abstract: Fine-tuning large language models (LLMs) using diverse datasets is crucial for enhancing their overall performance across various domains. In practical scenarios, existing methods based on modeling the mixture proportions of data composition often struggle with data whose domain labels are missing, imprecise or non-normalized, while methods based on data selection usually encounter difficulties in balancing multi-domain performance. To address these challenges, in this paper, we study the role of data diversity in enhancing the overall abilities of LLMs by empirically constructing contrastive data pools and theoretically deriving explanations for both inter- and intra-diversity. Building upon the insights gained, we propose a new method that gives the LLM a dual identity: an output model to cognitively probe and select data based on diversity reward, as well as an input model to be tuned with the selected data. Extensive experiments show that the proposed method notably boosts performance across domain-undetermined data and a series of foundational downstream tasks when applied to various advanced LLMs. We release our code and hope this study can shed light on the understanding of data diversity and advance feedback-driven data-model co-development for LLMs.
Authors: Jonathan Kim, Anna Podlasek, Kie Shidara, Feng Liu, Ahmed Alaa, Danilo Bernardo
Abstract: Large Language Models (LLMs) have attained human-level accuracy on medical question-answer (QA) benchmarks. However, their limitations in navigating open-ended clinical scenarios have recently been shown, raising concerns about the robustness and generalizability of LLM reasoning across diverse, real-world medical tasks. To probe potential LLM failure modes in clinical problem-solving, we present the medical abstraction and reasoning corpus (M-ARC). M-ARC assesses clinical reasoning through scenarios designed to exploit the Einstellung effect -- the fixation of thought arising from prior experience, targeting LLM inductive biases toward inflexible pattern matching from their training data rather than engaging in flexible reasoning. We find that LLMs, including current state-of-the-art o1 and Gemini models, perform poorly compared to physicians on M-ARC, often demonstrating lack of commonsense medical reasoning and a propensity to hallucinate. In addition, uncertainty estimation analyses indicate that LLMs exhibit overconfidence in their answers, despite their limited accuracy. The failure modes revealed by M-ARC in LLM medical reasoning underscore the need to exercise caution when deploying these models in clinical settings.
Authors: Rajiv Movva, Kenny Peng, Nikhil Garg, Jon Kleinberg, Emma Pierson
Abstract: We describe HypotheSAEs, a general method to hypothesize interpretable relationships between text data (e.g., headlines) and a target variable (e.g., clicks). HypotheSAEs has three steps: (1) train a sparse autoencoder on text embeddings to produce interpretable features describing the data distribution, (2) select features that predict the target variable, and (3) generate a natural language interpretation of each feature (e.g., "mentions being surprised or shocked") using an LLM. Each interpretation serves as a hypothesis about what predicts the target variable. Compared to baselines, our method better identifies reference hypotheses on synthetic datasets (at least +0.06 in F1) and produces more predictive hypotheses on real datasets (~twice as many significant findings), despite requiring 1-2 orders of magnitude less compute than recent LLM-based methods. HypotheSAEs also produces novel discoveries on two well-studied tasks: explaining partisan differences in Congressional speeches and identifying drivers of engagement with online headlines.
Authors: Bo Wen, Xin Zhang
Abstract: This paper presents SOLOMON, a novel Neuro-inspired Large Language Model (LLM) Reasoning Network architecture that enhances the adaptability of foundation models for domain-specific applications. Through a case study in semiconductor layout design, we demonstrate how SOLOMON enables swift adaptation of general-purpose LLMs to specialized tasks by leveraging Prompt Engineering and In-Context Learning techniques. Our experiments reveal the challenges LLMs face in spatial reasoning and applying domain knowledge to practical problems. Results show that SOLOMON instances significantly outperform their baseline LLM counterparts and achieve performance comparable to state-of-the-art reasoning model, o1-preview. We discuss future research directions for developing more adaptive AI systems that can continually learn, adapt, and evolve in response to new information and changing requirements.
Authors: Royson Lee, Minyoung Kim, Fady Rezk, Rui Li, Stylianos I. Venieris, Timothy Hospedales
Abstract: Federated learning (FL) has enabled the training of multilingual large language models (LLMs) on diverse and decentralized multilingual data, especially on low-resource languages. To improve client-specific performance, personalization via the use of parameter-efficient fine-tuning (PEFT) modules such as LoRA is common. This involves a personalization strategy (PS), such as the design of the PEFT adapter structures (e.g., in which layers to add LoRAs and what ranks) and choice of hyperparameters (e.g., learning rates) for fine-tuning. Instead of manual PS configuration, we propose FedP$^2$EFT, a federated learning-to-personalize method for multilingual LLMs in cross-device FL settings. Unlike most existing PEFT structure selection methods, which are prone to overfitting low-data regimes, FedP$^2$EFT collaboratively learns the optimal personalized PEFT structure for each client via Bayesian sparse rank selection. Evaluations on both simulated and real-world multilingual FL benchmarks demonstrate that FedP$^2$EFT largely outperforms existing personalized fine-tuning methods, while complementing a range of existing FL methods.
Authors: Simone Clemente, Zied Ben Houidi, Alexis Huet, Dario Rossi, Giulio Franzese, Pietro Michiardi
Abstract: Despite remarkable capabilities, large language models (LLMs) struggle to continually update their knowledge without catastrophic forgetting. In contrast, humans effortlessly integrate new information, detect conflicts with existing beliefs, and selectively update their mental models. This paper introduces a cognitive-inspired investigation paradigm to study continual knowledge updating in LLMs. We implement two key components inspired by human cognition: (1) Dissonance and Familiarity Awareness, analyzing model behavior to classify information as novel, familiar, or dissonant; and (2) Targeted Network Updates, which track neural activity to identify frequently used (stubborn) and rarely used (plastic) neurons. Through carefully designed experiments in controlled settings, we uncover a number of empirical findings demonstrating the potential of this approach. First, dissonance detection is feasible using simple activation and gradient features, suggesting potential for cognitive-inspired training. Second, we find that non-dissonant updates largely preserve prior knowledge regardless of targeting strategy, revealing inherent robustness in LLM knowledge integration. Most critically, we discover that dissonant updates prove catastrophically destructive to the model's knowledge base, indiscriminately affecting even information unrelated to the current updates. This suggests fundamental limitations in how neural networks handle contradictions and motivates the need for new approaches to knowledge updating that better mirror human cognitive mechanisms.
Authors: Chenyang Shao, Xinyuan Hu, Yutang Lin, Fengli Xu
Abstract: The rapid expansion of web content has made on-device AI assistants indispensable for helping users manage the increasing complexity of online tasks. The emergent reasoning ability in large language models offer a promising path for next-generation on-device AI agents. However, deploying full-scale Large Language Models (LLMs) on resource-limited local devices is challenging. In this paper, we propose Division-of-Thoughts (DoT), a collaborative reasoning framework leveraging the synergy between locally deployed Smaller-scale Language Models (SLMs) and cloud-based LLMs. DoT leverages a Task Decomposer to elicit the inherent planning abilities in language models to decompose user queries into smaller sub-tasks, which allows hybrid language models to fully exploit their respective strengths. Besides, DoT employs a Task Scheduler to analyze the pair-wise dependency of sub-tasks and create a dependency graph, facilitating parallel reasoning of sub-tasks and the identification of key steps. To allocate the appropriate model based on the difficulty of sub-tasks, DoT leverages a Plug-and-Play Adapter, which is an additional task head attached to the SLM that does not alter the SLM's parameters. To boost adapter's task allocation capability, we propose a self-reinforced training method that relies solely on task execution feedback. Extensive experiments on various benchmarks demonstrate that our DoT significantly reduces LLM costs while maintaining competitive reasoning accuracy. Specifically, DoT reduces the average reasoning time and API costs by 66.12% and 83.57%, while achieving comparable reasoning accuracy with the best baseline methods.
Authors: Tingyu Mo, Jacqueline C. K. Lam, Victor O. K. Li, Lawrence Y. L. Cheung
Abstract: Alzheimer's Disease (AD) is an irreversible neurodegenerative disease affecting 50 million people worldwide. Low-cost, accurate identification of key markers of AD is crucial for timely diagnosis and intervention. Language impairment is one of the earliest signs of cognitive decline, which can be used to discriminate AD patients from normal control individuals. Patient-interviewer dialogues may be used to detect such impairments, but they are often mixed with ambiguous, noisy, and irrelevant information, making the AD detection task difficult. Moreover, the limited availability of AD speech samples and variability in their speech styles pose significant challenges in developing robust speech-based AD detection models. To address these challenges, we propose DECT, a novel speech-based domain-specific approach leveraging large language models (LLMs) for fine-grained linguistic analysis and label-switched label-preserved data generation. Our study presents four novelties: We harness the summarizing capabilities of LLMs to identify and distill key Cognitive-Linguistic information from noisy speech transcripts, effectively filtering irrelevant information. We leverage the inherent linguistic knowledge of LLMs to extract linguistic markers from unstructured and heterogeneous audio transcripts. We exploit the compositional ability of LLMs to generate AD speech transcripts consisting of diverse linguistic patterns to overcome the speech data scarcity challenge and enhance the robustness of AD detection models. We use the augmented AD textual speech transcript dataset and a more fine-grained representation of AD textual speech transcript data to fine-tune the AD detection model. The results have shown that DECT demonstrates superior model performance with an 11% improvement in AD detection accuracy on the datasets from DementiaBank compared to the baselines.
Authors: Xiaorui Su, Shvat Messica, Yepeng Huang, Ruth Johnson, Lukas Fesser, Shanghua Gao, Faryad Sahneh, Marinka Zitnik
Abstract: Foundation models trained on patient electronic health records (EHRs) require tokenizing medical data into sequences of discrete vocabulary items. Existing tokenizers treat medical codes from EHRs as isolated textual tokens. However, each medical code is defined by its textual description, its position in ontological hierarchies, and its relationships to other codes, such as disease co-occurrences and drug-treatment associations. Medical vocabularies contain more than 600,000 codes with critical information for clinical reasoning. We introduce MedTok, a multimodal medical code tokenizer that uses the text descriptions and relational context of codes. MedTok processes text using a language model encoder and encodes the relational structure with a graph encoder. It then quantizes both modalities into a unified token space, preserving modality-specific and cross-modality information. We integrate MedTok into five EHR models and evaluate it on operational and clinical tasks across in-patient and out-patient datasets, including outcome prediction, diagnosis classification, drug recommendation, and risk stratification. Swapping standard EHR tokenizers with MedTok improves AUPRC across all EHR models, by 4.10% on MIMIC-III, 4.78% on MIMIC-IV, and 11.30% on EHRShot, with the largest gains in drug recommendation. Beyond EHR modeling, we demonstrate using MedTok tokenizer with medical QA systems. Our results demonstrate the potential of MedTok as a unified tokenizer for medical codes, improving tokenization for medical foundation models.
Authors: Xiao-Wen Yang, Xuan-Yi Zhu, Wen-Da Wei, Ding-Chu Zhang, Jie-Jing Shao, Zhi Zhou, Lan-Zhe Guo, Yu-Feng Li
Abstract: The integration of slow-thinking mechanisms into large language models (LLMs) offers a promising way toward achieving Level 2 AGI Reasoners, as exemplified by systems like OpenAI's o1. However, several significant challenges remain, including inefficient overthinking and an overreliance on auxiliary reward models. We point out that these limitations stem from LLMs' inability to internalize the search process, a key component of effective reasoning. A critical step toward addressing this issue is enabling LLMs to autonomously determine when and where to backtrack, a fundamental operation in traditional search algorithms. To this end, we propose a self-backtracking mechanism that equips LLMs with the ability to backtrack during both training and inference. This mechanism not only enhances reasoning ability but also efficiency by transforming slow-thinking processes into fast-thinking through self-improvement. Empirical evaluations demonstrate that our proposal significantly enhances the reasoning capabilities of LLMs, achieving a performance gain of over 40 percent compared to the optimal-path supervised fine-tuning method. We believe this study introduces a novel and promising pathway for developing more advanced and robust Reasoners.
Authors: Xuejiao Zhao, Siyan Liu, Su-Yin Yang, Chunyan Miao
Abstract: Retrieval-augmented generation (RAG) is a well-suited technique for retrieving privacy-sensitive Electronic Health Records (EHR). It can serve as a key module of the healthcare copilot, helping reduce misdiagnosis for healthcare practitioners and patients. However, the diagnostic accuracy and specificity of existing heuristic-based RAG models used in the medical domain are inadequate, particularly for diseases with similar manifestations. This paper proposes MedRAG, a RAG model enhanced by knowledge graph (KG)-elicited reasoning for the medical domain that retrieves diagnosis and treatment recommendations based on manifestations. MedRAG systematically constructs a comprehensive four-tier hierarchical diagnostic KG encompassing critical diagnostic differences of various diseases. These differences are dynamically integrated with similar EHRs retrieved from an EHR database, and reasoned within a large language model. This process enables more accurate and specific decision support, while also proactively providing follow-up questions to enhance personalized medical decision-making. MedRAG is evaluated on both a public dataset DDXPlus and a private chronic pain diagnostic dataset (CPDD) collected from Tan Tock Seng Hospital, and its performance is compared against various existing RAG methods. Experimental results show that, leveraging the information integration and relational abilities of the KG, our MedRAG provides more specific diagnostic insights and outperforms state-of-the-art models in reducing misdiagnosis rates. Our code will be available at https://github.com/SNOWTEAM2023/MedRAG
Authors: He Hu, Yucheng Zhou, Lianzhong You, Hongbo Xu, Qianning Wang, Zheng Lian, Fei Richard Yu, Fei Ma, Laizhong Cui
Abstract: With the integration of Multimodal large language models (MLLMs) into robotic systems and various AI applications, embedding emotional intelligence (EI) capabilities into these models is essential for enabling robots to effectively address human emotional needs and interact seamlessly in real-world scenarios. Existing static, text-based, or text-image benchmarks overlook the multimodal complexities of real-world interactions and fail to capture the dynamic, multimodal nature of emotional expressions, making them inadequate for evaluating MLLMs' EI. Based on established psychological theories of EI, we build EmoBench-M, a novel benchmark designed to evaluate the EI capability of MLLMs across 13 valuation scenarios from three key dimensions: foundational emotion recognition, conversational emotion understanding, and socially complex emotion analysis. Evaluations of both open-source and closed-source MLLMs on EmoBench-M reveal a significant performance gap between them and humans, highlighting the need to further advance their EI capabilities. All benchmark resources, including code and datasets, are publicly available at https://emo-gml.github.io/.
Authors: Edoardo Loru, Jacopo Nudo, Niccol\`o Di Marco, Matteo Cinelli, Walter Quattrociocchi
Abstract: Large Language Models (LLMs) are increasingly used to assess news credibility, yet little is known about how they make these judgments. While prior research has examined political bias in LLM outputs or their potential for automated fact-checking, their internal evaluation processes remain largely unexamined. Understanding how LLMs assess credibility provides insights into AI behavior and how credibility is structured and applied in large-scale language models. This study benchmarks the reliability and political classifications of state-of-the-art LLMs - Gemini 1.5 Flash (Google), GPT-4o mini (OpenAI), and LLaMA 3.1 (Meta) - against structured, expert-driven rating systems such as NewsGuard and Media Bias Fact Check. Beyond assessing classification performance, we analyze the linguistic markers that shape LLM decisions, identifying which words and concepts drive their evaluations. We uncover patterns in how LLMs associate credibility with specific linguistic features by examining keyword frequency, contextual determinants, and rank distributions. Beyond static classification, we introduce a framework in which LLMs refine their credibility assessments by retrieving external information, querying other models, and adapting their responses. This allows us to investigate whether their assessments reflect structured reasoning or rely primarily on prior learned associations.
Authors: Yu-Neng Chuang, Leisheng Yu, Guanchu Wang, Lizhe Zhang, Zirui Liu, Xuanting Cai, Yang Sui, Vladimir Braverman, Xia Hu
Abstract: Large language models (LLMs) are increasingly deployed and democratized on edge devices. To improve the efficiency of on-device deployment, small language models (SLMs) are often adopted due to their efficient decoding latency and reduced energy consumption. However, these SLMs often generate inaccurate responses when handling complex queries. One promising solution is uncertainty-based SLM routing, offloading high-stakes queries to stronger LLMs when resulting in low-confidence responses on SLM. This follows the principle of "If you lack confidence, seek stronger support" to enhance reliability. Relying on more powerful LLMs is yet effective but increases invocation costs. Therefore, striking a routing balance between efficiency and efficacy remains a critical challenge. Additionally, efficiently generalizing the routing strategy to new datasets remains under-explored. In this paper, we conduct a comprehensive investigation into benchmarking and generalization of uncertainty-driven routing strategies from SLMs to LLMs over 1500+ settings. Our findings highlight: First, uncertainty-correctness alignment in different uncertainty quantification (UQ) methods significantly impacts routing performance. Second, uncertainty distributions depend more on both the specific SLM and the chosen UQ method, rather than downstream data. Building on the insight, we propose a calibration data construction instruction pipeline and open-source a constructed hold-out set to enhance routing generalization on new downstream scenarios. The experimental results indicate calibration data effectively bootstraps routing performance without any new data.
Authors: Karolina Rudnicka
Abstract: The paper is an empirical case study of grammatical obsolescence in progress. The main studied variable is the purpose subordinator in order that, which is shown to be steadily decreasing in the frequency of use starting from the beginning of the twentieth century. This work applies a data-driven approach for the investigation and description of obsolescence, recently developed by the Rudnicka (2019). The methodology combines philological analysis with statistical methods used on data acquired from mega-corpora. Moving from the description of possible symptoms of obsolescence to different causes for it, the paper aims at presenting a comprehensive account of the studied phenomenon. Interestingly, a very significant role in the decline of in order that can be ascribed to the so-called higher-order processes, understood as processes influencing the constructional level from above. Two kinds of higher-order processes are shown to play an important role, namely i) an externally-motivated higher-order process exemplified by the drastic socio-cultural changes of the 19th and 20th centuries; ii) an internally-motivated higher-order processes instantiated by the rise of the to-infinitive (rise of infinite clauses).
Authors: Katarzyna Kobalczyk, Nicolas Astorga, Tennison Liu, Mihaela van der Schaar
Abstract: Despite the impressive performance of large language models (LLMs) across various benchmarks, their ability to address ambiguously specified problems--frequent in real-world interactions--remains underexplored. To address this gap, we introduce a formal definition of task ambiguity and frame the problem of task disambiguation through the lens of Bayesian Experimental Design. By posing clarifying questions, LLM agents can acquire additional task specifications, progressively narrowing the space of viable solutions and reducing the risk of generating unsatisfactory outputs. Yet, generating effective clarifying questions requires LLM agents to engage in a form of meta-cognitive reasoning, an ability LLMs may presently lack. Our proposed approach of active task disambiguation enables LLM agents to generate targeted questions maximizing the information gain. Effectively, this approach shifts the load from implicit to explicit reasoning about the space of viable solutions. Empirical results demonstrate that this form of question selection leads to more effective task disambiguation in comparison to approaches relying on reasoning solely within the space of questions.
Authors: Edward Hong Wang, Cynthia Xin Wen
Abstract: Recent advancements in large language models have demonstrated that extended inference through techniques can markedly improve performance, yet these gains come with increased computational costs and the propagation of inherent biases found in natural languages. This paper explores the design of a unified AI-centric language system that addresses these challenges by offering a more concise, unambiguous, and computationally efficient alternative to traditional human languages. We analyze the limitations of natural language such as gender bias, morphological irregularities, and contextual ambiguities and examine how these issues are exacerbated within current Transformer architectures, where redundant attention heads and token inefficiencies prevail. Drawing on insights from emergent artificial communication systems and constructed languages like Esperanto and Lojban, we propose a framework that translates diverse natural language inputs into a streamlined AI-friendly language, enabling more efficient model training and inference while reducing memory footprints. Finally, we outline a pathway for empirical validation through controlled experiments, paving the way for a universal interchange format that could revolutionize AI-to-AI and human-to-AI interactions by enhancing clarity, fairness, and overall performance.
Authors: Selim Furkan Tekin, Fatih Ilhan, Tiansheng Huang, Sihao Hu, Zachary Yahn, Ling Liu
Abstract: The advancement of Large Language Models (LLMs) and their finetuning strategies has triggered the renewed interests in multi-agent reinforcement learning. In this paper, we introduce a focal diversity-optimized multi-agent reinforcement learning approach, coined as MARL-Focal, with three unique characteristics. First, we develop an agent-fusion framework for encouraging multiple LLM based agents to collaborate in producing the final inference output for each LLM query. Second, we develop a focal-diversity optimized agent selection algorithm that can choose a small subset of the available agents based on how well they can complement one another to generate the query output. Finally, we design a conflict-resolution method to detect output inconsistency among multiple agents and produce our MARL-Focal output through reward-aware and policy-adaptive inference fusion. Extensive evaluations on five benchmarks show that MARL-Focal is cost-efficient and adversarial-robust. Our multi-agent fusion model achieves performance improvement of 5.51\% compared to the best individual LLM-agent and offers stronger robustness over the TruthfulQA benchmark. Code is available at https://github.com/sftekin/rl-focal
Authors: Zhaoyang Wang, Jinqi Jiang, Huichi Zhou, Wenhao Zheng, Xuchao Zhang, Chetan Bansal, Huaxiu Yao
Abstract: Recent Large Language Models (LLMs) have demonstrated satisfying general instruction following ability. However, small LLMs with about 7B parameters still struggle fine-grained format following (e.g., JSON format), which seriously hinder the advancements of their applications. Most existing methods focus on benchmarking general instruction following while overlook how to improve the specific format following ability for small LLMs. Besides, these methods often rely on evaluations based on advanced LLMs (e.g., GPT-4), which can introduce the intrinsic bias of LLMs and be costly due to the API calls. In this paper, we first curate a fully verifiable format following dataset VFF. In contrast to existing works often adopting external LLMs for instruction-following validations, every sample of VFF can be easily validated with a Python function. Further, we propose to leverage this verifiable feature to synthesize massive data for progressively training small LLMs, in order to improve their format following abilities. Experimental results highlight the prevalent limitations in the format following capabilities of 7B level open-source LLMs and demonstrate the effectiveness of our method in enhancing this essential ability.
Authors: Zijun Wu, Yongchang Hao, Lili Mou
Abstract: Large language models achieve state-of-the-art performance but are costly to fine-tune due to their size. Parameter-efficient fine-tuning methods, such as prompt tuning, address this by reducing trainable parameters while maintaining strong performance. However, prior methods tie prompt embeddings to the model's dimensionality, which may not scale well with larger LLMs and more customized LLMs. In this paper, we propose Ultra-Low-dimensional Prompt Tuning (ULPT), which optimizes prompts in a low-dimensional space (e.g., 2D) and use a random but frozen matrix for the up-projection. To enhance alignment, we introduce learnable shift and scale embeddings. ULPT drastically reduces the trainable parameters, e.g., 2D only using 2% parameters compared with vanilla prompt tuning while retaining most of the performance across 21 NLP tasks. Our theoretical analysis shows that random projections can capture high-rank structures effectively, and experimental results demonstrate ULPT's competitive performance over existing parameter-efficient methods.
Authors: Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, Yulia Tsvetkov
Abstract: This position paper argues that in many realistic (i.e., complex, contextualized, subjective) scenarios, one LLM is not enough to produce a reliable output. We challenge the status quo of relying solely on a single general-purpose LLM and argue for multi-LLM collaboration to better represent the extensive diversity of data, skills, and people. We first posit that a single LLM underrepresents real-world data distributions, heterogeneous skills, and pluralistic populations, and that such representation gaps cannot be trivially patched by further training a single LLM. We then organize existing multi-LLM collaboration methods into a hierarchy, based on the level of access and information exchange, ranging from API-level, text-level, logit-level, to weight-level collaboration. Based on these methods, we highlight how multi-LLM collaboration addresses challenges that a single LLM struggles with, such as reliability, democratization, and pluralism. Finally, we identify the limitations of existing multi-LLM methods and motivate future work. We envision multi-LLM collaboration as an essential path toward compositional intelligence and collaborative AI development.
Authors: Shangbin Feng, Zifeng Wang, Palash Goyal, Yike Wang, Weijia Shi, Huang Xia, Hamid Palangi, Luke Zettlemoyer, Yulia Tsvetkov, Chen-Yu Lee, Tomas Pfister
Abstract: We propose Heterogeneous Swarms, an algorithm to design multi-LLM systems by jointly optimizing model roles and weights. We represent multi-LLM systems as directed acyclic graphs (DAGs) of LLMs with topological message passing for collaborative generation. Given a pool of LLM experts and a utility function, Heterogeneous Swarms employs two iterative steps: role-step and weight-step. For role-step, we interpret model roles as learning a DAG that specifies the flow of inputs and outputs between LLMs. Starting from a swarm of random continuous adjacency matrices, we decode them into discrete DAGs, call the LLMs in topological order, evaluate on the utility function (e.g. accuracy on a task), and optimize the adjacency matrices with particle swarm optimization based on the utility score. For weight-step, we assess the contribution of individual LLMs in the multi-LLM systems and optimize model weights with swarm intelligence. We propose JFK-score to quantify the individual contribution of each LLM in the best-found DAG of the role-step, then optimize model weights with particle swarm optimization based on the JFK-score. Experiments demonstrate that Heterogeneous Swarms outperforms 15 role- and/or weight-based baselines by 18.5% on average across 12 tasks. Further analysis reveals that Heterogeneous Swarms discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of language models.
Authors: Shuhaib Mehri, Xiusi Chen, Heng Ji, Dilek Hakkani-T\"ur
Abstract: LLMs demonstrate remarkable capabilities in following natural language instructions, largely due to instruction-tuning on high-quality datasets. While synthetic data generation has emerged as a scalable approach for creating such datasets, maintaining consistent quality standards remains challenging. Recent approaches incorporate feedback to improve data quality, but typically operate at the sample level, generating and applying feedback for each response individually. In this work, we propose Reference-Level Feedback, a novel methodology that instead collects feedback based on high-quality reference samples from carefully curated seed data. We use this feedback to capture rich signals of desirable characteristics that can be propagated to newly synthesized data. We present REFED, a dataset of 10K instruction-response pairs synthesized using such feedback. We demonstrate the effectiveness of our approach by showing that Llama-3.1-8B-Instruct finetuned on REFED achieves state-of-the-art performance among similar-sized SFT-based models on AlpacaEval 2.0 and strong results on Arena-Hard. Through extensive experiments, we show that our approach consistently outperforms traditional sample-level feedback methods with significantly fewer feedback collections and improves performance across different model architectures.
Authors: Letian Peng, Chenyang An, Shibo Hao, Chengyu Dong, Jingbo Shang
Abstract: The generalization of language models (LMs) is undergoing active debates, contrasting their potential for general intelligence with their struggles with basic knowledge composition (e.g., reverse/transition curse). This paper uncovers the phenomenon of linear correlations in LMs during knowledge composition. For explanation, there exists a linear transformation between certain related knowledge that maps the next token prediction logits from one prompt to another, e.g., "X lives in the city of" $\rightarrow$ "X lives in the country of" for every given X. This mirrors the linearity in human knowledge composition, such as Paris $\rightarrow$ France. Our findings indicate that the linear transformation is resilient to large-scale fine-tuning, generalizing updated knowledge when aligned with real-world relationships, but causing hallucinations when it deviates. Empirical results suggest that linear correlation can serve as a potential identifier of LM's generalization. Finally, we show such linear correlations can be learned with a single feedforward network and pre-trained vocabulary representations, indicating LM generalization heavily relies on the latter.
Authors: Minseok Jung, Cynthia Fuertes Panizo, Liam Dugan, May Fung, Pin-Yu Chen, Paul Pu Liang
Abstract: The advancement of large language models (LLMs) has made it difficult to differentiate human-written text from AI-generated text. Several AI-text detectors have been developed in response, which typically utilize a fixed global threshold (e.g., {\theta} = 0.5) to classify machine-generated text. However, we find that one universal threshold can fail to account for subgroup-specific distributional variations. For example, when using a fixed threshold, detectors make more false positive errors on shorter human-written text than longer, and more positive classifications on neurotic writing styles than open among long text. These discrepancies can lead to misclassification that disproportionately affects certain groups. We address this critical limitation by introducing FairOPT, an algorithm for group-specific threshold optimization in AI-generated content classifiers. Our approach partitions data into subgroups based on attributes (e.g., text length and writing style) and learns decision thresholds for each group, which enables careful balancing of performance and fairness metrics within each subgroup. In experiments with four AI text classifiers on three datasets, FairOPT enhances overall F1 score and decreases balanced error rate (BER) discrepancy across subgroups. Our framework paves the way for more robust and fair classification criteria in AI-generated output detection.
Authors: Chenyang Huang, Hao Zhou, Cameron Jen, Kangjie Zheng, Osmar R. Za\"iane, Lili Mou
Abstract: Length-control summarization aims to condense long texts into a short one within a certain length limit. Previous approaches often use autoregressive (AR) models and treat the length requirement as a soft constraint, which may not always be satisfied. In this study, we propose a novel length-control decoding algorithm based on the Directed Acyclic Transformer (DAT). Our approach allows for multiple plausible sequence fragments and predicts a \emph{path} to connect them. In addition, we propose a Sequence Maximum a Posteriori (SeqMAP) decoding algorithm that marginalizes different possible paths and finds the most probable summary satisfying the length budget. Our algorithm is based on beam search, which further facilitates a reranker for performance improvement. Experimental results on the Gigaword and DUC2004 datasets demonstrate our state-of-the-art performance for length-control summarization.
Authors: Chenyang Huang, Fei Huang, Zaixiang Zheng, Osmar R. Za\"iane, Hao Zhou, Lili Mou
Abstract: Multilingual neural machine translation (MNMT) aims at using one single model for multiple translation directions. Recent work applies non-autoregressive Transformers to improve the efficiency of MNMT, but requires expensive knowledge distillation (KD) processes. To this end, we propose an M-DAT approach to non-autoregressive multilingual machine translation. Our system leverages the recent advance of the directed acyclic Transformer (DAT), which does not require KD. We further propose a pivot back-translation (PivotBT) approach to improve the generalization to unseen translation directions. Experiments show that our M-DAT achieves state-of-the-art performance in non-autoregressive MNMT.
Authors: Daphne Quillington, Kingsley Fairbrother, Xavier Tattershall, Irin Kabakum
Abstract: Optimization methodologies for training large-scale neural architectures often rely on uniform gradient propagation mechanisms that fail to align with hierarchical linguistic structures, limiting their capacity to generalize across diverse language distributions. A structured gradient refinement framework was introduced to incorporate multi-scale contextual adjustments, improving parameter adaptation through dynamic weighting strategies that enhanced representation coherence. Empirical evaluations demonstrated that structured propagation mechanisms contributed to reductions in gradient oscillations, resulting in more stable training dynamics and improved optimization efficiency. The comparative performance assessment indicated that models incorporating hierarchical propagation strategies exhibited greater robustness in long-range dependency retention and cross-domain adaptation. The hierarchical adjustment of weight updates provided an alternative to conventional backpropagation, reducing sensitivity to initialization conditions while improving overall convergence efficiency. The experimental results confirmed that structured gradient propagation influenced representation learning trajectories, aligning parameter updates with broader linguistic dependencies rather than isolated token-level relationships. Statistical evaluations indicated that structured optimization strategies mitigated overfitting while preserving adaptability across heterogeneous text distributions. The findings established that structured gradient propagation provided an empirically validated framework for refining hierarchical representation learning, supporting more effective integration of linguistic dependencies into optimization dynamics.
Authors: Hanyu Wang, Bochuan Cao, Yuanpu Cao, Jinghui Chen
Abstract: Large language models (LLMs) are known to struggle with consistently generating truthful responses. While various representation intervention techniques have been proposed, these methods typically apply a universal representation correction vector to all input queries, limiting their effectiveness against diverse queries in practice. In this study, we introduce TruthFlow, a novel method that leverages the Flow Matching technique for query-specific truthful representation correction. Specifically, TruthFlow first uses a flow model to learn query-specific correction vectors that transition representations from hallucinated to truthful states. Then, during inference, the trained flow model generates these correction vectors to enhance the truthfulness of LLM outputs. Experimental results demonstrate that TruthFlow significantly improves performance on open-ended generation tasks across various advanced LLMs evaluated on TruthfulQA. Moreover, the trained TruthFlow model exhibits strong transferability, performing effectively on other unseen hallucination benchmarks.
Authors: Sandra C. Sandoval, Christabel Acquaye, Kwesi Cobbina, Mohammad Nayeem Teli, Hal Daum\'e III
Abstract: We examine the representation of African American English (AAE) in large language models (LLMs), exploring (a) the perceptions Black Americans have of how effective these technologies are at producing authentic AAE, and (b) in what contexts Black Americans find this desirable. Through both a survey of Black Americans ($n=$ 104) and annotation of LLM-produced AAE by Black Americans ($n=$ 228), we find that Black Americans favor choice and autonomy in determining when AAE is appropriate in LLM output. They tend to prefer that LLMs default to communicating in Mainstream U.S. English in formal settings, with greater interest in AAE production in less formal settings. When LLMs were appropriately prompted and provided in context examples, our participants found their outputs to have a level of AAE authenticity on par with transcripts of Black American speech. Select code and data for our project can be found here: https://github.com/smelliecat/AAEMime.git
Authors: Runjin Chen, Gabriel Jacob Perin, Xuxi Chen, Xilun Chen, Yan Han, Nina S. T. Hirata, Junyuan Hong, Bhavya Kailkhura
Abstract: Alignment of large language models (LLMs) with human values and preferences, often achieved through fine-tuning based on human feedback, is essential for ensuring safe and responsible AI behaviors. However, the process typically requires substantial data and computation resources. Recent studies have revealed that alignment might be attainable at lower costs through simpler methods, such as in-context learning. This leads to the question: Is alignment predominantly superficial? In this paper, we delve into this question and provide a quantitative analysis. We formalize the concept of superficial knowledge, defining it as knowledge that can be acquired through easily token restyling, without affecting the model's ability to capture underlying causal relationships between tokens. We propose a method to extract and isolate superficial knowledge from aligned models, focusing on the shallow modifications to the final token selection process. By comparing models augmented only with superficial knowledge to fully aligned models, we quantify the superficial portion of alignment. Our findings reveal that while superficial knowledge constitutes a significant portion of alignment, particularly in safety and detoxification tasks, it is not the whole story. Tasks requiring reasoning and contextual understanding still rely on deeper knowledge. Additionally, we demonstrate two practical advantages of isolated superficial knowledge: (1) it can be transferred between models, enabling efficient offsite alignment of larger models using extracted superficial knowledge from smaller models, and (2) it is recoverable, allowing for the restoration of alignment in compromised models without sacrificing performance.
Authors: Weiwei Sun (Department of Computer Science and Technology, University of Cambridge), Xiaoxi Luo (Yuanpei College, Peking University)
Abstract: This paper is concerned with phonetic reconstruction of the consonant system of Middle Chinese. We propose to cast the problem as a Mixed Integer Programming problem, which is able to automatically explore homophonic information from ancient rhyme dictionaries and phonetic information from modern Chinese dialects, the descendants of Middle Chinese. Numerical evaluation on a wide range of synthetic and real data demonstrates the effectiveness and robustness of the new method. We apply the method to information from Guangyun and 20 modern Chinese dialects to obtain a new phonetic reconstruction result. A linguistically-motivated discussion of this result is also provided.
Authors: Lin Tian, Emily Booth, Francesco Bailo, Julian Droogan, Marian-Andrei Rizoiu
Abstract: In today's digital age, conspiracies and information campaigns can emerge rapidly and erode social and democratic cohesion. While recent deep learning approaches have made progress in modeling engagement through language and propagation models, they struggle with irregularly sampled data and early trajectory assessment. We present IC-Mamba, a novel state space model that forecasts social media engagement by modeling interval-censored data with integrated temporal embeddings. Our model excels at predicting engagement patterns within the crucial first 15-30 minutes of posting (RMSE 0.118-0.143), enabling rapid assessment of content reach. By incorporating interval-censored modeling into the state space framework, IC-Mamba captures fine-grained temporal dynamics of engagement growth, achieving a 4.72% improvement over state-of-the-art across multiple engagement metrics (likes, shares, comments, and emojis). Our experiments demonstrate IC-Mamba's effectiveness in forecasting both post-level dynamics and broader narrative patterns (F1 0.508-0.751 for narrative-level predictions). The model maintains strong predictive performance across extended time horizons, successfully forecasting opinion-level engagement up to 28 days ahead using observation windows of 3-10 days. These capabilities enable earlier identification of potentially problematic content, providing crucial lead time for designing and implementing countermeasures. Code is available at: https://github.com/ltian678/ic-mamba. An interactive dashboard demonstrating our results is available at: https://ic-mamba.behavioral-ds.science.
URLs: https://github.com/ltian678/ic-mamba., https://ic-mamba.behavioral-ds.science.
Authors: Soichiro Murakami, Peinan Zhang, Hidetaka Kamigaito, Hiroya Takamura, Manabu Okumura
Abstract: Effective linguistic choices that attract potential customers play crucial roles in advertising success. This study aims to explore the linguistic features of ad texts that influence human preferences. Although the creation of attractive ad texts is an active area of research, progress in understanding the specific linguistic features that affect attractiveness is hindered by several obstacles. First, human preferences are complex and influenced by multiple factors, including their content, such as brand names, and their linguistic styles, making analysis challenging. Second, publicly available ad text datasets that include human preferences are lacking, such as ad performance metrics and human feedback, which reflect people's interests. To address these problems, we present AdParaphrase, a paraphrase dataset that contains human preferences for pairs of ad texts that are semantically equivalent but differ in terms of wording and style. This dataset allows for preference analysis that focuses on the differences in linguistic features. Our analysis revealed that ad texts preferred by human judges have higher fluency, longer length, more nouns, and use of bracket symbols. Furthermore, we demonstrate that an ad text-generation model that considers these findings significantly improves the attractiveness of a given text. The dataset is publicly available at: https://github.com/CyberAgentAILab/AdParaphrase.
Authors: Antoine Dussolle, Andrea Carde\~na D\'iaz, Shota Sato, Peter Devine
Abstract: Instruction following is a core capability of modern Large language models (LLMs), making evaluating this capability essential to understanding these models. The Instruction Following Evaluation (IFEval) benchmark from the literature does this using objective criteria, offering a measure of LLM performance without subjective AI or human judgement. However, it only includes English instructions, limiting its ability to assess LLMs in other languages. We propose the Multilingual Instruction Following Evaluation (M-IFEval) benchmark, expanding the evaluation to French, Japanese, and Spanish, with both general and language-specific instructions. Applying this benchmark to 8 state-of-the-art LLMs, we find that benchmark performance across languages and instruction types can vary widely, underscoring the importance of a multilingual benchmark for evaluating LLMs in a diverse cultural context.
Authors: Yuwei Yin, Giuseppe Carenini
Abstract: Large language models (LLMs) achieve remarkable performance on challenging benchmarks that are often structured as multiple-choice question-answering (QA) tasks. Zero-shot Chain-of-Thought (CoT) prompting enhances reasoning in LLMs but provides only vague and generic guidance ("think step by step"). This paper introduces ARR, an intuitive and effective zero-shot prompting method that explicitly incorporates three key steps in QA solving: analyzing the intent of the question, retrieving relevant information, and reasoning step by step. Comprehensive experiments across diverse and challenging QA tasks demonstrate that ARR consistently improves the Baseline (without ARR prompting) and outperforms CoT. Ablation and case studies further validate the positive contributions of each component: analyzing, retrieving, and reasoning. Notably, intent analysis plays a vital role in ARR. Additionally, extensive evaluations across various model sizes, LLM series, and generation settings solidify the effectiveness, robustness, and generalizability of ARR.
Authors: Kazuya Matsuo, Yoko Ishii, Atsushi Otsuka, Ryo Ishii, Hiroaki Sugiyama, Masahiro Mizukami, Tsunehiro Arimoto, Narichika Nomoto, Yoshihide Sato, Tetsuya Yamaguchi
Abstract: This paper focuses on simulating text dialogues in which impressions between speakers improve during speed dating. This simulation involves selecting an utterance from multiple candidates generated by a text generation model that replicates a specific speaker's utterances, aiming to improve the impression of the speaker. Accurately selecting an utterance that improves the impression is crucial for the simulation. We believe that whether an utterance improves a dialogue partner's impression of the speaker may depend on the personalities of both parties. However, recent methods for utterance selection do not consider the impression per utterance or the personalities. To address this, we propose a method that predicts whether an utterance improves a partner's impression of the speaker, considering the personalities. The evaluation results showed that personalities are useful in predicting impression changes per utterance. Furthermore, we conducted a human evaluation of simulated dialogues using our method. The results showed that it could simulate dialogues more favorably received than those selected without considering personalities.
Authors: Sourabrata Mukherjee, Atul Kr. Ojha, John P. McCrae, Ondrej Dusek
Abstract: Text Style Transfer (TST) is the task of transforming a text to reflect a particular style while preserving its original content. Evaluating TST outputs is a multidimensional challenge, requiring the assessment of style transfer accuracy, content preservation, and naturalness. Using human evaluation is ideal but costly, same as in other natural language processing (NLP) tasks, however, automatic metrics for TST have not received as much attention as metrics for, e.g., machine translation or summarization. In this paper, we examine both set of existing and novel metrics from broader NLP tasks for TST evaluation, focusing on two popular subtasks-sentiment transfer and detoxification-in a multilingual context comprising English, Hindi, and Bengali. By conducting meta-evaluation through correlation with human judgments, we demonstrate the effectiveness of these metrics when used individually and in ensembles. Additionally, we investigate the potential of Large Language Models (LLMs) as tools for TST evaluation. Our findings highlight that certain advanced NLP metrics and experimental-hybrid-techniques, provide better insights than existing TST metrics for delivering more accurate, consistent, and reproducible TST evaluations.
Authors: Karolina Rudnicka
Abstract: The topic of "negative end" of change is, contrary to the fields of innovation and emergence, largely under-researched. Yet, it has lately started to gain an increasing attention from language scholars worldwide. The main focus of this article is threefold, namely to discuss the i) terminology; ii) concepts and iii) causes associated with the "negative end" of change in grammar. The article starts with an overview of research conducted on the topic. It then moves to situating phenomena referred to as loss, decline or obsolescence among processes of language change, before elaborating on the terminology and concepts behind it. The last part looks at possible causes for constructions to display a (gradual or rapid, but very consistent) decrease in the frequency of use over time, which continues until the construction disappears or there are only residual or fossilised forms left. Keywords: loss, obsolescence, decline, competition, higher
Authors: Ma\"el Kubli
Abstract: This paper presents a novel methodological framework for detecting and classifying latent constructs, including frames, narratives, and topics, from textual data using Open-Source Large Language Models (LLMs). The proposed hybrid approach combines automated summarization with human-in-the-loop validation to enhance the accuracy and interpretability of construct identification. By employing iterative sampling coupled with expert refinement, the framework guarantees methodological robustness and ensures conceptual precision. Applied to diverse data sets, including AI policy debates, newspaper articles on encryption, and the 20 Newsgroups data set, this approach demonstrates its versatility in systematically analyzing complex political discourses, media framing, and topic classification tasks.
Authors: Jungwoo Kim, Minsang Kim, Sungjin Lee
Abstract: The rapid evolution of Large Language Models (LLMs) has enabled the industry to develop various AI-based services. Instruction tuning is considered essential in adapting foundation models for target domains to provide high-quality services to customers. A key challenge in instruction tuning is obtaining high-quality instruction data. Self-Instruct, which automatically generates instruction data using ChatGPT APIs, alleviates the data scarcity problem. To improve the quality of instruction data, Self-Instruct discards many of the instructions generated from ChatGPT, even though it is inefficient in terms of cost owing to many useless API calls. To generate high-quality instruction data at a low cost, we propose a novel data generation framework, Self-Direct Instruction generation (SeDi-Instruct), which employs diversity-based filtering and iterative feedback task generation. Diversity-based filtering maintains model accuracy without excessively discarding low-quality generated instructions by enhancing the diversity of instructions in a batch. This reduces the cost of synthesizing instruction data. The iterative feedback task generation integrates instruction generation and training tasks and utilizes information obtained during the training to create high-quality instruction sets. Our results show that SeDi-Instruct enhances the accuracy of AI models by 5.2%, compared with traditional methods, while reducing data generation costs by 36%.
Authors: Hassane Kissane, Achim Schilling, Patrick Krauss
Abstract: This study investigates the internal representations of verb-particle combinations, called multi-word verbs, within transformer-based large language models (LLMs), specifically examining how these models capture lexical and syntactic properties at different neural network layers. Using the BERT architecture, we analyze the representations of its layers for two different verb-particle constructions: phrasal verbs like 'give up' and prepositional verbs like 'look at'. Our methodology includes training probing classifiers on the internal representations to classify these categories at both word and sentence levels. The results indicate that the model's middle layers achieve the highest classification accuracies. To further analyze the nature of these distinctions, we conduct a data separability test using the Generalized Discrimination Value (GDV). While GDV results show weak linear separability between the two verb types, probing classifiers still achieve high accuracy, suggesting that representations of these linguistic categories may be non-linearly separable. This aligns with previous research indicating that linguistic distinctions in neural networks are not always encoded in a linearly separable manner. These findings computationally support usage-based claims on the representation of verb-particle constructions and highlight the complex interaction between neural network architectures and linguistic structures.
Authors: Yuting Zeng, Weizhe Huang, Lei Jiang, Tongxuan Liu, Xitai Jin, Chen Tianying Tiana, Jing Li, Xiaohua Xu
Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across various natural language processing (NLP) scenarios, but they still face challenges when handling complex arithmetic and logical reasoning tasks. While Chain-Of-Thought (CoT) reasoning, self-consistency (SC) and self-correction strategies have attempted to guide models in sequential, multi-step reasoning, Multi-agent Debate (MAD) has emerged as a viable approach for enhancing the reasoning capabilities of LLMs. By increasing both the number of agents and the frequency of debates, the performance of LLMs improves significantly. However, this strategy results in a significant increase in token costs, presenting a barrier to scalability. To address this challenge, we introduce a novel sparsification strategy designed to reduce token costs within MAD. This approach minimizes ineffective exchanges of information and unproductive discussions among agents, thereby enhancing the overall efficiency of the debate process. We conduct comparative experiments on multiple datasets across various models, demonstrating that our approach significantly reduces the token costs in MAD to a considerable extent. Specifically, compared to MAD, our approach achieves an impressive reduction of up to 94.5\% in token costs while maintaining performance degradation below 2.0\%.
Authors: Masato Mita, Ryo Yoshida, Yohei Oseki
Abstract: Large language models exhibit general linguistic abilities but significantly differ from humans in their efficiency of language acquisition. This study proposes a method for integrating the developmental characteristics of working memory during the critical period, a stage when human language acquisition is particularly efficient, into language models. The proposed method introduces a mechanism that initially constrains working memory during the early stages of training and gradually relaxes this constraint in an exponential manner as learning progresses. Targeted syntactic evaluation shows that the proposed method outperforms conventional models without memory constraints or with static memory constraints. These findings not only provide new directions for designing data-efficient language models but also offer indirect evidence supporting the underlying mechanisms of the critical period hypothesis in human language acquisition.
Authors: Jing Yang, Max Glockner, Anderson Rocha, Iryna Gurevych
Abstract: Free-text explanations are expressive and easy to understand, but many datasets lack annotated explanation data, making it challenging to train models for explainable predictions. To address this, we investigate how to use existing explanation datasets for self-rationalization and evaluate models' out-of-distribution (OOD) performance. We fine-tune T5-Large and OLMo-7B models and assess the impact of fine-tuning data quality, the number of fine-tuning samples, and few-shot selection methods. The models are evaluated on 19 diverse OOD datasets across three tasks: natural language inference (NLI), fact-checking, and hallucination detection in abstractive summarization. For the generated explanation evaluation, we conduct a human study on 13 selected models and study its correlation with the Acceptability score (T5-11B) and three other LLM-based reference-free metrics. Human evaluation shows that the Acceptability score correlates most strongly with human judgments, demonstrating its effectiveness in evaluating free-text explanations. Our findings reveal: 1) few annotated examples effectively adapt models for OOD explanation generation; 2) compared to sample selection strategies, fine-tuning data source has a larger impact on OOD performance; and 3) models with higher label prediction accuracy tend to produce better explanations, as reflected by higher Acceptability scores.
Authors: Santiago Gonz\'alez-Silot, Andr\'es Montoro-Montarroso, Eugenio Mart\'inez C\'amara, Juan G\'omez-Romero
Abstract: The automatic detection of disinformation presents a significant challenge in the field of natural language processing. This task addresses a multifaceted societal and communication issue, which needs approaches that extend beyond the identification of general linguistic patterns through data-driven algorithms. In this research work, we hypothesise that text classification methods are not able to capture the nuances of disinformation and they often ground their decision in superfluous features. Hence, we apply a post-hoc explainability method (SHAP, SHapley Additive exPlanations) to identify spurious elements with high impact on the classification models. Our findings show that non-informative elements (e.g., URLs and emoticons) should be removed and named entities (e.g., Rwanda) should be pseudo-anonymized before training to avoid models' bias and increase their generalization capabilities. We evaluate this methodology with internal dataset and external dataset before and after applying extended data preprocessing and named entity replacement. The results show that our proposal enhances on average the performance of a disinformation classification method with external test data in 65.78% without a significant decrease of the internal test performance.
Authors: J. G\'omez-Romero, J. Cant\'on Correa, R. P\'erez Mercado, F. Prados Abad, M. Molina-Solana, W. Fajardo
Abstract: Telegram is a popular platform for public communication, generating large amounts of messages through its channels. pytopicgram is a Python library that helps researchers collect, organize, and analyze these Telegram messages. The library offers key features such as easy message retrieval, detailed channel information, engagement metrics, and topic identification using advanced modeling techniques. By simplifying data extraction and analysis, pytopicgram allows users to understand how content spreads and how audiences interact on Telegram. This paper describes the design, main features, and practical uses of \pytopicgram, showcasing its effectiveness for studying public conversations on Telegram.
Authors: Reihaneh Amooie, Wietse de Vries, Yun Hao, Jelske Dijkstra, Matt Coler, Martijn Wieling
Abstract: Automatic Speech Recognition (ASR) performance for low-resource languages is still far behind that of higher-resource languages such as English, due to a lack of sufficient labeled data. State-of-the-art methods deploy self-supervised transfer learning where a model pre-trained on large amounts of data is fine-tuned using little labeled data in a target low-resource language. In this paper, we present and examine a method for fine-tuning an SSL-based model in order to improve the performance for Frisian and its regional dialects (Clay Frisian, Wood Frisian, and South Frisian). We show that Frisian ASR performance can be improved by using multilingual (Frisian, Dutch, English and German) fine-tuning data and an auxiliary language identification task. In addition, our findings show that performance on dialectal speech suffers substantially, and, importantly, that this effect is moderated by the elicitation approach used to collect the dialectal data. Our findings also particularly suggest that relying solely on standard language data for ASR evaluation may underestimate real-world performance, particularly in languages with substantial dialectal variation.
Authors: Herbert Ullrich, Tom\'a\v{s} Mlyn\'a\v{r}, Jan Drchal
Abstract: In this paper, we explore the problem of Claim Extraction using one-to-many text generation methods, comparing LLMs, small summarization models finetuned for the task, and a previous NER-centric baseline QACG. As the current publications on Claim Extraction, Fact Extraction, Claim Generation and Check-worthy Claim Detection are quite scattered in their means and terminology, we compile their common objectives, releasing the FEVERFact dataset, with 17K atomic factual claims extracted from 4K contextualised Wikipedia sentences, adapted from the original FEVER. We compile the known objectives into an Evaluation framework of: Atomicity, Fluency, Decontextualization, Faithfulness checked for each generated claim separately, and Focus and Coverage measured against the full set of predicted claims for a single input. For each metric, we implement a scale using a reduction to an already-explored NLP task. We validate our metrics against human grading of generic claims, to see that the model ranking on $F_{fact}$, our hardest metric, did not change and the evaluation framework approximates human grading very closely in terms of $F_1$ and RMSE.
Authors: Jiayang Yu, Yihang Zhang, Bin Wang, Peiqin Lin, Yongkang Liu, Shi Feng
Abstract: Fine-tuning is a key approach for adapting language models to specific downstream tasks, but updating all model parameters becomes impractical as model sizes increase. Parameter-Efficient Fine-Tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), address this challenge by introducing additional adaptation parameters into pre-trained weight matrices. However, LoRA's performance varies across different insertion points within the model, highlighting potential parameter inefficiency due to unnecessary insertions. To this end, we propose SSMLoRA (State Space Model Low-Rank Adaptation), an extension of LoRA that incorporates a State Space Model (SSM) to interconnect low-rank matrices. SSMLoRA ensures that performance is maintained even with sparser insertions. SSMLoRA allows the model to not only map inputs to a low-rank space for better feature extraction but also leverage the computations from the previous low-rank space. Our method achieves comparable performance to LoRA on the General Language Understanding Evaluation (GLUE) benchmark while using only half the parameters. Additionally, due to its structure, SSMLoRA shows promise in handling tasks with longer input sequences. .You can find our code here:https://github.com/yuhkalhic/SSMLoRA.
Authors: Haohao Zhu, Junyu Lu, Zeyuan Zeng, Zewen Bai, Xiaokun Zhang, Liang Yang, Hongfei Lin
Abstract: Humor recognition aims to identify whether a specific speaker's text is humorous. Current methods for humor recognition mainly suffer from two limitations: (1) they solely focus on one aspect of humor commonalities, ignoring the multifaceted nature of humor; and (2) they typically overlook the critical role of speaker individuality, which is essential for a comprehensive understanding of humor expressions. To bridge these gaps, we introduce the Commonality and Individuality Incorporated Network for Humor Recognition (CIHR), a novel model designed to enhance humor recognition by integrating multifaceted humor commonalities with the distinctive individuality of speakers. The CIHR features a Humor Commonality Analysis module that explores various perspectives of multifaceted humor commonality within user texts, and a Speaker Individuality Extraction module that captures both static and dynamic aspects of a speaker's profile to accurately model their distinctive individuality. Additionally, Static and Dynamic Fusion modules are introduced to effectively incorporate the humor commonality with speaker's individuality in the humor recognition process. Extensive experiments demonstrate the effectiveness of CIHR, underscoring the importance of concurrently addressing both multifaceted humor commonality and distinctive speaker individuality in humor recognition.
Authors: Roman Vashurin (Mohamed bin Zayed University of Artificial Intelligence), Maiya Goloburda (Mohamed bin Zayed University of Artificial Intelligence), Preslav Nakov (Mohamed bin Zayed University of Artificial Intelligence), Artem Shelmanov (Mohamed bin Zayed University of Artificial Intelligence), Maxim Panov (Mohamed bin Zayed University of Artificial Intelligence)
Abstract: Uncertainty quantification (UQ) methods for Large Language Models (LLMs) encompasses a variety of approaches, with two major types being particularly prominent: information-based, which focus on model confidence expressed as token probabilities, and consistency-based, which assess the semantic relationship between multiple outputs generated using repeated sampling. Several recent methods have combined these two approaches and shown impressive performance in various applications. However, they sometimes fail to outperform much simpler baseline methods. Our investigation reveals distinctive characteristics of LLMs as probabilistic models, which help to explain why these UQ methods underperform in certain tasks. Based on these findings, we propose a new way of synthesizing model confidence and output consistency that leads to a family of efficient and robust UQ methods. We evaluate our approach across a variety of tasks such as question answering, abstractive summarization, and machine translation, demonstrating sizable improvements over state-of-the-art UQ approaches.
Authors: Gerrit J. J. van den Burg, Gen Suzuki, Wei Liu, Murat Sensoy
Abstract: Large language models (LLMs) are increasingly used as automated judges to evaluate recommendation systems, search engines, and other subjective tasks, where relying on human evaluators can be costly, time-consuming, and unscalable. LLMs offer an efficient solution for continuous, automated evaluation. However, since the systems that are built and improved with these judgments are ultimately designed for human use, it is crucial that LLM judgments align closely with human evaluators to ensure such systems remain human-centered. On the other hand, aligning LLM judgments with human evaluators is challenging due to individual variability and biases in human judgments. We propose a simple yet effective framework to align LLM judgments with individual human evaluators or their aggregated judgments, without retraining or fine-tuning the LLM. Our approach learns a linear mapping between the LLM's outputs and human judgments, achieving over 142% average improvement in agreement across 29 tasks with only a small number of calibration examples used for training. Notably, our method works in zero-shot and few-shot settings, exceeds inter-human agreement on four out of six tasks, and enables smaller LLMs to achieve performance comparable to that of larger models.
Authors: Geliang Ouyang, Jingyao Chen, Zhihe Nie, Yi Gui, Yao Wan, Hongyu Zhang, Dongping Chen
Abstract: Natural Language to Visualization (NL2Vis) seeks to convert natural-language descriptions into visual representations of given tables, empowering users to derive insights from large-scale data. Recent advancements in Large Language Models (LLMs) show promise in automating code generation to transform tabular data into accessible visualizations. However, they often struggle with complex queries that require reasoning across multiple tables. To address this limitation, we propose a collaborative agent workflow, termed nvAgent, for NL2Vis. Specifically, nvAgent comprises three agents: a processor agent for database processing and context filtering, a composer agent for planning visualization generation, and a validator agent for code translation and output verification. Comprehensive evaluations on the new VisEval benchmark demonstrate that nvAgent consistently surpasses state-of-the-art baselines, achieving a 7.88% improvement in single-table and a 9.23% improvement in multi-table scenarios. Qualitative analyses further highlight that nvAgent maintains nearly a 20% performance margin over previous models, underscoring its capacity to produce high-quality visual representations from complex, heterogeneous data sources.
Authors: Xiaoyu Deng, Ye Zhang, Tianmin Guo, Yongzhe Zhang, Zhengjian Kang, Hang Yang
Abstract: The astonishing performance of large language models (LLMs) and their remarkable achievements in production and daily life have led to their widespread application in collaborative tasks. However, current large models face challenges such as hallucination and lack of specificity in content generation in vertical domain tasks. Inspired by the contrast and classification mechanisms in human cognitive processes, this paper constructs an adversarial learning-based prompt framework named ChallengeMe, which includes three cascaded solutions: generation prompts, evaluation prompts, and feedback optimization. In this process, we designed seven core optimization dimensions and set the threshold for adversarial learning. The results of mixed case studies on the text summarization task show that the proposed framework can generate more accurate and fluent text summaries compared to the current advanced mainstream LLMs.
Authors: Kanghee Park, Timothy Zhou, Loris D'Antoni
Abstract: Large Language Models (LLMs) are often asked to generate structured outputs that obey precise syntactic rules, such as code snippets or formatted data. Grammar-constrained decoding (GCD) can guarantee that LLM outputs matches such rules by masking out tokens that will provably lead to outputs that do not belong to a specified context-free grammar (CFG). To guarantee soundness, GCD algorithms have to compute how a given LLM subword tokenizer can align with the tokens used by a given context-free grammar and compute token masks based on this information. Doing so efficiently is challenging and existing GCD algorithms require tens of minutes to preprocess common grammars. We present a new GCD algorithm together with an implementation that offers 17.71x faster offline preprocessing than existing approaches while preserving state-of-the-art efficiency in online mask computation.
Authors: Volker Emmrich
Abstract: This article explores the requirements for corpus compilation within the GiesKaNe project (University of Giessen and Kassel, Syntactic Basic Structures of New High German). The project is defined by three central characteristics: it is a reference corpus, a historical corpus, and a syntactically deeply annotated treebank. As a historical corpus, GiesKaNe aims to establish connections with both historical and contemporary corpora, ensuring its relevance across temporal and linguistic contexts. The compilation process strikes the balance between innovation and adherence to standards, addressing both internal project goals and the broader interests of the research community. The methodological complexity of such a project is managed through a complementary interplay of human expertise and machine-assisted processes. The article discusses foundational topics such as tokenization, normalization, sentence definition, tagging, parsing, and inter-annotator agreement, alongside advanced considerations. These include comparisons between grammatical models, annotation schemas, and established de facto annotation standards as well as the integration of human and machine collaboration. Notably, a novel method for machine-assisted classification of texts along the continuum of conceptual orality and literacy is proposed, offering new perspectives on text selection. Furthermore, the article introduces an approach to deriving de facto standard annotations from existing ones, mediating between standardization and innovation. In the course of describing the workflow the article demonstrates that even ambitious projects like GiesKaNe can be effectively implemented using existing research infrastructure, requiring no specialized annotation tools. Instead, it is shown that the workflow can be based on the strategic use of a simple spreadsheet and integrates the capabilities of the existing infrastructure.
Authors: Mukur Gupta, Noopur Bhatt, Suman Jana
Abstract: In this paper, we propose CodeSCM, a Structural Causal Model (SCM) for analyzing multi-modal code generation using large language models (LLMs). By applying interventions to CodeSCM, we measure the causal effects of different prompt modalities, such as natural language, code, and input-output examples, on the model. CodeSCM introduces latent mediator variables to separate the code and natural language semantics of a multi-modal code generation prompt. Using the principles of Causal Mediation Analysis on these mediators we quantify direct effects representing the model's spurious leanings. We find that, in addition to natural language instructions, input-output examples significantly influence code generation.
Authors: Steffen Eger, Yong Cao, Jennifer D'Souza, Andreas Geiger, Christian Greisinger, Stephanie Gross, Yufang Hou, Brigitte Krenn, Anne Lauscher, Yizhi Li, Chenghua Lin, Nafise Sadat Moosavi, Wei Zhao, Tristan Miller
Abstract: With the advent of large multimodal language models, science is now at a threshold of an AI-based technological transformation. Recently, a plethora of new AI models and tools has been proposed, promising to empower researchers and academics worldwide to conduct their research more effectively and efficiently. This includes all aspects of the research cycle, especially (1) searching for relevant literature; (2) generating research ideas and conducting experimentation; generating (3) text-based and (4) multimodal content (e.g., scientific figures and diagrams); and (5) AI-based automatic peer review. In this survey, we provide an in-depth overview over these exciting recent developments, which promise to fundamentally alter the scientific research process for good. Our survey covers the five aspects outlined above, indicating relevant datasets, methods and results (including evaluation) as well as limitations and scope for future research. Ethical concerns regarding shortcomings of these tools and potential for misuse (fake science, plagiarism, harms to research integrity) take a particularly prominent place in our discussion. We hope that our survey will not only become a reference guide for newcomers to the field but also a catalyst for new AI-based initiatives in the area of "AI4Science".
Authors: Yihe Deng, Yu Yang, Junkai Zhang, Wei Wang, Bo Li
Abstract: The rapid advancement of large language models (LLMs) has increased the need for guardrail models to ensure responsible use, particularly in detecting unsafe and illegal content. While substantial safety data exist in English, multilingual guardrail modeling remains underexplored due to the scarcity of open-source safety data in other languages. To address this gap, we propose a novel two-player Reinforcement Learning (RL) framework, where a generator and a guardrail model co-evolve adversarially to produce high-quality synthetic data for multilingual guardrail training. We theoretically formalize this interaction as a two-player game, proving convergence to a Nash equilibrium. Empirical evaluations show that our model \ours outperforms state-of-the-art models, achieving nearly 10% improvement over LlamaGuard3 (8B) on English benchmarks while being 4.5x faster at inference with a significantly smaller model (0.5B). We achieve substantial advancements in multilingual safety tasks, particularly in addressing the imbalance for lower-resource languages in a collected real dataset. Ablation studies emphasize the critical role of synthetic data generation in bridging the imbalance in open-source data between English and other languages. These findings establish a scalable and efficient approach to synthetic data generation, paving the way for improved multilingual guardrail models to enhance LLM safety. Code, model, and data will be open-sourced at https://github.com/yihedeng9/DuoGuard.
Authors: Ali Modarressi, Hanieh Deilamsalehy, Franck Dernoncourt, Trung Bui, Ryan A. Rossi, Seunghyun Yoon, Hinrich Sch\"utze
Abstract: Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 12 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 10 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information.
Authors: Haonan Wang, Minbin Huang, Runhui Huang, Lanqing Hong, Hang Xu, Tianyang Hu, Xiaodan Liang, Zhenguo Li, Hong Cheng, Kenji Kawaguchi
Abstract: Contrastive Language-Image Pre-training (CLIP) has become the standard for cross-modal image-text representation learning. Improving CLIP typically requires additional data and retraining with new loss functions, but these demands raise resource and time costs, limiting practical use. In this work, we introduce HELIP, a cost-effective strategy that improves CLIP models by exploiting challenging text-image pairs within existing datasets in continuous training. This eliminates the need for additional data or extensive retraining. Moreover, HELIP integrates effortlessly into current training pipelines with minimal code modifications, allowing for quick and seamless implementation. On comprehensive benchmarks, HELIP consistently boosts existing models. In particular, within just two epochs of training, it improves zero-shot classification accuracy on ImageNet for SLIP models pre-trained on CC3M, CC12M, and YFCC15M datasets by 3.05%, 4.47%, and 10.1% , respectively. In addition, on fine-grained classification datasets, HELIP improves the zero-shot performance of CLIP and SLIP by an average of 8.4% and 18.6%, and their linear probe performance by an average of 9.5% and 3.0%. The code is publicly available at: https://github.com/haonan3/HELIP-NACCL-2025.git.
Authors: Zining Zhu, Liang Zhao, Kangheng Lin, Jinze Yang, En Yu, Chenglong Liu, Haoran Wei, Jianjian Sun, Zheng Ge, Xiangyu Zhang
Abstract: This paper presents Perceptual Preference Optimization (PerPO), a perception alignment method aimed at addressing the visual discrimination challenges in generative pre-trained multimodal large language models (MLLMs). To align MLLMs with human visual perception process, PerPO employs discriminative rewarding to gather diverse negative samples, followed by listwise preference optimization to rank them.By utilizing the reward as a quantitative margin for ranking, our method effectively bridges generative preference optimization and discriminative empirical risk minimization. PerPO significantly enhances MLLMs' visual discrimination capabilities while maintaining their generative strengths, mitigates image-unconditional reward hacking, and ensures consistent performance across visual tasks. This work marks a crucial step towards more perceptually aligned and versatile MLLMs. We also hope that PerPO will encourage the community to rethink MLLM alignment strategies.
Authors: David Melhart, Matthew Barthet, Georgios N. Yannakakis
Abstract: Can out-of-the-box pretrained Large Language Models (LLMs) detect human affect successfully when observing a video? To address this question, for the first time, we evaluate comprehensively the capacity of popular LLMs to annotate and successfully predict continuous affect annotations of videos when prompted by a sequence of text and video frames in a multimodal fashion. Particularly in this paper, we test LLMs' ability to correctly label changes of in-game engagement in 80 minutes of annotated videogame footage from 20 first-person shooter games of the GameVibe corpus. We run over 2,400 experiments to investigate the impact of LLM architecture, model size, input modality, prompting strategy, and ground truth processing method on engagement prediction. Our findings suggest that while LLMs rightfully claim human-like performance across multiple domains, they generally fall behind capturing continuous experience annotations provided by humans. We examine some of the underlying causes for the relatively poor overall performance, highlight the cases where LLMs exceed expectations, and draw a roadmap for the further exploration of automated emotion labelling via LLMs.
Authors: Long Chen, Xiaotian Song, Andy Song, BaDong Chen, Jiancheng Lv, Yanan Sun
Abstract: Spiking Large Language Models have been shown as a good alternative to LLMs in various scenarios. Existing methods for creating Spiking LLMs, i.e., direct training and ANN-SNN conversion, often suffer from performance degradation and relatively high computational costs. To address these issues, we propose a novel Fast ANN-SNN conversion strategy (FAS) that transforms LLMs into spiking LLMs in two stages. The first stage employs a full-parameter fine-tuning of pre-trained models, so it does not need any direct training from scratch. The second stage introduces a coarse-to-fine calibration method to reduce conversion errors and improve accuracy. Our experiments on both language and vision-language tasks across four different scales of LLMs demonstrate that FAS can achieve state-of-the-art performance yet with significantly reduced inference latency and computational costs. For example, FAS only takes 8 timesteps to achieve an accuracy of 3% higher than that of the OPT-7B model, while reducing energy consumption by 96.63%.
Authors: Kunfeng Lai, Zhenheng Tang, Xinglin Pan, Peijie Dong, Xiang Liu, Haolan Chen, Li Shen, Bo Li, Xiaowen Chu
Abstract: Model merging aggregates Large Language Models (LLMs) finetuned on different tasks into a stronger one. However, parameter conflicts between models leads to performance degradation in averaging. While model routing addresses this issue by selecting individual models during inference, it imposes excessive storage and compute costs, and fails to leverage the common knowledge from different models. In this work, we observe that different layers exhibit varying levels of parameter conflicts. Building on this insight, we average layers with minimal parameter conflicts and use a novel task-level expert routing for layers with significant conflicts. To further reduce storage costs, inspired by task arithmetic sparsity, we decouple multiple fine-tuned experts into a dense expert and several sparse experts. Considering the out-of-distribution samples, we select and merge appropriate experts based on the task uncertainty of the input data. We conduct extensive experiments on both LLaMA and Qwen with varying parameter scales, and evaluate on real-world reasoning tasks. Results demonstrate that our method consistently achieves significant performance improvements while requiring less system cost compared to existing methods.
Authors: Ziyi Dong, Yao Xiao, Pengxu Wei, Liang Lin
Abstract: Groundbreaking advancements in text-to-image generation have recently been achieved with the emergence of diffusion models. These models exhibit a remarkable ability to generate highly artistic and intricately detailed images based on textual prompts. However, obtaining desired generation outcomes often necessitates repetitive trials of manipulating text prompts just like casting spells on a magic mirror, and the reason behind that is the limited capability of semantic understanding inherent in current image generation models. Specifically, existing diffusion models encode the text prompt input with a pre-trained encoder structure, which is usually trained on a limited number of image-caption pairs. The state-of-the-art large language models (LLMs) based on the decoder-only structure have shown a powerful semantic understanding capability as their architectures are more suitable for training on very large-scale unlabeled data. In this work, we propose to enhance text-to-image diffusion models by borrowing the strength of semantic understanding from large language models, and devise a simple yet effective adapter to allow the diffusion models to be compatible with the decoder-only structure. Meanwhile, we also provide a supporting theoretical analysis with various architectures (e.g., encoder-only, encoder-decoder, and decoder-only), and conduct extensive empirical evaluations to verify its effectiveness. The experimental results show that the enhanced models with our adapter module are superior to the stat-of-the-art models in terms of text-to-image generation quality and reliability.
Authors: Miaomiao Li, Hao Chen, Yang Wang, Tingyuan Zhu, Weijia Zhang, Kaijie Zhu, Kam-Fai Wong, Jindong Wang
Abstract: Generating synthetic datasets via large language models (LLMs) themselves has emerged as a promising approach to improve LLM performance. However, LLMs inherently reflect biases present in their training data, leading to a critical challenge: when these models generate synthetic data for training, they may propagate and amplify their inherent biases that can significantly impact model fairness and robustness on downstream tasks--a phenomenon we term bias inheritance. This work presents the first systematic investigation in understanding, analyzing, and mitigating bias inheritance. We study this problem by fine-tuning LLMs with a combined dataset consisting of original and LLM-augmented data, where bias ratio represents the proportion of augmented data. Through systematic experiments across 10 classification and generation tasks, we analyze how 6 different types of biases manifest at varying bias ratios. Our results reveal that bias inheritance has nuanced effects on downstream tasks, influencing both classification tasks and generation tasks differently. Then, our analysis identifies three key misalignment factors: misalignment of values, group data, and data distributions. Based on these insights, we propose three mitigation strategies: token-based, mask-based, and loss-based approaches. Experiments demonstrate that these strategies also work differently on various tasks and bias, indicating the substantial challenges to fully mitigate bias inheritance. We hope this work can provide valuable insights to the research of LLM data augmentation.
Authors: Xing Li, Zeyu Xing, Yiming Li, Linping Qu, Hui-Ling Zhen, Wulong Liu, Yiwu Yao, Sinno Jialin Pan, Mingxuan Yuan
Abstract: KV cache quantization can improve Large Language Models (LLMs) inference throughput and latency in long contexts and large batch-size scenarios while preserving LLMs effectiveness. However, current methods have three unsolved issues: overlooking layer-wise sensitivity to KV cache quantization, high overhead of online fine-grained decision-making, and low flexibility to different LLMs and constraints. Therefore, we thoroughly analyze the inherent correlation of layer-wise transformer attention patterns to KV cache quantization errors and study why key cache is more important than value cache for quantization error reduction. We further propose a simple yet effective framework KVTuner to adaptively search for the optimal hardware-friendly layer-wise KV quantization precision pairs for coarse-grained KV cache with multi-objective optimization and directly utilize the offline searched configurations during online inference. To reduce the computational cost of offline calibration, we utilize the intra-layer KV precision pair pruning and inter-layer clustering to reduce the search space. Experimental results show that we can achieve nearly lossless 3.25-bit mixed precision KV cache quantization for LLMs like Llama-3.1-8B-Instruct and 4.0-bit for sensitive models like Qwen2.5-7B-Instruct on mathematical reasoning tasks. The maximum inference throughput can be improved by 38.3% compared with KV8 quantization over various context lengths.
Authors: Khushboo Verma, Alan Michels, Ergi Gumusaneli, Shilpa Chitnis, Smita Sinha Kumar, Christopher Thompson, Lena Esmail, Guruprasath Srinivasan, Chandini Panchada, Sushovan Guha, Satwant Kumar
Abstract: Referral workflow inefficiencies, including misaligned referrals and delays, contribute to suboptimal patient outcomes and higher healthcare costs. In this study, we investigated the possibility of predicting procedural needs based on primary care diagnostic entries, thereby improving referral accuracy, streamlining workflows, and providing better care to patients. A de-identified dataset of 2,086 orthopedic referrals from the University of Texas Health at Tyler was analyzed using machine learning models built on Base General Embeddings (BGE) for semantic extraction. To ensure real-world applicability, noise tolerance experiments were conducted, and oversampling techniques were employed to mitigate class imbalance. The selected optimum and parsimonious embedding model demonstrated high predictive accuracy (ROC-AUC: 0.874, Matthews Correlation Coefficient (MCC): 0.540), effectively distinguishing patients requiring surgical intervention. Dimensionality reduction techniques confirmed the model's ability to capture meaningful clinical relationships. A threshold sensitivity analysis identified an optimal decision threshold (0.30) to balance precision and recall, maximizing referral efficiency. In the predictive modeling analysis, the procedure rate increased from 11.27% to an optimal 60.1%, representing a 433% improvement with significant implications for operational efficiency and healthcare revenue. The results of our study demonstrate that referral optimization can enhance primary and surgical care integration. Through this approach, precise and timely predictions of procedural requirements can be made, thereby minimizing delays, improving surgical planning, and reducing administrative burdens. In addition, the findings highlight the potential of clinical decision support as a scalable solution for improving patient outcomes and the efficiency of the healthcare system.
Authors: Daman Arora, Andrea Zanette
Abstract: Scaling model size and training data has led to great advances in the performance of Large Language Models (LLMs). However, the diminishing returns of this approach necessitate alternative methods to improve model capabilities, particularly in tasks requiring advanced reasoning. Large reasoning models, which leverage long chain-of-thoughts, bring unprecedented breakthroughs in problem-solving capabilities but at a substantial deployment cost associated to longer generations. Reducing inference costs is crucial for the economic feasibility, user experience, and environmental sustainability of these models. In this work, we propose to train large reasoning models to reason efficiently. More precisely, we use reinforcement learning (RL) to train reasoning models to dynamically allocate inference-time compute based on task complexity. Our method incentivizes models to minimize unnecessary computational overhead while maintaining accuracy, thereby achieving substantial efficiency gains. It enables the derivation of a family of reasoning models with varying efficiency levels, controlled via a single hyperparameter. Experiments on two open-weight large reasoning models demonstrate significant reductions in inference cost while preserving most of the accuracy.
Authors: Zony Yu, Yuqiao Wen, Lili Mou
Abstract: Knowledge distillation (KD) is a popular method of transferring knowledge from a large "teacher" model to a small "student" model. KD can be divided into two categories: prediction matching and intermediate-layer matching. We explore an intriguing phenomenon: layer-selection strategy does not matter (much) in intermediate-layer matching. In this paper, we show that seemingly nonsensical matching strategies such as matching the teacher's layers in reverse still result in surprisingly good student performance. We provide an interpretation for this phenomenon by examining the angles between teacher layers viewed from the student's perspective.
Authors: Ahmad Rashid, Ruotian Wu, Rongqi Fan, Hongliang Li, Agustinus Kristiadi, Pascal Poupart
Abstract: Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training like in standard RLHF methods. However, they rely on a reward model to score each candidate token generated by the language model at inference, incurring significant test-time overhead. Additionally, the reward model is usually only trained to score full sequences, which can lead to sub-optimal choices for partial sequences. In this work, we present a novel reward model architecture that is trained, using a Bradley-Terry loss, to prefer the optimal expansion of a sequence with just a \emph{single call} to the reward model at each step of the generation process. That is, a score for all possible candidate tokens is generated simultaneously, leading to efficient inference. We theoretically analyze various RGTG reward models and demonstrate that prior techniques prefer sub-optimal sequences compared to our method during inference. Empirically, our reward model leads to significantly faster inference than other RGTG methods. It requires fewer calls to the reward model and performs competitively compared to previous RGTG and offline RLHF methods.
Authors: So Yeon Min, Yue Wu, Jimin Sun, Max Kaufmann, Fahim Tajwar, Yonatan Bisk, Ruslan Salakhutdinov
Abstract: Human intelligence involves metacognitive abilities like self-regulation, recognizing limitations, and seeking assistance only when needed. While LLM Agents excel in many domains, they often lack this awareness. Overconfident agents risk catastrophic failures, while those that seek help excessively hinder efficiency. A key challenge is enabling agents with a limited intervention budget $C$ is to decide when to request assistance. In this paper, we propose an offline framework that trains a "helper" policy to request interventions, such as more powerful models or test-time compute, by combining LLM-based process reward models (PRMs) with tabular reinforcement learning. Using state transitions collected offline, we score optimal intervention timing with PRMs and train the helper model on these labeled trajectories. This offline approach significantly reduces costly intervention calls during training. Furthermore, the integration of PRMs with tabular RL enhances robustness to off-policy data while avoiding the inefficiencies of deep RL. We empirically find that our method delivers optimal helper behavior.
Authors: Tal Haklay, Hadas Orgad, David Bau, Aaron Mueller, Yonatan Belinkov
Abstract: A widely used strategy to discover and understand language model mechanisms is circuit analysis. A circuit is a minimal subgraph of a model's computation graph that executes a specific task. We identify a gap in existing circuit discovery methods: they assume circuits are position-invariant, treating model components as equally relevant across input positions. This limits their ability to capture cross-positional interactions or mechanisms that vary across positions. To address this gap, we propose two improvements to incorporate positionality into circuits, even on tasks containing variable-length examples. First, we extend edge attribution patching, a gradient-based method for circuit discovery, to differentiate between token positions. Second, we introduce the concept of a dataset schema, which defines token spans with similar semantics across examples, enabling position-aware circuit discovery in datasets with variable length examples. We additionally develop an automated pipeline for schema generation and application using large language models. Our approach enables fully automated discovery of position-sensitive circuits, yielding better trade-offs between circuit size and faithfulness compared to prior work.
Authors: Brian Formento, Chuan Sheng Foo, See-Kiong Ng
Abstract: A fundamental issue in deep learning has been adversarial robustness. As these systems have scaled, such issues have persisted. Currently, large language models (LLMs) with billions of parameters suffer from adversarial attacks just like their earlier, smaller counterparts. However, the threat models have changed. Previously, having gray-box access, where input embeddings or output logits/probabilities were visible to the user, might have been reasonable. However, with the introduction of closed-source models, no information about the model is available apart from the generated output. This means that current black-box attacks can only utilize the final prediction to detect if an attack is successful. In this work, we investigate and demonstrate the potential of attack guidance, akin to using output probabilities, while having only black-box access in a classification setting. This is achieved through the ability to elicit confidence from the model. We empirically show that the elicited confidence is calibrated and not hallucinated for current LLMs. By minimizing the elicited confidence, we can therefore increase the likelihood of misclassification. Our new proposed paradigm demonstrates promising state-of-the-art results on three datasets across two models (LLaMA-3-8B-Instruct and Mistral-7B-Instruct-V0.3) when comparing our technique to existing hard-label black-box attack methods that introduce word-level substitutions.
Authors: Junde Wu, Jiayuan Zhu, Yuyuan Liu
Abstract: We introduce Agentic Reasoning, a framework that enhances large language model (LLM) reasoning by integrating external tool-using agents. Unlike conventional LLM-based reasoning approaches, which rely solely on internal inference, Agentic Reasoning dynamically engages web search, code execution, and structured reasoning-context memory to solve complex problems requiring deep research and multi-step logical deduction. Our framework introduces the Mind Map agent, which constructs a structured knowledge graph to track logical relationships, improving deductive reasoning. Additionally, the integration of web-search and coding agents enables real-time retrieval and computational analysis, enhancing reasoning accuracy and decision-making. Evaluations on PhD-level scientific reasoning (GPQA) and domain-specific deep research tasks demonstrate that our approach significantly outperforms existing models, including leading retrieval-augmented generation (RAG) systems and closed-source LLMs. Moreover, our results indicate that agentic reasoning improves expert-level knowledge synthesis, test-time scalability, and structured problem-solving. The code is at: https://github.com/theworldofagents/Agentic-Reasoning.
URLs: https://github.com/theworldofagents/Agentic-Reasoning.
Authors: Ruiyang Ren, Yuhao Wang, Junyi Li, Jinhao Jiang, Wayne Xin Zhao, Wenjie Wang, Tat-Seng Chua
Abstract: In the era of vast digital information, the sheer volume and heterogeneity of available information present significant challenges for intricate information seeking. Users frequently face multistep web search tasks that involve navigating vast and varied data sources. This complexity demands every step remains comprehensive, accurate, and relevant. However, traditional search methods often struggle to balance the need for localized precision with the broader context required for holistic understanding, leaving critical facets of intricate queries underexplored. In this paper, we introduce an LLM-based search assistant that adopts a new information seeking paradigm with holistically guided Monte Carlo tree search (HG-MCTS). We reformulate the task as a progressive information collection process with a knowledge memory and unite an adaptive checklist with multi-perspective reward modeling in MCTS. The adaptive checklist provides explicit sub-goals to guide the MCTS process toward comprehensive coverage of complex user queries. Simultaneously, our multi-perspective reward modeling offers both exploration and retrieval rewards, along with progress feedback that tracks completed and remaining sub-goals, refining the checklist as the tree search progresses. By striking a balance between localized tree expansion and global guidance, HG-MCTS reduces redundancy in search paths and ensures that all crucial aspects of an intricate query are properly addressed. Extensive experiments on real-world intricate information seeking tasks demonstrate that HG-MCTS acquires thorough knowledge collections and delivers more accurate final responses compared with existing baselines.
Authors: Wonjun Lee, Doehyeon Lee, Eugene Choi, Sangyoon Yu, Ashkan Yousefpour, Haon Park, Bumsub Ham, Suhyun Kim
Abstract: Current Vision Language Models (VLMs) remain vulnerable to malicious prompts that induce harmful outputs. Existing safety benchmarks for VLMs primarily rely on automated evaluation methods, but these methods struggle to detect implicit harmful content or produce inaccurate evaluations. Therefore, we found that existing benchmarks have low levels of harmfulness, ambiguous data, and limited diversity in image-text pair combinations. To address these issues, we propose the ELITE {\em benchmark}, a high-quality safety evaluation benchmark for VLMs, underpinned by our enhanced evaluation method, the ELITE {\em evaluator}. The ELITE evaluator explicitly incorporates a toxicity score to accurately assess harmfulness in multimodal contexts, where VLMs often provide specific, convincing, but unharmful descriptions of images. We filter out ambiguous and low-quality image-text pairs from existing benchmarks using the ELITE evaluator and generate diverse combinations of safe and unsafe image-text pairs. Our experiments demonstrate that the ELITE evaluator achieves superior alignment with human evaluations compared to prior automated methods, and the ELITE benchmark offers enhanced benchmark quality and diversity. By introducing ELITE, we pave the way for safer, more robust VLMs, contributing essential tools for evaluating and mitigating safety risks in real-world applications.
Authors: Iason Ioannis Panagos, Giorgos Sfikas, Christophoros Nikou
Abstract: Visual speech recognition (VSR), which decodes spoken words from video data, offers significant benefits, particularly when audio is unavailable. However, the high dimensionality of video data leads to prohibitive computational costs that demand powerful hardware, limiting VSR deployment on resource-constrained devices. This work addresses this limitation by developing lightweight VSR architectures. Leveraging efficient operation design paradigms, we create compact yet powerful models with reduced resource requirements and minimal accuracy loss. We train and evaluate our models on a large-scale public dataset for recognition of words from video sequences, demonstrating their effectiveness for practical applications. We also conduct an extensive array of ablative experiments to thoroughly analyze the size and complexity of each model. Code and trained models will be made publicly available.
Authors: Liang Ze Wong
Abstract: In this paper, we investigate the ability of single-layer attention-only transformers (i.e. attention layers) to memorize facts contained in databases from a linear-algebraic perspective. We associate with each database a 3-tensor, propose the rank of this tensor as a measure of the size of the database, and provide bounds on the rank in terms of properties of the database. We also define a 3-tensor corresponding to an attention layer, and empirically demonstrate the relationship between its rank and database rank on a dataset of toy models and random databases. By highlighting the roles played by the value-output and query-key weights, and the effects of argmax and softmax on rank, our results shed light on the `additive motif' of factual recall in transformers, while also suggesting a way of increasing layer capacity without increasing the number of parameters.
Authors: Tushar Pandey, Ara Ghukasyan, Oktay Goktas, Santosh Kumar Radha
Abstract: Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, yet their performance is highly dependent on the prompting strategy and model scale. While reinforcement learning and fine-tuning have been deployed to boost reasoning, these approaches incur substantial computational and data overhead. In this work, we introduce Adaptive Graph of Thoughts (AGoT), a dynamic, graph-based inference framework that enhances LLM reasoning solely at test time. Rather than relying on fixed-step methods like Chain of Thought (CoT) or Tree of Thoughts (ToT), AGoT recursively decomposes complex queries into structured subproblems, forming an dynamic directed acyclic graph (DAG) of interdependent reasoning steps. By selectively expanding only those subproblems that require further analysis, AGoT unifies the strengths of chain, tree, and graph paradigms into a cohesive framework that allocates computation where it is most needed. We validate our approach on diverse benchmarks spanning multi-hop retrieval, scientific reasoning, and mathematical problem-solving, achieving up to 46.2% improvement on scientific reasoning tasks (GPQA) - comparable to gains achieved through computationally intensive reinforcement learning approaches and outperforming state-of-the-art iterative approaches. These results suggest that dynamic decomposition and structured recursion offer a scalable, cost-effective alternative to post-training modifications, paving the way for more robust, general-purpose reasoning in LLMs.
Authors: Thierry Bossy, Julien Vignoud, Tahseen Rabbani, Juan R. Troncoso Pastoriza, Martin Jaggi
Abstract: Federated learning (FL) is a popular paradigm for collaborative training which avoids direct data exposure between clients. However, data privacy issues still remain: FL-trained large language models are capable of memorizing and completing phrases and sentences contained in training data when given with their prefixes. Thus, it is possible for adversarial and honest-but-curious clients to recover training data of other participants simply through targeted prompting. In this work, we demonstrate that a popular and simple fine-tuning strategy, low-rank adaptation (LoRA), reduces memorization during FL up to a factor of 10. We study this effect by performing a medical question-answering fine-tuning task and injecting multiple replicas of out-of-distribution sensitive sequences drawn from an external clinical dataset. We observe a reduction in memorization for a wide variety of Llama 2 and 3 models, and find that LoRA can reduce memorization in centralized learning as well. Furthermore, we show that LoRA can be combined with other privacy-preserving techniques such as gradient clipping and Gaussian noising, secure aggregation, and Goldfish loss to further improve record-level privacy while maintaining performance.
Authors: Rohit Saxena, Aryo Pradipta Gema, Pasquale Minervini
Abstract: Understanding time from visual representations is a fundamental cognitive skill, yet it remains a challenge for multimodal large language models (MLLMs). In this work, we investigate the capabilities of MLLMs in interpreting time and date through analogue clocks and yearly calendars. To facilitate this, we curated a structured dataset comprising two subsets: 1) $\textit{ClockQA}$, which comprises various types of clock styles$-$standard, black-dial, no-second-hand, Roman numeral, and arrow-hand clocks$-$paired with time related questions; and 2) $\textit{CalendarQA}$, which consists of yearly calendar images with questions ranging from commonly known dates (e.g., Christmas, New Year's Day) to computationally derived ones (e.g., the 100th or 153rd day of the year). We aim to analyse how MLLMs can perform visual recognition, numerical reasoning, and temporal inference when presented with time-related visual data. Our evaluations show that despite recent advancements, reliably understanding time remains a significant challenge for MLLMs.
Authors: Kush R. Varshney
Abstract: The Parry-Lord oral-formulaic theory was a breakthrough in understanding how oral narrative poetry is learned, composed, and transmitted by illiterate bards. In this paper, we provide an annotated reading of the mechanism underlying this theory from the lens of large language models (LLMs) and generative artificial intelligence (AI). We point out the the similarities and differences between oral composition and LLM generation, and comment on the implications to society and AI policy.
Authors: Parjanya Prajakta Prashant, Kaustubh Ponkshe, Babak Salimi
Abstract: Large language models (LLMs) demonstrate impressive capabilities across many tasks yet risk reproducing copyrighted content verbatim, raising legal and ethical concerns. Although methods like differential privacy or neuron editing can reduce memorization, they typically require costly retraining or direct access to model weights and may degrade performance. To address these challenges, we propose TokenSwap, a lightweight, post-hoc approach that replaces the probabilities of grammar-related tokens with those from a small auxiliary model (e.g., DistilGPT-2). We run extensive experiments on commercial grade models such as Pythia-6.9b and LLaMA-3-8b and demonstrate that our method effectively reduces well-known cases of memorized generation by upto 10x with little to no impact on downstream tasks. Our approach offers a uniquely accessible and effective solution to users of real-world systems.
Authors: Jonas Geiping, Sean McLeish, Neel Jain, John Kirchenbauer, Siddharth Singh, Brian R. Bartoldson, Bhavya Kailkhura, Abhinav Bhatele, Tom Goldstein
Abstract: We study a novel language model architecture that is capable of scaling test-time computation by implicitly reasoning in latent space. Our model works by iterating a recurrent block, thereby unrolling to arbitrary depth at test-time. This stands in contrast to mainstream reasoning models that scale up compute by producing more tokens. Unlike approaches based on chain-of-thought, our approach does not require any specialized training data, can work with small context windows, and can capture types of reasoning that are not easily represented in words. We scale a proof-of-concept model to 3.5 billion parameters and 800 billion tokens. We show that the resulting model can improve its performance on reasoning benchmarks, sometimes dramatically, up to a computation load equivalent to 50 billion parameters.
Authors: Jan Ludziejewski, Maciej Pi\'oro, Jakub Krajewski, Maciej Stefaniak, Micha{\l} Krutul, Jan Ma{\l}a\'snicki, Marek Cygan, Piotr Sankowski, Kamil Adamczewski, Piotr Mi{\l}o\'s, Sebastian Jaszczur
Abstract: Mixture of Experts (MoE) architectures have significantly increased computational efficiency in both research and real-world applications of large-scale machine learning models. However, their scalability and efficiency under memory constraints remain relatively underexplored. In this work, we present joint scaling laws for dense and MoE models, incorporating key factors such as the number of active parameters, dataset size, and the number of experts. Our findings provide a principled framework for selecting the optimal MoE configuration under fixed memory and compute budgets. Surprisingly, we show that MoE models can be more memory-efficient than dense models, contradicting conventional wisdom. To derive and validate the theoretical predictions of our scaling laws, we conduct over 280 experiments with up to 2.7B active parameters and up to 5B total parameters. These results offer actionable insights for designing and deploying MoE models in practical large-scale training scenarios.
Authors: Lasha Abzianidze
Abstract: In this position paper, we propose a reasoning framework that can model the reasoning process underlying natural language inferences. The framework is based on the semantic tableau method, a well-studied proof system in formal logic. Like the semantic tableau, the framework is driven by refutation -- something is proved if and only if its counterexample was not refuted. Despite being rooted in formal logic, the framework shares similarities with the mental models, a theory on the psychology of reasoning. We will show how the reasoning framework can facilitate the collection of comprehensive and structured explanations for existing naturalistic inference problems. To make the suggestion more concrete, we propose a method of semi-automatically obtaining structured explanations from the formal proofs of a reliable and high-performing logic-based inference system. Taking advantage of the in-depth information available in the generated formal proofs, we show how it can be used to define natural language reasoning tasks with structured explanations. The proposed tasks can be ordered according to difficulty defined in terms of the granularity of explanations. We argue that the tasks that contain a natural sketch of the proofs will suffer from substantially fewer shortcomings than the existing explainable reasoning tasks (or datasets).
Authors: Arian Askari, Roxana Petcu, Chuan Meng, Mohammad Aliannejadi, Amin Abolghasemi, Evangelos Kanoulas, Suzan Verberne
Abstract: Identifying user intents in information-seeking dialogs is crucial for a system to meet user's information needs. Intent prediction (IP) is challenging and demands sufficient dialogs with human-labeled intents for training. However, manually annotating intents is resource-intensive. While large language models (LLMs) have been shown to be effective in generating synthetic data, there is no study on using LLMs to generate intent-aware information-seeking dialogs. In this paper, we focus on leveraging LLMs for zero-shot generation of large-scale, open-domain, and intent-aware information-seeking dialogs. We propose SOLID, which has novel self-seeding and multi-intent self-instructing schemes. The former improves the generation quality by using the LLM's own knowledge scope to initiate dialog generation; the latter prompts the LLM to generate utterances sequentially, and mitigates the need for manual prompt design by asking the LLM to autonomously adapt its prompt instruction when generating complex multi-intent utterances. Furthermore, we propose SOLID-RL, which is further trained to generate a dialog in one step on the data generated by SOLID. We propose a length-based quality estimation mechanism to assign varying weights to SOLID-generated dialogs based on their quality during the training process of SOLID-RL. We use SOLID and SOLID-RL to generate more than 300k intent-aware dialogs, surpassing the size of existing datasets. Experiments show that IP methods trained on dialogs generated by SOLID and SOLID-RL achieve better IP quality than ones trained on human-generated dialogs.
Authors: Kevin Wu, Eric Wu, James Zou
Abstract: Retrieval augmented generation (RAG) is frequently used to mitigate hallucinations and provide up-to-date knowledge for large language models (LLMs). However, given that document retrieval is an imprecise task and sometimes results in erroneous or even harmful content being presented in context, this raises the question of how LLMs handle retrieved information: If the provided content is incorrect, does the model know to ignore it, or does it recapitulate the error? Conversely, when the model's initial response is incorrect, does it always know to use the retrieved information to correct itself, or does it insist on its wrong prior response? To answer this, we curate a dataset of over 1200 questions across six domains (e.g., drug dosages, Olympic records, locations) along with content relevant to answering each question. We further apply precise perturbations to the answers in the content that range from subtle to blatant errors. We benchmark six top-performing LLMs, including GPT-4o, on this dataset and find that LLMs are susceptible to adopting incorrect retrieved content, overriding their own correct prior knowledge over 60% of the time. However, the more unrealistic the retrieved content is (i.e. more deviated from truth), the less likely the model is to adopt it. Also, the less confident a model is in its initial response (via measuring token probabilities), the more likely it is to adopt the information in the retrieved content. We exploit this finding and demonstrate simple methods for improving model accuracy where there is conflicting retrieved content. Our results highlight a difficult task and benchmark for LLMs -- namely, their ability to correctly discern when it is wrong in light of correct retrieved content and to reject cases when the provided content is incorrect.
Authors: Huanxuan Liao, Shizhu He, Yao Xu, Yuanzhe Zhang, Yanchao Hao, Shengping Liu, Kang Liu, Jun Zhao
Abstract: Large language models (LLMs) have acquired the ability to solve general tasks by utilizing instruction finetuning (IFT). However, IFT still relies heavily on instance training of extensive task data, which greatly limits the adaptability of LLMs to real-world scenarios where labeled task instances are scarce and broader task generalization becomes paramount. Contrary to LLMs, humans acquire skills and complete tasks not merely through repeated practice but also by understanding and following instructional guidelines. This paper is dedicated to simulating human learning to address the shortcomings of instance training, focusing on instruction learning to enhance cross-task generalization. Within this context, we introduce Task Adapters Generation from Instructions (TAGI), which automatically constructs the task-specific model in a parameter generation manner based on the given task instructions without retraining for unseen tasks. Specifically, we utilize knowledge distillation to enhance the consistency between TAGI developed through Learning with Instruction and task-specific models developed through Training with Instance, by aligning the labels, output logits, and adapter parameters between them. TAGI is endowed with cross-task generalization capabilities through a two-stage training process that includes hypernetwork pretraining and finetuning. We evaluate TAGI on the Super-Natural Instructions and P3 datasets. The experimental results demonstrate that TAGI can match or even outperform traditional meta-trained models and other hypernetwork models, while significantly reducing computational requirements.
Authors: Jingyi Sun, Pepa Atanasova, Isabelle Augenstein
Abstract: Explaining the decision-making process of machine learning models is crucial for ensuring their reliability and transparency for end users. One popular explanation form highlights key input features, such as i) tokens (e.g., Shapley Values and Integrated Gradients), ii) interactions between tokens (e.g., Bivariate Shapley and Attention-based methods), or iii) interactions between spans of the input (e.g., Louvain Span Interactions). However, these explanation types have only been studied in isolation, making it difficult to judge their respective applicability. To bridge this gap, we develop a unified framework that facilitates an automated and direct comparison between highlight and interactive explanations comprised of four diagnostic properties. We conduct an extensive analysis across these three types of input feature explanations -- each utilizing three different explanation techniques -- across two datasets and two models, and reveal that each explanation has distinct strengths across the different diagnostic properties. Nevertheless, interactive span explanations outperform other types of input feature explanations across most diagnostic properties. Despite being relatively understudied, our analysis underscores the need for further research to improve methods generating these explanation types. Additionally, integrating them with other explanation types that perform better in certain characteristics could further enhance their overall effectiveness.
Authors: Yan Meng, Di Wu, Christof Monz
Abstract: The massive amounts of web-mined parallel data contain large amounts of noise. Semantic misalignment, as the primary source of the noise, poses a challenge for training machine translation systems. In this paper, we first introduce a process for simulating misalignment controlled by semantic similarity, which closely resembles misaligned sentences in real-world web-crawled corpora. Under our simulated misalignment noise settings, we quantitatively analyze its impact on machine translation and demonstrate the limited effectiveness of widely used pre-filters for noise detection. This underscores the necessity of more fine-grained ways to handle hard-to-detect misalignment noise. With an observation of the increasing reliability of the model's self-knowledge for distinguishing misaligned and clean data at the token level, we propose self-correction, an approach that gradually increases trust in the model's self-knowledge to correct the training supervision. Comprehensive experiments show that our method significantly improves translation performance both in the presence of simulated misalignment noise and when applied to real-world, noisy web-mined datasets, across a range of translation tasks.
Authors: Yusuke Ide, Yuto Nishida, Justin Vasselli, Miyu Oba, Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe
Abstract: The grammatical knowledge of language models (LMs) is often measured using a benchmark of linguistic minimal pairs, where the LMs are presented with a pair of acceptable and unacceptable sentences and required to judge which is more acceptable. Conventional approaches directly compare sentence probabilities assigned by LMs, but recent large language models (LLMs) are trained to perform tasks via prompting, and thus, the raw probabilities they assign may not fully reflect their grammatical knowledge. In this study, we attempt to derive more accurate acceptability judgments from LLMs using prompts and templates. Through extensive experiments in English and Chinese, we compare nine judgment methods and find two of them, a probability readout method -- in-template LP and a prompt-based method -- Yes/No probability computing, achieve higher accuracy than the conventional ones. Our analysis reveals that these methods excel in different linguistic phenomena, suggesting they access different aspects of LLMs' knowledge. We also find that ensembling the two methods outperforms single methods. Consequently, we recommend these techniques, either individually or ensembled, as more effective alternatives to conventional approaches for assessing grammatical knowledge in LLMs.
Authors: Max Glockner, Yufang Hou, Preslav Nakov, Iryna Gurevych
Abstract: Health-related misinformation claims often falsely cite a credible biomedical publication as evidence. These publications only superficially seem to support the false claim, when logical fallacies are applied. In this work, we aim to detect and to highlight such fallacies, which requires assessing the exact content of the misrepresented publications. To achieve this, we introduce MissciPlus, an extension of the fallacy detection dataset Missci. MissciPlus extends Missci by grounding the applied fallacies in real-world passages from misrepresented studies. This creates a realistic test-bed for detecting and verbalizing fallacies under real-world input conditions, and enables new and realistic passage-retrieval tasks. MissciPlus is the first logical fallacy dataset which pairs the real-world misrepresented evidence with incorrect claims, identical to the input to evidence-based fact-checking models. With MissciPlus, we i) benchmark retrieval models in identifying passages that support claims only with fallacious reasoning, ii) evaluate how well LLMs verbalize fallacious reasoning based on misrepresented scientific passages, and iii) assess the effectiveness of fact-checking models in refuting claims that misrepresent biomedical research. Our findings show that current fact-checking models struggle to use misrepresented scientific passages to refute misinformation. Moreover, these passages can mislead LLMs into accepting false claims as true.
Authors: Yeonjun In, Sungchul Kim, Ryan A. Rossi, Md Mehrab Tanjim, Tong Yu, Ritwik Sinha, Chanyoung Park
Abstract: The retrieval augmented generation (RAG) framework addresses an ambiguity in user queries in QA systems by retrieving passages that cover all plausible interpretations and generating comprehensive responses based on the passages. However, our preliminary studies reveal that a single retrieval process often suffers from low quality results, as the retrieved passages frequently fail to capture all plausible interpretations. Although the iterative RAG approach has been proposed to address this problem, it comes at the cost of significantly reduced efficiency. To address these issues, we propose the diversify-verify-adapt (DIVA) framework. DIVA first diversifies the retrieved passages to encompass diverse interpretations. Subsequently, DIVA verifies the quality of the passages and adapts the most suitable approach tailored to their quality. This approach improves the QA systems accuracy and robustness by handling low quality retrieval issue in ambiguous questions, while enhancing efficiency.
Authors: Huanxuan Liao, Shizhu He, Yao Xu, Yuanzhe Zhang, Kang Liu, Jun Zhao
Abstract: In this paper, we propose $\textbf{Ne}$ural-$\textbf{Sy}$mbolic $\textbf{C}$ollaborative $\textbf{D}$istillation ($\textbf{NesyCD}$), a novel knowledge distillation method for learning the complex reasoning abilities of Large Language Models (LLMs, e.g., \textgreater 13B). We argue that complex reasoning tasks are difficult for Small Language Models (SLMs, e.g., $\leq$ 7B), as these tasks demand not only general cognitive abilities but also specialized knowledge, which is often sparse and difficult for these neural-based SLMs to effectively capture. Therefore, NesyCD distills the general capabilities and specialized knowledge in LLMs using different manners. On the one hand, we distill only general abilities from teacher LLMs into the student SLMs of parameterized neural networks. On the other hand, for the specialized abilities and uncommon knowledge of a complex reasoning task, we employ a symbolic knowledge distillation approach to obtain and store the specialized knowledge within a symbolic knowledge base (KB). By decoupling general and specialized capabilities, the proposed NesyCD can achieve superior performance cost-effectively, utilizing smaller models and blending parameterized neural networks with symbolic KB. Moreover, the specialized KB generalizes well and is comprehended and manipulated by humans. Our experiments show that NesyCD significantly boosts SLMs' complex reasoning performance on in-domain (BBH, GSM8K) and out-of-domain (AGIEval, ARC) datasets. Notably, our approach enabled the LLaMA3-8B and Qwen2-7B to surpass GPT-3.5-turbo in performance and come close to matching LLaMA3-70B, despite the latter having nine times more parameters. Our code will be available at https://github.com/Xnhyacinth/NesyCD.
Authors: Tongxuan Liu, Wenjiang Xu, Weizhe Huang, Yuting Zeng, Jiaxing Wang, Xingyu Wang, Hailong Yang, Jing Li
Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks but their performance in complex logical reasoning tasks remains unsatisfactory. Although some prompting methods, such as Chain-of-Thought, can improve the reasoning ability of LLMs to some extent, they suffer from an unfaithful issue where derived conclusions may not align with the generated reasoning chain. To address this issue, some studies employ the approach of propositional logic to further enhance logical reasoning abilities of LLMs. However, the potential omissions in the extraction of logical expressions in these methods can cause information loss in the logical reasoning process, thereby generating incorrect results. To this end, we propose Logic-of-Thought (LoT) prompting which employs propositional logic to generate expanded logical information descriptions and utilizes them as an additional augmentation to original contexts, thereby ensuring information completeness and enhancing logical reasoning ability. LoT is orthogonal to existing prompting methods and can be seamlessly integrated with them. Extensive experiments demonstrate that LoT boosts the performance of various prompting methods with a striking margin across five logical reasoning tasks. In particular, LoT enhances Chain-of-Thought's performance on the ReClor dataset by +4.35%, improves Chain-of-Thought with Self-Consistency's performance on the RuleTaker dataset by +3.52%, and boosts performance of Tree-of-Thoughts on the ProofWriter dataset by +8%.
Authors: Jin Hwa Lee, Thomas Jiralerspong, Lei Yu, Yoshua Bengio, Emily Cheng
Abstract: By virtue of linguistic compositionality, few syntactic rules and a finite lexicon can generate an unbounded number of sentences. That is, language, though seemingly high-dimensional, can be explained using relatively few degrees of freedom. An open question is whether contemporary language models (LMs) reflect the intrinsic simplicity of language that is enabled by compositionality. We take a geometric view of this problem by relating the degree of compositionality in a dataset to the intrinsic dimension (ID) of its representations under an LM, a measure of feature complexity. We find not only that the degree of dataset compositionality is reflected in representations' ID, but that the relationship between compositionality and geometric complexity arises due to learned linguistic features over training. Finally, our analyses reveal a striking contrast between nonlinear and linear dimensionality, showing they respectively encode semantic and superficial aspects of linguistic composition.
Authors: Chongyu Fan, Jiancheng Liu, Licong Lin, Jinghan Jia, Ruiqi Zhang, Song Mei, Sijia Liu
Abstract: This work studies the problem of large language model (LLM) unlearning, aiming to remove unwanted data influences (e.g., copyrighted or harmful content) while preserving model utility. Despite the increasing demand for unlearning, a technically-grounded optimization framework is lacking. Gradient ascent (GA)-type methods, though widely used, are suboptimal as they reverse the learning process without controlling optimization divergence (i.e., deviation from the pre-trained state), leading to risks of over-forgetting and potential model collapse. Negative preference optimization (NPO) has been proposed to address this issue and is considered one of the state-of-the-art LLM unlearning approaches. In this work, we revisit NPO and identify another critical issue: reference model bias. This bias arises from using the reference model (i.e., the model prior to unlearning) to evaluate the unlearning success, which can compromise NPO's effectiveness. Specifically, it leads to (a) uneven allocation of optimization power across forget data with varying difficulty levels and (b) ineffective gradient weight smoothing during the early stages of unlearning optimization. To overcome these challenges, we propose a simple yet effective unlearning optimization framework, called SimNPO, showing that `simplicity' in removing the reliance on a reference model (through the lens of simple preference optimization) benefits unlearning. We provide deeper insights into SimNPO's advantages through an analysis based on mixtures of Markov chains. Extensive experiments further validate SimNPO's efficacy on benchmarks like TOFU and MUSE, as well as its robustness against relearning attacks. Codes are available at https://github.com/OPTML-Group/Unlearn-Simple.
Authors: Jiamu Zheng, Jinghuai Zhang, Tianyu Du, Xuhong Zhang, Jianwei Yin, Tao Lin
Abstract: Collaborative learning of large language models (LLMs) has emerged as a new paradigm for utilizing private data from different parties to guarantee efficiency and privacy. Meanwhile, Knowledge Editing (KE) for LLMs has also garnered increased attention due to its ability to manipulate the behaviors of LLMs explicitly, yet leaves the collaborative KE case (in which knowledge edits of multiple parties are aggregated in a privacy-preserving and continual manner) unexamined. To this end, this manuscript dives into the first investigation of collaborative KE, in which we start by carefully identifying the unique three challenges therein, including knowledge overlap, knowledge conflict, and knowledge forgetting. We then propose a non-destructive collaborative KE framework, COLLABEDIT, which employs a novel model merging mechanism to mimic the global KE behavior while preventing the severe performance drop. Extensive experiments on two canonical datasets demonstrate the superiority of COLLABEDIT compared to other destructive baselines, and results shed light on addressing three collaborative KE challenges and future applications. Our code is available at https://github.com/LINs-lab/CollabEdit.
Authors: Yuxiang Wang, Jianzhong Qi, Junhao Gan
Abstract: Question answering on free-form tables (a.k.a. TableQA) is a challenging task because of the flexible structure and complex schema of tables. Recent studies use Large Language Models (LLMs) for this task, exploiting their capability in understanding the questions and tabular data, which are typically given in natural language and contain many textual fields, respectively. While this approach has shown promising results, it overlooks the challenges brought by numerical values which are common in tabular data, and LLMs are known to struggle with such values. We aim to address this issue, and we propose a model named TabLaP that uses LLMs as a planner rather than an answer generator. This approach exploits LLMs' capability in multi-step reasoning while leaving the actual numerical calculations to a Python interpreter for accurate calculation. Recognizing the inaccurate nature of LLMs, we further make a first attempt to quantify the trustworthiness of the answers produced by TabLaP, such that users can use TabLaP in a regret-aware manner. Experimental results on two benchmark datasets show that TabLaP is substantially more accurate than the state-of-the-art models, improving the answer accuracy by 5.7% and 5.8% on the two datasets, respectively.
Authors: Nura Aljaafari, Danilo S. Carvalho, Andr\'e Freitas
Abstract: Understanding the internal mechanisms of large language models (LLMs) is integral to enhancing their reliability, interpretability, and inference processes. We present Constituent-Aware Pooling (CAP), a methodology designed to analyse how LLMs process compositional linguistic structures. Grounded in principles of compositionality, mechanistic interpretability, and information theory, CAP systematically intervenes in model activations through constituent-based pooling at various model levels. Our experiments on inverse definition modelling, hypernym and synonym prediction reveal critical insights into transformers' limitations in handling compositional abstractions. No specific layer integrates tokens into unified semantic representations based on their constituent parts. We observe fragmented information processing, which intensifies with model size, suggesting that larger models struggle more with these interventions and exhibit greater information dispersion. This fragmentation likely stems from transformers' training objectives and architectural design, preventing systematic and cohesive representations. Our findings highlight fundamental limitations in current transformer architectures regarding compositional semantics processing and model interpretability, underscoring the critical need for novel approaches in LLM design to address these challenges.
Authors: Nan Xu, Xuezhe Ma
Abstract: Interestingly, LLMs yet struggle with some basic tasks that humans find trivial to handle, e.g., counting the number of character r's in the word "strawberry". There are several popular conjectures (e.g., tokenization, architecture and training data) regarding the reason for deficiency of LLMs in simple word-based counting problems, sharing the similar belief that such failure stems from model pretraining hence probably inevitable during deployment. In this paper, we carefully design multiple evaluation settings to investigate validity of prevalent conjectures. Meanwhile, we measure transferability of advanced mathematical and coding reasoning capabilities from specialized LLMs to simple counting tasks. Although specialized LLMs suffer from counting problems as well, we find conjectures about inherent deficiency of LLMs invalid and further seek opportunities to elicit knowledge and capabilities from LLMs that are beneficial to counting tasks. Compared with strategies such as finetuning and in-context learning that are commonly adopted to enhance performance on new or challenging tasks, we show that engaging reasoning is the most robust and efficient way to help LLMs better perceive tasks with more accurate responses. We hope our conjecture validation design could provide insights into the study of future critical failure modes of LLMs. Based on challenges in transferring advanced capabilities to much simpler tasks, we call for more attention to model capability acquisition and evaluation. We also highlight the importance of cultivating consciousness of "reasoning before responding" during model pretraining.
Authors: Yoichi Ishibashi, Taro Yano, Masafumi Oyamada
Abstract: Large Language Models (LLMs) have shown remarkable performance improvements and are rapidly gaining adoption in industry. However, the methods for improving LLMs are still designed by humans, which restricts the invention of new model-improving algorithms to human expertise and imagination. To address this, we propose the Self-Developing framework, which enables LLMs to autonomously generate and learn model-improvement algorithms. In this framework, the seed model generates, applies, and learns model-improving algorithms, continuously improving both the seed model and the algorithms themselves. Among model-improving strategies, we focus on model merging algorithms. In mathematical reasoning tasks, Self-Developing discovers novel merging strategies and outperforms human-designed methods. On GSM8k, the discovered algorithms improve the seed model by 6% and surpass human-designed methods by 4.3%. Moreover, they exhibit strong transferability, achieving a 7.4% performance gain on out-of-domain models. These results suggest that LLMs can autonomously develop effective model-improvement techniques beyond human intuition.
Authors: Yifan Peng, Krishna C. Puvvada, Zhehuai Chen, Piotr Zelasko, He Huang, Kunal Dhawan, Ke Hu, Shinji Watanabe, Jagadeesh Balam, Boris Ginsburg
Abstract: Recent studies have augmented large language models (LLMs) with speech capabilities, leading to the development of speech language models (SpeechLMs). Earlier SpeechLMs focused on single-turn speech-based question answering (QA), where user input comprised a speech context and a text question. More recent studies have extended this to multi-turn conversations, though they often require complex, multi-stage supervised fine-tuning (SFT) with diverse data. Another critical challenge with SpeechLMs is catastrophic forgetting, where models optimized for speech tasks suffer significant degradation in text-only performance. To mitigate these issues, we propose a novel single-stage joint speech-text SFT approach on the low-rank adaptation (LoRA) of the LLM backbone. Our joint SFT combines text-only SFT data with three types of speech-related data: speech recognition and translation, speech-based QA, and mixed-modal SFT. Compared to previous SpeechLMs with 7B or 13B parameters, our 3B model demonstrates superior performance across various speech benchmarks while preserving the original capabilities on text-only tasks. Furthermore, our model shows emergent abilities of effectively handling previously unseen prompts and tasks, including multi-turn, mixed-modal inputs.
Authors: Langlin Huang, Mengyu Bu, Yang Feng
Abstract: Byte-based machine translation systems have shown significant potential in massively multilingual settings. Unicode encoding, which maps each character to specific byte(s), eliminates the emergence of unknown words, even in new languages. This avoids out-of-vocabulary risk in multilingual translation and enables broad language scalability. However, byte-level tokenization results in sequences that are hard to interpret due to limited semantic information per byte. Local contextualization has proven effective in assigning initial semantics to tokens, improving sentence comprehension. Nevertheless, variations in encoding rules across languages necessitate an adaptive approach for effective contextualization. To this end, we propose Mixture of Contextualization Experts (MoCE), adaptively selecting and mixing attention heads, which are treated as contextualization experts. This enhances the flexibility of contextualization scales and allows models to search for better contextualization combinations. Experiment results show that our method outperforms existing methods without extensive manual adjustment of hyper-parameters and surpasses subword-based models with fewer parameters in Ted-59 dataset. Our code is available at https://github.com/ictnlp/MoCE.
Authors: Yanxi Chen, Xuchen Pan, Yaliang Li, Bolin Ding, Jingren Zhou
Abstract: We propose two simple yet principled algorithms that enjoy provable scaling laws for the test-time compute of large language models (LLMs), which require a black-box LLM and nothing else (e.g., no external verifier or reward model) for a minimalistic implementation. (i) The first one is a two-stage knockout-style algorithm: given an input problem, it first generates multiple candidate solutions, and then aggregate them for a final output, via a knockout tournament where pairwise comparisons among the candidates are conducted. Assuming that the LLM can generate a correct solution with non-zero probability and do better than a random guess in comparing a pair of correct and incorrect solutions, we prove theoretically that the failure probability of this algorithm decays to zero exponentially or by a power law (depending on the specific way of scaling) as its test-time compute grows. (ii) The second one is a two-stage league-style algorithm, where each candidate solution is evaluated by its average win rate against multiple opponents, rather than eliminated upon loss to a single opponent. Under certain technical assumptions that are analogous to but more robust than those required by the knockout-style algorithm, we prove theoretically that the failure probability of the league-style algorithm also decays to zero exponentially as its test-time compute grows. Through extensive experiments with two challenging benchmarks, namely GPQA and MMLU-Pro, we validate the proposed theories and demonstrate the outstanding scaling properties of both algorithms.
Authors: Xinmeng Hou, Lingyue Fu, Chenhao Meng, Kounianhua Du, Wuqi Wang, Hai Hu
Abstract: Aspect-Opinion Pair Extraction (AOPE) and Aspect Sentiment Triplet Extraction (ASTE) have drawn growing attention in NLP. However, most existing approaches extract aspects and opinions independently, optionally adding pairwise relations, often leading to error propagation and high time complexity. To address these challenges and being inspired by transition-based dependency parsing, we propose the first transition-based model for AOPE and ASTE that performs aspect and opinion extraction jointly, which also better captures position-aware aspect-opinion relations and mitigates entity-level bias. By integrating contrastive-augmented optimization, our model delivers more accurate action predictions and jointly optimizes separate subtasks in linear time. Extensive experiments on 4 commonly used ASTE/AOPE datasets show that, while performing worse when trained on a single dataset than some previous models, our model achieves the best performance on both ASTE and AOPE if trained on combined datasets, outperforming the strongest previous models in F1-measures (often by a large margin). We hypothesize that this is due to our model's ability to learn transition actions from multiple datasets and domains. Our code is available at https://anonymous.4open.science/r/trans_aste-8FCF.
Authors: Junwei Yin, Min Gao, Kai Shu, Wentao Li, Yinqiu Huang, Zongwei Wang
Abstract: The rapid proliferation of fake news on social media threatens social stability, creating an urgent demand for more effective detection methods. While many promising approaches have emerged, most rely on content analysis with limited semantic depth, leading to suboptimal comprehension of news content.To address this limitation, capturing broader-range semantics is essential yet challenging, as it introduces two primary types of noise: fully connecting sentences in news graphs often adds unnecessary structural noise, while highly similar but authenticity-irrelevant sentences introduce feature noise, complicating the detection process. To tackle these issues, we propose BREAK, a broad-range semantics model for fake news detection that leverages a fully connected graph to capture comprehensive semantics while employing dual denoising modules to minimize both structural and feature noise. The semantic structure denoising module balances the graph's connectivity by iteratively refining it between two bounds: a sequence-based structure as a lower bound and a fully connected graph as the upper bound. This refinement uncovers label-relevant semantic interrelations structures. Meanwhile, the semantic feature denoising module reduces noise from similar semantics by diversifying representations, aligning distinct outputs from the denoised graph and sequence encoders using KL-divergence to achieve feature diversification in high-dimensional space. The two modules are jointly optimized in a bi-level framework, enhancing the integration of denoised semantics into a comprehensive representation for detection. Extensive experiments across four datasets demonstrate that BREAK significantly outperforms existing fake news detection methods.
Authors: Zehong Wang, Sidney Liu, Zheyuan Zhang, Tianyi Ma, Chuxu Zhang, Yanfang Ye
Abstract: Graphs are ubiquitous structures found in numerous real-world applications, such as drug discovery, recommender systems, and social network analysis. To model graph-structured data, graph neural networks (GNNs) have become a popular tool. However, existing GNN architectures encounter challenges in cross-graph learning where multiple graphs have different feature spaces. To address this, recent approaches introduce text-attributed graphs (TAGs), where each node is associated with a textual description, which can be projected into a unified feature space using textual encoders. While promising, this method relies heavily on the availability of text-attributed graph data, which is difficult to obtain in practice. To bridge this gap, we propose a novel method named Topology-Aware Node description Synthesis (TANS), leveraging large language models (LLMs) to convert existing graphs into text-attributed graphs. The key idea is to integrate topological information into LLMs to explain how graph topology influences node semantics. We evaluate our TANS on text-rich, text-limited, and text-free graphs, demonstrating its applicability. Notably, on text-free graphs, our method significantly outperforms existing approaches that manually design node features, showcasing the potential of LLMs for preprocessing graph-structured data in the absence of textual information. The code and data are available at https://github.com/Zehong-Wang/TANS.
Authors: Behzad Shayegh, Hobie H. -B. Lee, Xiaodan Zhu, Jackie Chi Kit Cheung, Lili Mou
Abstract: We address unsupervised dependency parsing by building an ensemble of diverse existing models through post hoc aggregation of their output dependency parse structures. We observe that these ensembles often suffer from low robustness against weak ensemble components due to error accumulation. To tackle this problem, we propose an efficient ensemble-selection approach that considers error diversity and avoids error accumulation. Results demonstrate that our approach outperforms each individual model as well as previous ensemble techniques. Additionally, our experiments show that the proposed ensemble-selection method significantly enhances the performance and robustness of our ensemble, surpassing previously proposed strategies, which have not accounted for error diversity.
Authors: Santhosh Thottingal
Abstract: This paper introduces an approach to question answering over knowledge bases like Wikipedia and Wikidata by performing "question-to-question" matching and retrieval from a dense vector embedding store. Instead of embedding document content, we generate a comprehensive set of questions for each logical content unit using an instruction-tuned LLM. These questions are vector-embedded and stored, mapping to the corresponding content. Vector embedding of user queries are then matched against this question vector store. The highest similarity score leads to direct retrieval of the associated article content, eliminating the need for answer generation. Our method achieves high cosine similarity ( > 0.9 ) for relevant question pairs, enabling highly precise retrieval. This approach offers several advantages including computational efficiency, rapid response times, and increased scalability. We demonstrate its effectiveness on Wikipedia and Wikidata, including multimedia content through structured fact retrieval from Wikidata, opening up new pathways for multimodal question answering.
Authors: Giorgio Robino
Abstract: This study introduces Conversation Routines (CR), a structured prompt engineering framework for developing task-oriented dialog systems using Large Language Models (LLMs). While LLMs demonstrate remarkable natural language understanding capabilities, engineering them to reliably execute complex business workflows remains challenging. The proposed CR framework enables the development of Conversation Agentic Systems (CAS) through natural language specifications, embedding task-oriented logic within LLM prompts. This approach provides a systematic methodology for designing and implementing complex conversational workflows while maintaining behavioral consistency. We demonstrate the framework's effectiveness through two proof-of-concept implementations: a Train Ticket Booking System and an Interactive Troubleshooting Copilot. These case studies validate CR's capability to encode sophisticated behavioral patterns and decision logic while preserving natural conversational flexibility. Results show that CR enables domain experts to design conversational workflows in natural language while leveraging custom functions (tools) developed by software engineers, creating an efficient division of responsibilities where developers focus on core API implementation and domain experts handle conversation design. While the framework shows promise in accessibility and adaptability, we identify key challenges including computational overhead, non-deterministic behavior, and domain-specific logic optimization. Future research directions include CR evaluation methods based on prompt engineering frameworks driven by goal-oriented grading criteria, improving scalability for complex multi-agent interactions, and enhancing system robustness to address the identified limitations across diverse business applications.
Authors: Guangzhi Sun, Xiao Zhan, Shutong Feng, Philip C. Woodland, Jose Such
Abstract: Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption. Current LLM safety benchmarks often focus solely on the refusal of individual problematic queries, which overlooks the importance of the context where the query occurs and may cause undesired refusal of queries under safe contexts that diminish user experience. Addressing this gap, we introduce CASE-Bench, a Context-Aware SafEty Benchmark that integrates context into safety assessments of LLMs. CASE-Bench assigns distinct, formally described contexts to categorized queries based on Contextual Integrity theory. Additionally, in contrast to previous studies which mainly rely on majority voting from just a few annotators, we recruited a sufficient number of annotators necessary to ensure the detection of statistically significant differences among the experimental conditions based on power analysis. Our extensive analysis using CASE-Bench on various open-source and commercial LLMs reveals a substantial and significant influence of context on human judgments (p<0.0001 from a z-test), underscoring the necessity of context in safety evaluations. We also identify notable mismatches between human judgments and LLM responses, particularly in commercial models within safe contexts.
Authors: Menglong Cui, Pengzhi Gao, Wei Liu, Jian Luan, Bin Wang
Abstract: Large language models (LLMs) have shown continuously improving multilingual capabilities, and even small-scale open-source models have demonstrated rapid performance enhancement. In this paper, we systematically explore the abilities of open LLMs with less than ten billion parameters to handle multilingual machine translation (MT) tasks. We conduct comprehensive evaluations on six popular LLMs and find that models like Gemma2-9B exhibit impressive multilingual translation capabilities. We then introduce the Parallel-First Monolingual-Second (PFMS) data mixing strategy in the continual pretraining stage to further enhance the MT performance and present GemmaX2-28, a 9B model achieving top-tier multilingual translation performance across 28 languages. Specifically, GemmaX2-28 consistently outperforms the state-of-the-art (SOTA) models such as TowerInstruct and XALMA and achieves competitive performance with Google Translate and GPT-4-turbo.
Authors: Dylan Zhang, Qirun Dai, Hao Peng
Abstract: High-quality supervised fine-tuning (SFT) data are crucial for eliciting strong capabilities from pretrained large language models (LLMs). Typically, instructions are paired with multiple responses sampled from other LLMs, which are often out of the distribution of the target model to be fine-tuned. This, at scale, can lead to diminishing returns and even hurt the models' performance and robustness. We propose **GRAPE**, a novel SFT framework that accounts for the unique characteristics of the target model. For each instruction, it gathers responses from various LLMs and selects the one with the highest probability measured by the target model, indicating that it aligns most closely with the target model's pretrained distribution; it then proceeds with standard SFT training. We first evaluate GRAPE with a controlled experiment, where we sample various solutions for each question in UltraInteract from multiple models and fine-tune commonly used LMs like LLaMA3.1-8B, Mistral-7B, and Qwen2.5-7B on GRAPE-selected data. GRAPE significantly outperforms strong baselines, including distilling from the strongest model with an absolute gain of up to 13.8%, averaged across benchmarks, and training on 3x more data with a maximum performance improvement of 17.3%. GRAPE's strong performance generalizes to realistic settings. We experiment with the post-training data used for Tulu3 and Olmo-2. GRAPE outperforms strong baselines trained on 4.5 times more data by 6.1% and a state-of-the-art data selection approach by 3% on average performance. Remarkably, using 1/3 of the data and half the number of epochs, GRAPE enables LLaMA3.1-8B to surpass the performance of Tulu3-SFT by 3.5%.
Authors: Joseph Cho, Mrudang Mathur, Cyril Zakka, Dhamanpreet Kaur, Matthew Leipzig, Alex Dalal, Aravind Krishnan, Eubee Koo, Karen Wai, Cindy S. Zhao, Rohan Shad, Robyn Fong, Ross Wightman, Akshay Chaudhari, William Hiesinger
Abstract: Deep learning algorithms require extensive data to achieve robust performance. However, data availability is often restricted in the medical domain due to patient privacy concerns. Synthetic data presents a possible solution to these challenges. Recently, image generative models have found increasing use for medical applications but are often designed for singular medical specialties and imaging modalities, thus limiting their broader utility. To address this, we introduce MediSyn: a text-guided, latent diffusion model capable of generating synthetic images from 6 medical specialties and 10 image types. The synthetic images are validated by expert clinicians for alignment with their corresponding text prompts. Furthermore, a direct comparison of the synthetic images against the real images confirms that our model synthesizes novel images and, crucially, may preserve patient privacy. Finally, classifiers trained on a mixture of synthetic and real data achieve similar performance to those trained on twice the amount of real data. Our findings highlight the immense potential for generalist image generative models to accelerate algorithmic research and development in medicine.
Authors: Cong Lu, Shengran Hu, Jeff Clune
Abstract: Go-Explore is a powerful family of algorithms designed to solve hard-exploration problems built on the principle of archiving discovered states, and iteratively returning to and exploring from the most promising states. This approach has led to superhuman performance across a wide variety of challenging problems including Atari games and robotic control, but requires manually designing heuristics to guide exploration (i.e., determine which states to save and explore from, and what actions to consider next), which is time-consuming and infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE) which greatly extends the scope of the original Go-Explore by replacing these handcrafted heuristics with the intelligence and internalized human notions of interestingness captured by giant pretrained foundation models (FMs). This provides IGE with a human-like ability to instinctively identify how interesting or promising any new state is (e.g., discovering new objects, locations, or behaviors), even in complex environments where heuristics are hard to define. Moreover, IGE offers the exciting opportunity to recognize and capitalize on serendipitous discoveries -- states encountered during exploration that are valuable in terms of exploration, yet where what makes them interesting was not anticipated by the human user. We evaluate our algorithm on a diverse range of language and vision-based tasks that require search and exploration. Across these tasks, IGE strongly exceeds classic reinforcement learning and graph search baselines, and also succeeds where prior state-of-the-art FM agents like Reflexion completely fail. Overall, Intelligent Go-Explore combines the tremendous strengths of FMs and the powerful Go-Explore algorithm, opening up a new frontier of research into creating more generally capable agents with impressive exploration capabilities.
Authors: Fran\c{c}ois Roewer-Despr\'es, Jinyue Feng, Zining Zhu, Frank Rudzicz
Abstract: We present ACCORD, a framework and benchmark suite for disentangling the commonsense grounding and reasoning abilities of large language models (LLMs) through controlled, multi-hop counterfactuals. ACCORD introduces formal elements to commonsense reasoning to explicitly control and quantify reasoning complexity beyond the typical 1 or 2 hops. Uniquely, ACCORD can automatically generate benchmarks of arbitrary reasoning complexity, and so it scales with future LLM improvements. Benchmarking state-of-the-art LLMs -- including GPT-4o (2024-05-13), Llama-3-70B-Instruct, and Mixtral-8x22B-Instruct-v0.1 -- shows performance degrading to random chance with only moderate scaling, leaving substantial headroom for improvement. We release a leaderboard of the benchmark suite tested in this work, as well as code for automatically generating more complex benchmarks.
Authors: Yuzi Yan, Yibo Miao, Jialian Li, Yipin Zhang, Jian Xie, Zhijie Deng, Dong Yan
Abstract: Aligning large language models (LLMs) with human preferences has gained significant attention, with Proximal Policy Optimization (PPO) as a standard yet computationally expensive method and Direct Preference Optimization (DPO) as a more efficient alternative. While DPO offers simplicity, it remains underutilized in state-of-the-art LLMs, suggesting potential limitations. In this work, we revisit DPO, analyzing its theoretical foundations and empirical performance to bridge this gap. We identify three key properties, termed 3D properties, that emerge from DPO's learning process: Drastic drop in rejected response likelihood, Degradation into response suppression, and Dispersion effect on unseen responses. We show that these issues arise from DPO's optimization dynamics, where the interaction between chosen and rejected response gradients leads to instability. Our findings are supported by experiments on both a controlled toy model and real-world LLM tasks, including mathematical problem-solving and instruction following. To address these challenges, we propose simple regularization techniques that improve training stability and performance. Additionally, we examine how preference data distribution impacts DPO's effectiveness, offering insights into how alignment models handle out-of-domain (OOD) data. Our work connects these observations to broader research and provides a theoretical explanation for DPO's limitations. We hope these insights will guide future advancements in reward-model-free preference learning, bringing it closer to reward-model-based approaches.
Authors: Teun van der Weij, Felix Hofst\"atter, Ollie Jaffe, Samuel F. Brown, Francis Rhys Ward
Abstract: Trustworthy capability evaluations are crucial for ensuring the safety of AI systems, and are becoming a key component of AI regulation. However, the developers of an AI system, or the AI system itself, may have incentives for evaluations to understate the AI's actual capability. These conflicting interests lead to the problem of sandbagging, which we define as strategic underperformance on an evaluation. In this paper we assess sandbagging capabilities in contemporary language models (LMs). We prompt frontier LMs, like GPT-4 and Claude 3 Opus, to selectively underperform on dangerous capability evaluations, while maintaining performance on general (harmless) capability evaluations. Moreover, we find that models can be fine-tuned, on a synthetic dataset, to hide specific capabilities unless given a password. This behaviour generalizes to high-quality, held-out benchmarks such as WMDP. In addition, we show that both frontier and smaller models can be prompted or password-locked to target specific scores on a capability evaluation. We have mediocre success in password-locking a model to mimic the answers a weaker model would give. Overall, our results suggest that capability evaluations are vulnerable to sandbagging. This vulnerability decreases the trustworthiness of evaluations, and thereby undermines important safety decisions regarding the development and deployment of advanced AI systems.
Authors: Nan Xu, Fei Wang, Sheng Zhang, Hoifung Poon, Muhao Chen
Abstract: Motivated by in-context learning (ICL) capabilities of Large Language Models (LLMs), multimodal LLMs with additional visual modality are also exhibited with similar ICL abilities when multiple image-text pairs are provided as demonstrations. However, relatively less work has been done to investigate the principles behind how and why multimodal ICL works. We conduct a systematic and principled evaluation of multimodal ICL for models of different scales on a broad spectrum of new yet critical tasks. Through perturbations over different modality information, we show that modalities matter differently across tasks in multimodal ICL. Guided by task-specific modality impact, we recommend modality-driven demonstration strategies to boost ICL performance. We also find that models may follow inductive biases from multimodal ICL even if they are rarely seen in or contradict semantic priors from pretraining data. Our principled analysis provides a comprehensive way of understanding the role of demonstrations in multimodal in-context learning, and sheds light on effectively improving multimodal ICL on a wide range of tasks.
Authors: Zhengyuan Zhu, Daniel Lee, Hong Zhang, Sai Sree Harsha, Loic Feujio, Akash Maharaj, Yunyao Li
Abstract: Recent advancements in retrieval-augmented generation (RAG) have demonstrated impressive performance in the question-answering (QA) task. However, most previous works predominantly focus on text-based answers. While some studies address multimodal data, they still fall short in generating comprehensive multimodal answers, particularly for explaining concepts or providing step-by-step tutorials on how to accomplish specific goals. This capability is especially valuable for applications such as enterprise chatbots and settings such as customer service and educational systems, where the answers are sourced from multimodal data. In this paper, we introduce a simple and effective framework named MuRAR (Multimodal Retrieval and Answer Refinement). MuRAR enhances text-based answers by retrieving relevant multimodal data and refining the responses to create coherent multimodal answers. This framework can be easily extended to support multimodal answers in enterprise chatbots with minimal modifications. Human evaluation results indicate that multimodal answers generated by MuRAR are more useful and readable compared to plain text answers.
Authors: Yuan Tian, Tianyi Zhang
Abstract: Recent advances in large language models (LLMs) have transformed software development by automatically generating code from natural language. Yet challenges remain in generating fully correct code that aligns with user intent. Our study reveals that LLMs tend to pay less attention to user prompts as more code tokens are generated. We hypothesize that this attention dilution issue is an important reason for code generation errors. To mitigate this issue, we propose Selective Prompt Anchoring (SPA) to guide code LLMs to pay more attention to user intent when generating code. We evaluate SPA using six base LLMs across six benchmarks. Our results demonstrate that SPA enhances Pass@1 by up to 12.9%, consistently outperforming SOTA code generation methods in all settings. Our code is available at https://github.com/magic-YuanTian/Selective-Prompt-Anchoring.
URLs: https://github.com/magic-YuanTian/Selective-Prompt-Anchoring.
Authors: Kaiwen Zheng, Yongxin Chen, Hanzi Mao, Ming-Yu Liu, Jun Zhu, Qinsheng Zhang
Abstract: Masked diffusion models (MDMs) have emerged as a popular research topic for generative modeling of discrete data, thanks to their superior performance over other discrete diffusion models, and are rivaling the auto-regressive models (ARMs) for language modeling tasks. The recent effort in simplifying the masked diffusion framework further leads to alignment with continuous-space diffusion models and more principled training and sampling recipes. In this paper, however, we reveal that both training and sampling of MDMs are theoretically free from the time variable, arguably the key signature of diffusion models, and are instead equivalent to masked models. The connection on the sampling aspect is drawn by our proposed first-hitting sampler (FHS). Specifically, we show that the FHS is theoretically equivalent to MDMs' original generation process while significantly alleviating the time-consuming categorical sampling and achieving a 20$\times$ speedup. In addition, our investigation raises doubts about whether MDMs can truly beat ARMs in text generation. We identify, for the first time, an underlying numerical issue, even with the commonly used 32-bit floating-point precision, which results in inaccurate categorical sampling. We show that it lowers the effective temperature both theoretically and empirically, and the resulting decrease in token diversity makes previous evaluations, which assess the generation quality solely through the incomplete generative perplexity metric, somewhat unfair.
Authors: Mingjie Liu, Yun-Da Tsai, Wenfei Zhou, Haoxing Ren
Abstract: Despite the significant progress made in code generation with large language models, challenges persist, especially with hardware description languages such as Verilog. This paper first presents an analysis of fine-tuned LLMs on Verilog coding, with synthetic data from prior methods. We identify two main issues: difficulties in handling non-textual representations (Karnaugh maps, state-transition diagrams and waveforms) and significant variability during training with models randomly making "minor" mistakes. To address these limitations, we enhance data curation by creating correct-by-construction data targeting non-textual representations. Additionally, we introduce an automated framework that generates error reports from various model checkpoints and injects these errors into open-source code to create targeted code repair data. Our fine-tuned Starcoder2-15B outperforms prior state-of-the-art results by 3.8%, 10.9%, 6.6% for pass@1 on VerilogEval-Machine, VerilogEval-Human, and RTLLM.
Authors: Guobin Shen, Dongcheng Zhao, Yiting Dong, Xiang He, Yi Zeng
Abstract: As large language models (LLMs) become integral to various applications, ensuring both their safety and utility is paramount. Jailbreak attacks, which manipulate LLMs into generating harmful content, pose significant challenges to this balance. Existing defenses, such as prompt engineering and safety fine-tuning, often introduce computational overhead, increase inference latency, and lack runtime flexibility. Moreover, overly restrictive safety measures can degrade model utility by causing refusals of benign queries. In this paper, we introduce Jailbreak Antidote, a method that enables real-time adjustment of LLM safety preferences by manipulating a sparse subset of the model's internal states during inference. By shifting the model's hidden representations along a safety direction with varying strengths, we achieve flexible control over the safety-utility balance without additional token overhead or inference delays. Our analysis reveals that safety-related information in LLMs is sparsely distributed; adjusting approximately 5% of the internal state is as effective as modifying the entire state. Extensive experiments on nine LLMs (ranging from 2 billion to 72 billion parameters), evaluated against ten jailbreak attack methods and compared with six defense strategies, validate the effectiveness and efficiency of our approach. By directly manipulating internal states during reasoning, Jailbreak Antidote offers a lightweight, scalable solution that enhances LLM safety while preserving utility, opening new possibilities for real-time safety mechanisms in widely-deployed AI systems.
Authors: Yuzhe Yang, Yifei Zhang, Yan Hu, Yilin Guo, Ruoli Gan, Yueru He, Mingcong Lei, Xiao Zhang, Haining Wang, Qianqian Xie, Jimin Huang, Honghai Yu, Benyou Wang
Abstract: This paper introduces the UCFE: User-Centric Financial Expertise benchmark, an innovative framework designed to evaluate the ability of large language models (LLMs) to handle complex real-world financial tasks. UCFE benchmark adopts a hybrid approach that combines human expert evaluations with dynamic, task-specific interactions to simulate the complexities of evolving financial scenarios. Firstly, we conducted a user study involving 804 participants, collecting their feedback on financial tasks. Secondly, based on this feedback, we created our dataset that encompasses a wide range of user intents and interactions. This dataset serves as the foundation for benchmarking 11 LLMs services using the LLM-as-Judge methodology. Our results show a significant alignment between benchmark scores and human preferences, with a Pearson correlation coefficient of 0.78, confirming the effectiveness of the UCFE dataset and our evaluation approach. UCFE benchmark not only reveals the potential of LLMs in the financial domain but also provides a robust framework for assessing their performance and user satisfaction.
Authors: Mayank Nagda, Phil Ostheimer, Sophie Fellenz
Abstract: Topic models are a popular approach for extracting semantic information from large document collections. However, recent studies suggest that the topics generated by these models often do not align well with human intentions. Although metadata such as labels and authorship information are available, it has not yet been effectively incorporated into neural topic models. To address this gap, we introduce FANToM, a novel method to align neural topic models with both labels and authorship information. FANToM allows for the inclusion of this metadata when available, producing interpretable topics and author distributions for each topic. Our approach demonstrates greater expressiveness than conventional topic models by learning the alignment between labels, topics, and authors. Experimental results show that FANToM improves existing models in terms of both topic quality and alignment. Additionally, it identifies author interests and similarities.
Authors: Justin Deschenaux, Caglar Gulcehre
Abstract: Autoregressive (AR) Large Language Models (LLMs) have demonstrated significant success across numerous tasks. However, the AR modeling paradigm presents certain limitations; for instance, contemporary autoregressive LLMs are trained to generate one token at a time, which can result in noticeable latency. Recent advances have indicated that search and repeated sampling can enhance performance in various applications, such as theorem proving, code generation, and alignment, by utilizing greater computational resources during inference. In this study, we demonstrate that diffusion language models are capable of generating at least 32 tokens simultaneously, while exceeding the performance of AR models in text quality and on the LAMBADA natural language understanding benchmark. This outcome is achieved through a novel distillation method for discrete diffusion models, which reduces the number of inference steps by a factor of 32-64. Practically, at the 1.3B parameters scale, diffusion models, even without caching, can generate tokens at a rate that is up to 8 times faster than AR models employing KV-caching, and we anticipate further improvements with the inclusion of caching. Moreover, we demonstrate the efficacy of our approach for diffusion language models with up to 860M parameters.
Authors: Gary A. McCully, John D. Hastings, Shengjie Xu
Abstract: This study examines the impact of data snooping on neural networks used to detect vulnerabilities in lifted code, and builds on previous research that used word2vec and unidirectional and bidirectional transformer-based embeddings. The research specifically focuses on how model performance is affected when embedding models are trained with datasets, which include samples used for neural network training and validation. The results show that introducing data snooping did not significantly alter model performance, suggesting that data snooping had a minimal impact or that samples randomly dropped as part of the methodology contained hidden features critical to achieving optimal performance. In addition, the findings reinforce the conclusions of previous research, which found that models trained with GPT-2 embeddings consistently outperformed neural networks trained with other embeddings. The fact that this holds even when data snooping is introduced into the embedding model indicates GPT-2's robustness in representing complex code features, even under less-than-ideal conditions.
Authors: Muru Zhang, Mayank Mishra, Zhongzhu Zhou, William Brandon, Jue Wang, Yoon Kim, Jonathan Ragan-Kelley, Shuaiwen Leon Song, Ben Athiwaratkun, Tri Dao
Abstract: Large language model inference is both memory-intensive and time-consuming, often requiring distributed algorithms to efficiently scale. Various model parallelism strategies are used in multi-gpu training and inference to partition computation across multiple devices, reducing memory load and computation time. However, using model parallelism necessitates communication of information between GPUs, which has been a major bottleneck and limits the gains obtained by scaling up the number of devices. We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models that enables straightforward overlapping that effectively hides the latency of communication. Our insight is that in addition to systems optimization, one can also redesign the model architecture to decouple communication from computation. While Ladder Residual can allow communication-computation decoupling in conventional parallelism patterns, we focus on Tensor Parallelism in this paper, which is particularly bottlenecked by its heavy communication. For a Transformer model with 70B parameters, applying Ladder Residual to all its layers can achieve 29% end-to-end wall clock speed up at inference time with TP sharding over 8 devices. We refer the resulting Transformer model as the Ladder Transformer. We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline. We also show that it is possible to convert parts of the Llama-3.1 8B model to our Ladder Residual architecture with minimal accuracy degradation by only retraining for 3B tokens. We release our code for training and inference for easier replication of experiments.
Authors: Alexan Ayrapetyan, Sofia Kostandian, Ara Yeroyan, Mher Yerznkanyan, Nikolay Karpov, Nune Tadevosyan, Vitaly Lavrukhin, Boris Ginsburg
Abstract: This study explores methods to increase data volume for low-resource languages using techniques such as crowdsourcing, pseudo-labeling, advanced data preprocessing and various permissive data sources such as audiobooks, Common Voice, YouTube. While these methods are well-explored for highresource languages, their application for low-resource languages remains underexplored. Using Armenian and Georgian as case studies, we demonstrate how linguistic and resource-specific characteristics influence the success of these methods. This work provides practical guidance for researchers to choose cost-effective and quality-driven dataset extension strategies for low-resource languages. The key takeaway from various data extension approaches is that paid crowd-sourcing offers the best balance between cost and quality, outperforming volunteer crowd-sourcing, open-source audiobooks, and unlabeled data usage. Ablation study shows that models trained on the expanded datasets outperform existing baselines and achieve 5.73% for Gergian and 9.9% for Armenian ASR word error rate using a relatively small FastConformer architecture. We open-sourced both the Armenian and Georgian models to allow further research and practical applications.
Authors: Tianbo Yang, Mingqi Yang, Hongyi Zhao, Tianshuo Yang
Abstract: Developing the logic necessary to solve mathematical problems or write mathematical proofs is one of the more difficult objectives for large language models (LLMS). Currently, the most popular methods in literature consists of fine-tuning the model on written mathematical content such as academic publications and textbooks, so that the model can learn to emulate the style of mathematical writing. In this project, we explore the effectiveness of using retrieval augmented generation (RAG) to address gaps in the mathematical reasoning of LLMs. We develop LemmaHead, a RAG knowledge base that supplements queries to the model with relevant mathematical context, with particular focus on context from published textbooks. To measure our model's performance in mathematical reasoning, our testing paradigm focuses on the task of automated theorem proving via generating proofs to a given mathematical claim in the Lean formal language.
Authors: Seffi Cohen, Niv Goldshlager, Nurit Cohen-Inger, Bracha Shapira, Lior Rokach
Abstract: Large Language Models (LLMs) have shown remarkable capabilities across various natural language processing tasks but often struggle to excel uniformly in diverse or complex domains. We propose a novel ensemble method - Diverse Fingerprint Ensemble (DFPE), which leverages the complementary strengths of multiple LLMs to achieve more robust performance. Our approach involves: (1) clustering models based on response "fingerprints" patterns, (2) applying a quantile-based filtering mechanism to remove underperforming models at a per-subject level, and (3) assigning adaptive weights to remaining models based on their subject-wise validation accuracy. In experiments on the Massive Multitask Language Understanding (MMLU) benchmark, DFPE outperforms the best single model by 3% overall accuracy and 5% in discipline-level accuracy. This method increases the robustness and generalization of LLMs and underscores how model selection, diversity preservation, and performance-driven weighting can effectively address challenging, multi-faceted language understanding tasks.
Authors: Jeonghoon Kim, Byeongchan Lee, Cheonbok Park, Yeontaek Oh, Beomjun Kim, Taehwan Yoo, Seongjin Shin, Dongyoon Han, Jinwoo Shin, Kang Min Yoo
Abstract: Designing Transformer architectures with the optimal layer normalization (LN) strategy that ensures large-scale training stability and expedite convergence has remained elusive, even in this era of large language models (LLMs). To this end, we present a comprehensive analytical foundation for understanding how different LN strategies influence training dynamics in large-scale Transformer training. Until recently, Pre-LN and Post-LN have long dominated standard practices despite their limitations in large-scale training. However, several open-source large-scale models have recently begun silently adopting a third strategy without much explanation. This strategy places layer normalization (LN) peripherally around sublayers, a design we term Peri-LN. While Peri-LN has demonstrated promising empirical performance, its precise mechanisms and benefits remain almost unexplored. Our in-depth analysis shows that Peri-LN strikes an ideal balance in variance growth -- unlike Pre-LN and Post-LN, which are prone to vanishing gradients and ``massive activations.'' To validate our theoretical insight, we conduct large-scale experiments on Transformers up to 3.2B parameters, showing that Peri-LN consistently achieves more balanced variance growth, steadier gradient flow, and convergence stability. Our results suggest that Peri-LN warrants broader consideration for large-scale Transformer architectures, providing renewed insights into the optimal placement and application of LN.