Authors: Trilok Padhi, Ramneet Kaur, Adam D. Cobb, Manoj Acharya, Anirban Roy, Colin Samplawski, Brian Matejek, Alexander M. Berenbeim, Nathaniel D. Bastian, Susmit Jha
Abstract: We introduce a novel approach for calibrating uncertainty quantification (UQ) tailored for multi-modal large language models (LLMs). Existing state-of-the-art UQ methods rely on consistency among multiple responses generated by the LLM on an input query under diverse settings. However, these approaches often report higher confidence in scenarios where the LLM is consistently incorrect. This leads to a poorly calibrated confidence with respect to accuracy. To address this, we leverage cross-modal consistency in addition to self-consistency to improve the calibration of the multi-modal models. Specifically, we ground the textual responses to the visual inputs. The confidence from the grounding model is used to calibrate the overall confidence. Given that using a grounding model adds its own uncertainty in the pipeline, we apply temperature scaling - a widely accepted parametric calibration technique - to calibrate the grounding model's confidence in the accuracy of generated responses. We evaluate the proposed approach across multiple multi-modal tasks, such as medical question answering (Slake) and visual question answering (VQAv2), considering multi-modal models such as LLaVA-Med and LLaVA. The experiments demonstrate that the proposed framework achieves significantly improved calibration on both tasks.
Authors: Gianluca Manzo, Julia Ive
Abstract: Automating chest radiograph interpretation using Deep Learning (DL) models has the potential to significantly improve clinical workflows, decision-making, and large-scale health screening. However, in medical settings, merely optimising predictive performance is insufficient, as the quantification of uncertainty is equally crucial. This paper investigates the relationship between predictive uncertainty, derived from Bayesian Deep Learning approximations, and human/linguistic uncertainty, as estimated from free-text radiology reports labelled by rule-based labellers. Utilising BERT as the model of choice, this study evaluates different binarisation methods for uncertainty labels and explores the efficacy of Monte Carlo Dropout and Deep Ensembles in estimating predictive uncertainty. The results demonstrate good model performance, but also a modest correlation between predictive and linguistic uncertainty, highlighting the challenges in aligning machine uncertainty with human interpretation nuances. Our findings suggest that while Bayesian approximations provide valuable uncertainty estimates, further refinement is necessary to fully capture and utilise the subtleties of human uncertainty in clinical applications.
Authors: Lucia Zheng, Neel Guha, Javokhir Arifov, Sarah Zhang, Michal Skreta, Christopher D. Manning, Peter Henderson, Daniel E. Ho
Abstract: As the legal community increasingly examines the use of large language models (LLMs) for various legal applications, legal AI developers have turned to retrieval-augmented LLMs ("RAG" systems) to improve system performance and robustness. An obstacle to the development of specialized RAG systems is the lack of realistic legal RAG benchmarks which capture the complexity of both legal retrieval and downstream legal question-answering. To address this, we introduce two novel legal RAG benchmarks: Bar Exam QA and Housing Statute QA. Our tasks correspond to real-world legal research tasks, and were produced through annotation processes which resemble legal research. We describe the construction of these benchmarks and the performance of existing retriever pipelines. Our results suggest that legal RAG remains a challenging application, thus motivating future research.
Authors: Jiale Liu, Yifan Zeng, Shaokun Zhang, Chi Zhang, Malte H{\o}jmark-Bertelsen, Marie Normann Gadeberg, Huazheng Wang, Qingyun Wu
Abstract: LLM-based optimization has shown remarkable potential in enhancing agentic systems. However, the conventional approach of prompting LLM optimizer with the whole training trajectories on training dataset in a single pass becomes untenable as datasets grow, leading to context window overflow and degraded pattern recognition. To address these challenges, we propose Fine-Grained Optimization (FGO), a scalable framework that divides large optimization tasks into manageable subsets, performs targeted optimizations, and systematically combines optimized components through progressive merging. Evaluation across ALFWorld, LogisticsQA, and GAIA benchmarks demonstrate that FGO outperforms existing approaches by 1.6-8.6% while reducing average prompt token consumption by 56.3%. Our framework provides a practical solution for scaling up LLM-based optimization of increasingly sophisticated agent systems. Further analysis demonstrates that FGO achieves the most consistent performance gain in all training dataset sizes, showcasing its scalability and efficiency.
Authors: Qianchu Liu, Sheng Zhang, Guanghui Qin, Timothy Ossowski, Yu Gu, Ying Jin, Sid Kiblawi, Sam Preston, Mu Wei, Paul Vozila, Tristan Naumann, Hoifung Poon
Abstract: Recent proprietary models (e.g., o3) have begun to demonstrate strong multimodal reasoning capabilities. Yet, most existing open-source research concentrates on training text-only reasoning models, with evaluations limited to mainly mathematical and general-domain tasks. Therefore, it remains unclear how to effectively extend reasoning capabilities beyond text input and general domains. This paper explores a fundamental research question: Is reasoning generalizable across modalities and domains? Our findings support an affirmative answer: General-domain text-based post-training can enable such strong generalizable reasoning. Leveraging this finding, we introduce X-Reasoner, a vision-language model post-trained solely on general-domain text for generalizable reasoning, using a two-stage approach: an initial supervised fine-tuning phase with distilled long chain-of-thoughts, followed by reinforcement learning with verifiable rewards. Experiments show that X-Reasoner successfully transfers reasoning capabilities to both multimodal and out-of-domain settings, outperforming existing state-of-the-art models trained with in-domain and multimodal data across various general and medical benchmarks (Figure 1). Additionally, we find that X-Reasoner's performance in specialized domains can be further enhanced through continued training on domain-specific text-only data. Building upon this, we introduce X-Reasoner-Med, a medical-specialized variant that achieves new state of the art on numerous text-only and multimodal medical benchmarks.
Authors: Darren Yow-Bang Wang, Zhengyuan Shen, Soumya Smruti Mishra, Zhichao Xu, Yifei Teng, Haibo Ding
Abstract: Structured outputs are essential for large language models (LLMs) in critical applications like agents and information extraction. Despite their capabilities, LLMs often generate outputs that deviate from predefined schemas, significantly hampering reliable application development. We present SLOT (Structured LLM Output Transformer), a model-agnostic approach that transforms unstructured LLM outputs into precise structured formats. While existing solutions predominantly rely on constrained decoding techniques or are tightly coupled with specific models, SLOT employs a fine-tuned lightweight language model as a post-processing layer, achieving flexibility across various LLMs and schema specifications. We introduce a systematic pipeline for data curation and synthesis alongside a formal evaluation methodology that quantifies both schema accuracy and content fidelity. Our results demonstrate that fine-tuned Mistral-7B model with constrained decoding achieves near perfect schema accuracy (99.5%) and content similarity (94.0%), outperforming Claude-3.5-Sonnet by substantial margins (+25 and +20 percentage points, respectively). Notably, even compact models like Llama-3.2-1B can match or exceed the structured output capabilities of much larger proprietary models when equipped with SLOT, enabling reliable structured generation in resource-constrained environments.
Authors: Xu Huang, Yuefeng Huang, Weiwen Liu, Xingshan Zeng, Yasheng Wang, Ruiming Tang, Hong Xie, Defu Lian
Abstract: Tool invocation is a crucial mechanism for extending the capabilities of Large Language Models (LLMs) and has recently garnered significant attention. It enables LLMs to solve complex problems through tool calls while accessing up-to-date world knowledge. However, existing work primarily focuses on the fundamental ability of LLMs to invoke tools for problem-solving, without considering personalized constraints in tool invocation. In this work, we introduce the concept of Personalized Tool Invocation and define two key tasks: Tool Preference and Profile-dependent Query. Tool Preference addresses user preferences when selecting among functionally similar tools, while Profile-dependent Query considers cases where a user query lacks certain tool parameters, requiring the model to infer them from the user profile. To tackle these challenges, we propose PTool, a data synthesis framework designed for personalized tool invocation. Additionally, we construct \textbf{PTBench}, the first benchmark for evaluating personalized tool invocation. We then fine-tune various open-source models, demonstrating the effectiveness of our framework and providing valuable insights. Our benchmark is public at https://github.com/hyfshadow/PTBench.
Authors: Mengxian Lyu, Xiaohan Li, Ziyi Chen, Jinqian Pan, Cheng Peng, Sankalp Talankar, Yonghui Wu
Abstract: Natural language generation (NLG) is the key technology to achieve generative artificial intelligence (AI). With the breakthroughs in large language models (LLMs), NLG has been widely used in various medical applications, demonstrating the potential to enhance clinical workflows, support clinical decision-making, and improve clinical documentation. Heterogeneous and diverse medical data modalities, such as medical text, images, and knowledge bases, are utilized in NLG. Researchers have proposed many generative models and applied them in a number of healthcare applications. There is a need for a comprehensive review of NLG methods and applications in the medical domain. In this study, we systematically reviewed 113 scientific publications from a total of 3,988 NLG-related articles identified using a literature search, focusing on data modality, model architecture, clinical applications, and evaluation methods. Following PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) guidelines, we categorize key methods, identify clinical applications, and assess their capabilities, limitations, and emerging challenges. This timely review covers the key NLG technologies and medical applications and provides valuable insights for future studies to leverage NLG to transform medical discovery and healthcare.
Authors: Mingruo Yuan, Ben Kao, Tien-Hsuan Wu, Michael M. K. Cheung, Henry W. H. Chan, Anne S. Y. Cheung, Felix W. H. Chan, Yongxi Chen
Abstract: Access to legal information is fundamental to access to justice. Yet accessibility refers not only to making legal documents available to the public, but also rendering legal information comprehensible to them. A vexing problem in bringing legal information to the public is how to turn formal legal documents such as legislation and judgments, which are often highly technical, to easily navigable and comprehensible knowledge to those without legal education. In this study, we formulate a three-step approach for bringing legal knowledge to laypersons, tackling the issues of navigability and comprehensibility. First, we translate selected sections of the law into snippets (called CLIC-pages), each being a small piece of article that focuses on explaining certain technical legal concept in layperson's terms. Second, we construct a Legal Question Bank (LQB), which is a collection of legal questions whose answers can be found in the CLIC-pages. Third, we design an interactive CLIC Recommender (CRec). Given a user's verbal description of a legal situation that requires a legal solution, CRec interprets the user's input and shortlists questions from the question bank that are most likely relevant to the given legal situation and recommends their corresponding CLIC pages where relevant legal knowledge can be found. In this paper we focus on the technical aspects of creating an LQB. We show how large-scale pre-trained language models, such as GPT-3, can be used to generate legal questions. We compare machine-generated questions (MGQs) against human-composed questions (HCQs) and find that MGQs are more scalable, cost-effective, and more diversified, while HCQs are more precise. We also show a prototype of CRec and illustrate through an example how our 3-step approach effectively brings relevant legal knowledge to the public.
Authors: Vihaan Miriyala, Smrithi Bukkapatnam, Lavanya Prahallad
Abstract: We explore the use of Chain-of-Thought (CoT) prompting with large language models (LLMs) to improve the accuracy of granular sentiment categorization in app store reviews. Traditional numeric and polarity-based ratings often fail to capture the nuanced sentiment embedded in user feedback. We evaluated the effectiveness of CoT prompting versus simple prompting on 2000 Amazon app reviews by comparing each method's predictions to human judgements. CoT prompting improved classification accuracy from 84% to 93% highlighting the benefit of explicit reasoning in enhancing sentiment analysis performance.
Authors: Variath Madhupal Gautham Nair, Vishal Varma Dantuluri
Abstract: Existing large language models (LLMs) are advancing rapidly and produce outstanding results in image generation tasks, yet their content safety checks remain vulnerable to prompt-based jailbreaks. Through preliminary testing on platforms such as ChatGPT, MetaAI, and Grok, we observed that even short, natural prompts could lead to the generation of compromising images ranging from realistic depictions of forged documents to manipulated images of public figures. We introduce Unmasking the Canvas (UTC Benchmark; UTCB), a dynamic and scalable benchmark dataset to evaluate LLM vulnerability in image generation. Our methodology combines structured prompt engineering, multilingual obfuscation (e.g., Zulu, Gaelic, Base64), and evaluation using Groq-hosted LLaMA-3. The pipeline supports both zero-shot and fallback prompting strategies, risk scoring, and automated tagging. All generations are stored with rich metadata and curated into Bronze (non-verified), Silver (LLM-aided verification), and Gold (manually verified) tiers. UTCB is designed to evolve over time with new data sources, prompt templates, and model behaviors. Warning: This paper includes visual examples of adversarial inputs designed to test model safety. All outputs have been redacted to ensure responsible disclosure.
Authors: Manas Satish Bedmutha, Feng Chen, Andrea Hartzler, Trevor Cohen, Nadir Weibel
Abstract: Effective communication between providers and their patients influences health and care outcomes. The effectiveness of such conversations has been linked not only to the exchange of clinical information, but also to a range of interpersonal behaviors; commonly referred to as social signals, which are often conveyed through non-verbal cues and shape the quality of the patient-provider relationship. Recent advances in large language models (LLMs) have demonstrated an increasing ability to infer emotional and social behaviors even when analyzing only textual information. As automation increases also in clinical settings, such as for transcription of patient-provider conversations, there is growing potential for LLMs to automatically analyze and extract social behaviors from these interactions. To explore the foundational capabilities of LLMs in tracking social signals in clinical dialogue, we designed task-specific prompts and evaluated model performance across multiple architectures and prompting styles using a highly imbalanced, annotated dataset spanning 20 distinct social signals such as provider dominance, patient warmth, etc. We present the first system capable of tracking all these 20 coded signals, and uncover patterns in LLM behavior. Further analysis of model configurations and clinical context provides insights for enhancing LLM performance on social signal processing tasks in healthcare settings.
Authors: Maria Marina, Nikolay Ivanov, Sergey Pletenev, Mikhail Salnikov, Daria Galimzianova, Nikita Krayko, Vasily Konovalov, Alexander Panchenko, Viktor Moskvoretskii
Abstract: Large Language Models~(LLMs) are prone to hallucinations, and Retrieval-Augmented Generation (RAG) helps mitigate this, but at a high computational cost while risking misinformation. Adaptive retrieval aims to retrieve only when necessary, but existing approaches rely on LLM-based uncertainty estimation, which remain inefficient and impractical. In this study, we introduce lightweight LLM-independent adaptive retrieval methods based on external information. We investigated 27 features, organized into 7 groups, and their hybrid combinations. We evaluated these methods on 6 QA datasets, assessing the QA performance and efficiency. The results show that our approach matches the performance of complex LLM-based methods while achieving significant efficiency gains, demonstrating the potential of external information for adaptive retrieval.
Authors: Sofia Jamil, Aryan Dabad, Bollampalli Areen Reddy, Sriparna Saha, Rajiv Misra, Adil A. Shakur
Abstract: In the realm of cancer treatment, summarizing adverse drug events (ADEs) reported by patients using prescribed drugs is crucial for enhancing pharmacovigilance practices and improving drug-related decision-making. While the volume and complexity of pharmacovigilance data have increased, existing research in this field has predominantly focused on general diseases rather than specifically addressing cancer. This work introduces the task of grouped summarization of adverse drug events reported by multiple patients using the same drug for cancer treatment. To address the challenge of limited resources in cancer pharmacovigilance, we present the MultiLabeled Cancer Adverse Drug Reaction and Summarization (MCADRS) dataset. This dataset includes pharmacovigilance posts detailing patient concerns regarding drug efficacy and adverse effects, along with extracted labels for drug names, adverse drug events, severity, and adversity of reactions, as well as summaries of ADEs for each drug. Additionally, we propose the Grouping and Abstractive Summarization of Cancer Adverse Drug events (GASCADE) framework, a novel pipeline that combines the information extraction capabilities of Large Language Models (LLMs) with the summarization power of the encoder-decoder T5 model. Our work is the first to apply alignment techniques, including advanced algorithms like Direct Preference Optimization, to encoder-decoder models using synthetic datasets for summarization tasks. Through extensive experiments, we demonstrate the superior performance of GASCADE across various metrics, validated through both automated assessments and human evaluations. This multitasking approach enhances drug-related decision-making and fosters a deeper understanding of patient concerns, paving the way for advancements in personalized and responsive cancer care. The code and dataset used in this work are publicly available.
Authors: Dario Garcia-Gasulla, Jordi Bayarri-Planas, Ashwin Kumar Gururajan, Enrique Lopez-Cuena, Adrian Tormos, Daniel Hinjos, Pablo Bernabeu-Perez, Anna Arias-Duart, Pablo Agustin Martin-Torres, Marta Gonzalez-Mallo, Sergio Alvarez-Napagao, Eduard Ayguad\'e-Parra, Ulises Cort\'es
Abstract: Purpose: With advancements in Large Language Models (LLMs) for healthcare, the need arises for competitive open-source models to protect the public interest. This work contributes to the field of open medical LLMs by optimizing key stages of data preprocessing and training, while showing how to improve model safety (through DPO) and efficacy (through RAG). The evaluation methodology used, which includes four different types of tests, defines a new standard for the field. The resultant models, shown to be competitive with the best private alternatives, are released with a permisive license. Methods: Building on top of strong base models like Llama 3.1 and Qwen 2.5, Aloe Beta uses a custom dataset to enhance public data with synthetic Chain of Thought examples. The models undergo alignment with Direct Preference Optimization, emphasizing ethical and policy-aligned performance in the presence of jailbreaking attacks. Evaluation includes close-ended, open-ended, safety and human assessments, to maximize the reliability of results. Results: Recommendations are made across the entire pipeline, backed by the solid performance of the Aloe Family. These models deliver competitive performance across healthcare benchmarks and medical fields, and are often preferred by healthcare professionals. On bias and toxicity, the Aloe Beta models significantly improve safety, showing resilience to unseen jailbreaking attacks. For a responsible release, a detailed risk assessment specific to healthcare is attached to the Aloe Family models. Conclusion: The Aloe Beta models, and the recipe that leads to them, are a significant contribution to the open-source medical LLM field, offering top-of-the-line performance while maintaining high ethical requirements. This work sets a new standard for developing and reporting aligned LLMs in healthcare.
Authors: David Exler, Mark Schutera, Markus Reischl, Luca Rettenberger
Abstract: With the increasing prevalence of artificial intelligence, careful evaluation of inherent biases needs to be conducted to form the basis for alleviating the effects these predispositions can have on users. Large language models (LLMs) are predominantly used by many as a primary source of information for various topics. LLMs frequently make factual errors, fabricate data (hallucinations), or present biases, exposing users to misinformation and influencing opinions. Educating users on their risks is key to responsible use, as bias, unlike hallucinations, cannot be caught through data verification. We quantify the political bias of popular LLMs in the context of the recent vote of the German Bundestag using the score produced by the Wahl-O-Mat. This metric measures the alignment between an individual's political views and the positions of German political parties. We compare the models' alignment scores to identify factors influencing their political preferences. Doing so, we discover a bias toward left-leaning parties, most dominant in larger LLMs. Also, we find that the language we use to communicate with the models affects their political views. Additionally, we analyze the influence of a model's origin and release date and compare the results to the outcome of the recent vote of the Bundestag. Our results imply that LLMs are prone to exhibiting political bias. Large corporations with the necessary means to develop LLMs, thus, knowingly or unknowingly, have a responsibility to contain these biases, as they can influence each voter's decision-making process and inform public opinion in general and at scale.
Authors: Aidar Valeev (Research Center of the Artificial Intelligence Institute, Innopolis University, Russia), Roman Garaev (Research Center of the Artificial Intelligence Institute, Innopolis University, Russia), Vadim Lomshakov (St. Petersburg Department of the Steklov Institute of Mathematics, Russia), Irina Piontkovskaya (Huawei Noah's Ark Lab), Vladimir Ivanov (Research Center of the Artificial Intelligence Institute, Innopolis University, Russia), Israel Adewuyi (Research Center of the Artificial Intelligence Institute, Innopolis University, Russia)
Abstract: Large Language Models demonstrate the ability to solve various programming tasks, including code generation. Typically, the performance of LLMs is measured on benchmarks with small or medium-sized context windows of thousands of lines of code. At the same time, in real-world software projects, repositories can span up to millions of LoC. This paper closes this gap by contributing to the long context code generation benchmark (YABLoCo). The benchmark featured a test set of 215 functions selected from four large repositories with thousands of functions. The dataset contained metadata of functions, contexts of the functions with different levels of dependencies, docstrings, functions bodies, and call graphs for each repository. This paper presents three key aspects of the contribution. First, the benchmark aims at function body generation in large repositories in C and C++, two languages not covered by previous benchmarks. Second, the benchmark contains large repositories from 200K to 2,000K LoC. Third, we contribute a scalable evaluation pipeline for efficient computing of the target metrics and a tool for visual analysis of generated code. Overall, these three aspects allow for evaluating code generation in large repositories in C and C++.
Authors: Xiaoyu Xu, Minxin Du, Qingqing Ye, Haibo Hu
Abstract: Large language models (LLMs) trained over extensive corpora risk memorizing sensitive, copyrighted, or toxic content. To address this, we propose OBLIVIATE, a robust unlearning framework that removes targeted data while preserving model utility. The framework follows a structured process: extracting target tokens, building retain sets, and fine-tuning with a tailored loss function comprising three components -- masking, distillation, and world fact. Using low-rank adapters (LoRA), it ensures efficiency without compromising unlearning quality. We conduct experiments on multiple datasets, including the Harry Potter series, WMDP, and TOFU, using a comprehensive suite of metrics: forget quality (new document-level memorization score), model utility, and fluency. Results demonstrate its effectiveness in resisting membership inference attacks, minimizing the impact on retained data, and maintaining robustness across diverse scenarios.
Authors: Ilya Koziev
Abstract: The quality of natural language texts in fine-tuning datasets plays a critical role in the performance of generative models, particularly in computational creativity tasks such as poem or song lyric generation. Fluency defects in generated poems significantly reduce their value. However, training texts are often sourced from internet-based platforms without stringent quality control, posing a challenge for data engineers to manage defect levels effectively. To address this issue, we propose the use of automated linguistic anomaly detection to identify and filter out low-quality texts from training datasets for creative models. In this paper, we present a comprehensive comparison of unsupervised and supervised text anomaly detection approaches, utilizing both synthetic and human-labeled datasets. We also introduce the RUPOR dataset, a collection of Russian-language human-labeled poems designed for cross-sentence grammatical error detection, and provide the full evaluation code. Our work aims to empower the community with tools and insights to improve the quality of training datasets for generative models in creative domains.
Authors: Yehui Tang, Yichun Yin, Yaoyuan Wang, Hang Zhou, Yu Pan, Wei Guo, Ziyang Zhang, Miao Rang, Fangcheng Liu, Naifu Zhang, Binghan Li, Yonghan Dong, Xiaojun Meng, Yasheng Wang, Dong Li, Yin Li, Dandan Tu, Can Chen, Youliang Yan, Fisher Yu, Ruiming Tang, Yunhe Wang, Botian Huang, Bo Wang, Boxiao Liu, Changzheng Zhang, Da Kuang, Fei Liu, Gang Huang, Jiansheng Wei, Jiarui Qin, Jie Ran, Jinpeng Li, Jun Zhao, Liang Dai, Lin Li, Liqun Deng, Peifeng Qin, Pengyuan Zeng, Qiang Gu, Shaohua Tang, Shengjun Cheng, Tao Gao, Tao Yu, Tianshu Li, Tianyu Bi, Wei He, Weikai Mao, Wenyong Huang, Wulong Liu, Xiabing Li, Xianzhi Yu, Xueyu Wu, Xu He, Yangkai Du, Yan Xu, Ye Tian, Yimeng Wu, Yongbing Huang, Yong Tian, Yong Zhu, Yue Li, Yufei Wang, Yuhang Gai, Yujun Li, Yu Luo, Yunsheng Ni, Yusen Sun, Zelin Chen, Zhe Liu, Zhicheng Liu, Zhipeng Tu, Zilin Ding, Zongyuan Zhan
Abstract: Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.
Authors: Josh McGiff, Nikola S. Nikolov
Abstract: Generative language modelling has surged in popularity with the emergence of services such as ChatGPT and Google Gemini. While these models have demonstrated transformative potential in productivity and communication, they overwhelmingly cater to high-resource languages like English. This has amplified concerns over linguistic inequality in natural language processing (NLP). This paper presents the first systematic review focused specifically on strategies to address data scarcity in generative language modelling for low-resource languages (LRL). Drawing from 54 studies, we identify, categorise and evaluate technical approaches, including monolingual data augmentation, back-translation, multilingual training, and prompt engineering, across generative tasks. We also analyse trends in architecture choices, language family representation, and evaluation methods. Our findings highlight a strong reliance on transformer-based models, a concentration on a small subset of LRLs, and a lack of consistent evaluation across studies. We conclude with recommendations for extending these methods to a wider range of LRLs and outline open challenges in building equitable generative language systems. Ultimately, this review aims to support researchers and developers in building inclusive AI tools for underrepresented languages, a necessary step toward empowering LRL speakers and the preservation of linguistic diversity in a world increasingly shaped by large-scale language technologies.
Authors: Hao Sun, Zile Qiao, Jiayan Guo, Xuanbo Fan, Yingyan Hou, Yong Jiang, Pengjun Xie, Fei Huang, Yan Zhang
Abstract: Effective information searching is essential for enhancing the reasoning and generation capabilities of large language models (LLMs). Recent research has explored using reinforcement learning (RL) to improve LLMs' search capabilities by interacting with live search engines in real-world environments. While these approaches show promising results, they face two major challenges: (1) Uncontrolled Document Quality: The quality of documents returned by search engines is often unpredictable, introducing noise and instability into the training process. (2) Prohibitively High API Costs: RL training requires frequent rollouts, potentially involving hundreds of thousands of search requests, which incur substantial API expenses and severely constrain scalability. To address these challenges, we introduce ZeroSearch, a reinforcement learning framework that incentivizes the search capabilities of LLMs without interacting with real search engines. Our approach begins with lightweight supervised fine-tuning to transform the LLM into a retrieval module capable of generating both relevant and noisy documents in response to a query. During RL training, we employ a curriculum-based rollout strategy that incrementally degrades the quality of generated documents, progressively eliciting the model's reasoning ability by exposing it to increasingly challenging retrieval scenarios. Extensive experiments demonstrate that ZeroSearch effectively incentivizes the search capabilities of LLMs using a 3B LLM as the retrieval module. Remarkably, a 7B retrieval module achieves comparable performance to the real search engine, while a 14B retrieval module even surpasses it. Furthermore, it generalizes well across both base and instruction-tuned models of various parameter sizes and is compatible with a wide range of RL algorithms.
Authors: Md Fahim Anjum
Abstract: Large Language Models (LLM) with reasoning capabilities offer a promising path for improving candidate evaluation in planning frameworks, but their relative performance against traditional non-reasoning models remains largely underexplored. In this study, we benchmark a distilled 1.5B parameter reasoning model (DeepSeek-R1) against several state-of-the-art non-reasoning LLMs within a generator-discriminator LLM planning framework for the text-to-SQL task. For this, we introduce a novel method for extracting soft scores from the chain-of-thought (CoT) outputs from reasoning that enables fine-grained ranking of candidates. Our central hypothesis is that reasoning models are more effective discriminators than non-reasoning LLMs. Our results show that distilled DeepSeek-R1-1.5B achieves up to $87\%$ higher F1 and $3.7\%$ better discrimination accuracy than CodeLlama-7B, as well as $3.7\%$ higher execution accuracy than CodeLlama-13B, despite having significantly fewer parameters. Furthermore, we find that there is a limit to the logical capabilities of reasoning models, and only providing more context or allowing more compute budget for reasoning is not enough to improve their discrimination performance. Finally, we demonstrate that, unlike non-reasoning LLMs, reasoning models find generation more challenging than discrimination and may underperform as generators compared to smaller non-reasoning LLMs. Our work highlights the potential of reasoning models as discriminators in agentic frameworks, far outweighing their capabilities as generators, offering insights into their optimal role within LLM planning infrastructures.
Authors: Hyun Lee, Chris Yi, Maminur Islam, B. D. S. Aritra
Abstract: Large Language Models (LLMs) have demonstrated strong capabilities in various natural language processing tasks; however, their application to graph-related problems remains limited, primarily due to scalability constraints and the absence of dedicated mechanisms for processing graph structures. Existing approaches predominantly integrate LLMs with Graph Neural Networks (GNNs), using GNNs as feature encoders or auxiliary components. However, directly encoding graph structures within LLMs has been underexplored, particularly in the context of large-scale graphs where token limitations hinder effective representation. To address these challenges, we propose SDM-InstructGLM, a novel instruction-tuned Graph Language Model (InstructGLM) framework that enhances scalability and efficiency without relying on GNNs. Our method introduces a similarity-degree-based biased random walk mechanism, which selectively samples and encodes graph information based on node-feature similarity and degree centrality, ensuring an adaptive and structured representation within the LLM. This approach significantly improves token efficiency, mitigates information loss due to random sampling, and enhances performance on graph-based tasks such as node classification and link prediction. Furthermore, our results demonstrate the feasibility of LLM-only graph processing, enabling scalable and interpretable Graph Language Models (GLMs) optimized through instruction-based fine-tuning. This work paves the way for GNN-free approaches to graph learning, leveraging LLMs as standalone graph reasoning models. Our source code is available on GitHub.
Authors: Euntae Choi, Sumin Song, Woosang Lim, Sungjoo Yoo
Abstract: Large Language Models (LLMs) face deployment challenges due to high computational costs, and while Post-Training Quantization (PTQ) offers a solution, existing rotation-based methods struggle at very low bit-widths like 2-bit. We introduce a novel, training-free approach to construct an improved rotation matrix, addressing the limitations of current methods. The key contributions include leveraging the Walsh-Hadamard transform with sequency ordering, which clusters similar frequency components to reduce quantization error compared to standard Hadamard matrices, significantly improving performance. Furthermore, we propose a Grouped Sequency-arranged Rotation (GSR) using block-diagonal matrices with smaller Walsh blocks, effectively isolating outlier impacts and achieving performance comparable to optimization-based methods without requiring any training. Our method demonstrates robust performance on reasoning tasks and Perplexity (PPL) score on WikiText-2. Our method also enhances results even when applied over existing learned rotation techniques.
Authors: Ganghua Wang, Zhaorun Chen, Bo Li, Haifeng Xu
Abstract: As foundation models continue to scale, the size of trained models grows exponentially, presenting significant challenges for their evaluation. Current evaluation practices involve curating increasingly large datasets to assess the performance of large language models (LLMs). However, there is a lack of systematic analysis and guidance on determining the sufficiency of test data or selecting informative samples for evaluation. This paper introduces a certifiable and cost-efficient evaluation framework for LLMs. Our framework adapts to different evaluation objectives and outputs confidence intervals that contain true values with high probability. We use ``test sample complexity'' to quantify the number of test points needed for a certifiable evaluation and derive tight bounds on test sample complexity. Based on the developed theory, we develop a partition-based algorithm, named Cer-Eval, that adaptively selects test points to minimize the cost of LLM evaluation. Real-world experiments demonstrate that Cer-Eval can save 20% to 40% test points across various benchmarks, while maintaining an estimation error level comparable to the current evaluation process and providing a 95% confidence guarantee.
Authors: Yogesh Gajula
Abstract: E-commerce platforms generate vast volumes of user feedback, such as star ratings, written reviews, and comments. However, most recommendation engines rely primarily on numerical scores, often overlooking the nuanced opinions embedded in free text. This paper comprehensively reviews sentiment-aware recommendation systems from a natural language processing perspective, covering advancements from 2023 to early 2025. It highlights the benefits of integrating sentiment analysis into e-commerce recommenders to enhance prediction accuracy and explainability through detailed opinion extraction. Our survey categorizes recent work into four main approaches: deep learning classifiers that combine sentiment embeddings with user item interactions, transformer based methods for nuanced feature extraction, graph neural networks that propagate sentiment signals, and conversational recommenders that adapt in real time to user feedback. We summarize model architectures and demonstrate how sentiment flows through recommendation pipelines, impacting dialogue-based suggestions. Key challenges include handling noisy or sarcastic text, dynamic user preferences, and bias mitigation. Finally, we outline research gaps and provide a roadmap for developing smarter, fairer, and more user-centric recommendation tools.
Authors: Gerrit Gro{\ss}mann, Larisa Ivanova, Sai Leela Poduru, Mohaddeseh Tabrizian, Islam Mesabah, David A. Selby, Sebastian J. Vollmer
Abstract: According to Yuval Noah Harari, large-scale human cooperation is driven by shared narratives that encode common beliefs and values. This study explores whether such narratives can similarly nudge LLM agents toward collaboration. We use a finitely repeated public goods game in which LLM agents choose either cooperative or egoistic spending strategies. We prime agents with stories highlighting teamwork to different degrees and test how this influences negotiation outcomes. Our experiments explore four questions:(1) How do narratives influence negotiation behavior? (2) What differs when agents share the same story versus different ones? (3) What happens when the agent numbers grow? (4) Are agents resilient against self-serving negotiators? We find that story-based priming significantly affects negotiation strategies and success rates. Common stories improve collaboration, benefiting each agent. By contrast, priming agents with different stories reverses this effect, and those agents primed toward self-interest prevail. We hypothesize that these results carry implications for multi-agent system design and AI alignment.
Authors: Prudhviraj Naidu, Zixian Wang, Leon Bergen, Ramamohan Paturi
Abstract: We train Transformer-based language models on ten foundational algorithmic tasks and observe pronounced phase transitions in their loss curves that deviate from established power-law scaling trends. Over large ranges of compute, the validation loss barely improves, then abruptly decreases. Probing the models' internal representations reveals the learning of quiet features during the stagnant phase, followed by sudden acquisition of loud features that coincide with the sharp drop in loss. Our ablation experiments show that disrupting a single learned feature can dramatically degrade performance, providing evidence of their causal role in task performance. These findings challenge the prevailing assumption that next-token predictive loss reliably tracks incremental progress; instead, key internal features may be developing below the surface until they coalesce, triggering a rapid performance gain.
Authors: Tuochao Chen, Nicholas Batchelder, Alisa Liu, Noah Smith, Shyamnath Gollakota
Abstract: We introduce LlamaPIE, the first real-time proactive assistant designed to enhance human conversations through discreet, concise guidance delivered via hearable devices. Unlike traditional language models that require explicit user invocation, this assistant operates in the background, anticipating user needs without interrupting conversations. We address several challenges, including determining when to respond, crafting concise responses that enhance conversations, leveraging knowledge of the user for context-aware assistance, and real-time, on-device processing. To achieve this, we construct a semi-synthetic dialogue dataset and propose a two-model pipeline: a small model that decides when to respond and a larger model that generates the response. We evaluate our approach on real-world datasets, demonstrating its effectiveness in providing helpful, unobtrusive assistance. User studies with our assistant, implemented on Apple Silicon M2 hardware, show a strong preference for the proactive assistant over both a baseline with no assistance and a reactive model, highlighting the potential of LlamaPie to enhance live conversations.
Authors: Nouar Aldahoul, Hazem Ibrahim, Matteo Varvello, Aaron Kaufman, Talal Rahwan, Yasir Zaki
Abstract: Large Language Models (LLMs) are a transformational technology, fundamentally changing how people obtain information and interact with the world. As people become increasingly reliant on them for an enormous variety of tasks, a body of academic research has developed to examine these models for inherent biases, especially political biases, often finding them small. We challenge this prevailing wisdom. First, by comparing 31 LLMs to legislators, judges, and a nationally representative sample of U.S. voters, we show that LLMs' apparently small overall partisan preference is the net result of offsetting extreme views on specific topics, much like moderate voters. Second, in a randomized experiment, we show that LLMs can promulgate their preferences into political persuasiveness even in information-seeking contexts: voters randomized to discuss political issues with an LLM chatbot are as much as 5 percentage points more likely to express the same preferences as that chatbot. Contrary to expectations, these persuasive effects are not moderated by familiarity with LLMs, news consumption, or interest in politics. LLMs, especially those controlled by private companies or governments, may become a powerful and targeted vector for political influence.
Authors: Trinh T. L. Vuong, Jin Tae Kwak
Abstract: We present VideoPath-LLaVA, the first large multimodal model (LMM) in computational pathology that integrates three distinct image scenarios, single patch images, automatically keyframe-extracted clips, and manually segmented video pathology images, to mimic the natural diagnostic process of pathologists. By generating detailed histological descriptions and culminating in a definitive sign-out diagnosis, VideoPath-LLaVA bridges visual narratives with diagnostic reasoning. Central to our approach is the VideoPath-Instruct dataset, comprising 4278 video and diagnosis-specific chain-of-thought instructional pairs sourced from educational histopathology videos on YouTube. Although high-quality data is critical for enhancing diagnostic reasoning, its creation is time-intensive and limited in volume. To overcome this challenge, we transfer knowledge from existing single-image instruction datasets to train on weakly annotated, keyframe-extracted clips, followed by fine-tuning on manually segmented videos. VideoPath-LLaVA establishes a new benchmark in pathology video analysis and offers a promising foundation for future AI systems that support clinical decision-making through integrated visual and diagnostic reasoning. Our code, data, and model are publicly available at https://github.com/trinhvg/VideoPath-LLaVA.
Authors: Kai Ruan, Mowen Huang, Ji-Rong Wen, Hao Sun
Abstract: Large Language Models (LLMs) show potential for complex reasoning, yet their capacity for emergent coordination in Multi-Agent Systems (MAS) when operating under strict constraints-such as limited local perception and communication, characteristic of natural swarms-remains largely unexplored, particularly concerning the nuances of swarm intelligence. Existing benchmarks often do not fully capture the unique challenges of decentralized coordination that arise when agents operate with incomplete spatio-temporal information. To bridge this gap, we introduce SwarmBench, a novel benchmark designed to systematically evaluate the swarm intelligence capabilities of LLMs acting as decentralized agents. SwarmBench features five foundational MAS coordination tasks within a configurable 2D grid environment, forcing agents to rely primarily on local sensory input (k x k view) and local communication. We propose metrics for coordination effectiveness and analyze emergent group dynamics. Evaluating several leading LLMs in a zero-shot setting, we find significant performance variations across tasks, highlighting the difficulties posed by local information constraints. While some coordination emerges, results indicate limitations in robust planning and strategy formation under uncertainty in these decentralized scenarios. Assessing LLMs under swarm-like conditions is crucial for realizing their potential in future decentralized systems. We release SwarmBench as an open, extensible toolkit-built upon a customizable and scalable physical system with defined mechanical properties. It provides environments, prompts, evaluation scripts, and the comprehensive experimental datasets generated, aiming to foster reproducible research into LLM-based MAS coordination and the theoretical underpinnings of Embodied MAS. Our code repository is available at https://github.com/x66ccff/swarmbench.
Authors: Shigeki Karita, Yuma Koizumi, Heiga Zen, Haruko Ishikawa, Robin Scheibler, Michiel Bacchiani
Abstract: Training data cleaning is a new application for generative model-based speech restoration (SR). This paper introduces Miipher-2, an SR model designed for million-hour scale data, for training data cleaning for large-scale generative models like large language models. Key challenges addressed include generalization to unseen languages, operation without explicit conditioning (e.g., text, speaker ID), and computational efficiency. Miipher-2 utilizes a frozen, pre-trained Universal Speech Model (USM), supporting over 300 languages, as a robust, conditioning-free feature extractor. To optimize efficiency and minimize memory, Miipher-2 incorporates parallel adapters for predicting clean USM features from noisy inputs and employs the WaneFit neural vocoder for waveform synthesis. These components were trained on 3,000 hours of multi-lingual, studio-quality recordings with augmented degradations, while USM parameters remained fixed. Experimental results demonstrate Miipher-2's superior or comparable performance to conventional SR models in word-error-rate, speaker similarity, and both objective and subjective sound quality scores across all tested languages. Miipher-2 operates efficiently on consumer-grade accelerators, achieving a real-time factor of 0.0078, enabling the processing of a million-hour speech dataset in approximately three days using only 100 such accelerators.
Authors: Qi Liu, Xinhao Zheng, Renqiu Xia, Xingzhi Qi, Qinxiang Cao, Junchi Yan
Abstract: As a seemingly self-explanatory task, problem-solving has been a significant component of science and engineering. However, a general yet concrete formulation of problem-solving itself is missing. With the recent development of AI-based problem-solving agents, the demand for process-level verifiability is rapidly increasing yet underexplored. To fill these gaps, we present a principled formulation of problem-solving as a deterministic Markov decision process; a novel framework, FPS (Formal Problem-Solving), which utilizes existing FTP (formal theorem proving) environments to perform process-verified problem-solving; and D-FPS (Deductive FPS), decoupling solving and answer verification for better human-alignment. The expressiveness, soundness and completeness of the frameworks are proven. We construct three benchmarks on problem-solving: FormalMath500, a formalization of a subset of the MATH500 benchmark; MiniF2F-Solving and PutnamBench-Solving, adaptations of FTP benchmarks MiniF2F and PutnamBench. For faithful, interpretable, and human-aligned evaluation, we propose RPE (Restricted Propositional Equivalence), a symbolic approach to determine the correctness of answers by formal verification. We evaluate four prevalent FTP models and two prompting methods as baselines, solving at most 23.77% of FormalMath500, 27.47% of MiniF2F-Solving, and 0.31% of PutnamBench-Solving.
Authors: Elif Akata, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, Eric Schulz
Abstract: LLMs are increasingly used in applications where they interact with humans and other agents. We propose to use behavioural game theory to study LLM's cooperation and coordination behaviour. We let different LLMs play finitely repeated $2\times2$ games with each other, with human-like strategies, and actual human players. Our results show that LLMs perform particularly well at self-interested games like the iterated Prisoner's Dilemma family. However, they behave sub-optimally in games that require coordination, like the Battle of the Sexes. We verify that these behavioural signatures are stable across robustness checks. We additionally show how GPT-4's behaviour can be modulated by providing additional information about its opponent and by using a "social chain-of-thought" (SCoT) strategy. This also leads to better scores and more successful coordination when interacting with human players. These results enrich our understanding of LLM's social behaviour and pave the way for a behavioural game theory for machines.
Authors: Qingyue Wang, Yanhe Fu, Yanan Cao, Shuai Wang, Zhiliang Tian, Liang Ding
Abstract: Recently, large language models (LLMs), such as GPT-4, stand out remarkable conversational abilities, enabling them to engage in dynamic and contextually relevant dialogues across a wide range of topics. However, given a long conversation, these chatbots fail to recall past information and tend to generate inconsistent responses. To address this, we propose to recursively generate summaries/ memory using large language models (LLMs) to enhance long-term memory ability. Specifically, our method first stimulates LLMs to memorize small dialogue contexts and then recursively produce new memory using previous memory and following contexts. Finally, the chatbot can easily generate a highly consistent response with the help of the latest memory. We evaluate our method on both open and closed LLMs, and the experiments on the widely-used public dataset show that our method can generate more consistent responses in a long-context conversation. Also, we show that our strategy could nicely complement both long-context (e.g., 8K and 16K) and retrieval-enhanced LLMs, bringing further long-term dialogue performance. Notably, our method is a potential solution to enable the LLM to model the extremely long context. The code and scripts will be released later.
Authors: Jingyuan Ma, Damai Dai, Zihang Yuan, Rui li, Weilin Luo, Bin Wang, Qun Liu, Lei Sha, Zhifang Sui
Abstract: Recent research have demonstrated LLMs' impressive performance in math and reasoning. However, the capacity of LLMs to address math problems under unconventional conditions, such as internal inconsistencies and flawed assumptions, remains largely unexplored. In this paper, we propose a novel benchmark Unreasonable Math Problem (UMP) designed to assess LLMs' ability to recognize and respond to unreasonability in math problem. The benchmark consists of a carefully curated collection of unreasonable math questions across diverse types. Based on extensive experiments covering 19 LLMs, we observe that even state-of-the-art models such as GPT-4o achieve only limited performance of 0.6 in UMP, while reasoning models such as DeepSeek-R1 are prone to overthinking and unstable. We further explore strategies for improving the recognition of unreasonable inputs, shedding light on both the possibility and limitations of LLMs in this challenging setting.
Authors: Zi-Yi Dou, Cheng-Fu Yang, Xueqing Wu, Kai-Wei Chang, Nanyun Peng
Abstract: Finetuning language agents with reasoning-action trajectories is effective, but obtaining these trajectories from human annotations or stronger models is costly and sometimes impractical. In this paper, we investigate the use of self-training in language agents, which can generate supervision from the agent itself, offering a promising alternative without relying on human or stronger model demonstrations. Self-training, however, requires high-quality model-generated samples, which are hard to obtain for challenging language agent tasks. To address this, we present Reflection-Reinforced Self-Training (Re-ReST), which uses a \textit{reflector} to refine low-quality generated samples during self-training. The reflector takes the agent's output and feedback from an external environment (e.g., unit test results in code generation) to produce improved samples. This technique enhances the quality of inferior samples and efficiently enriches the self-training dataset with higher-quality samples. We conduct extensive experiments on open-source language agents across tasks, including multi-hop question answering, sequential decision-making, code generation, visual question answering, and text-to-image generation. The results demonstrate the effectiveness of self-training and Re-ReST in language agent tasks, with self-training improving baselines by 7.6\% on HotpotQA and 28.4\% on AlfWorld, and Re-ReST further boosting performance by 2.0\% and 14.1\%, respectively. Our studies also confirm the efficiency of using a reflector to generate high-quality samples for self-training. Moreover, we demonstrate a method to employ reflection during inference without ground-truth feedback, addressing the limitation of previous reflection work. Our code is released at https://github.com/PlusLabNLP/Re-ReST.
Authors: Hussein Jawad, Yassine Chenik, Nicolas J. -B. Brunel
Abstract: The rapid adoption of Large Language Models (LLMs) has exposed critical security and ethical vulnerabilities, particularly their susceptibility to adversarial manipulations. This paper introduces QROA, a novel black-box jailbreak method designed to identify adversarial suffixes that can bypass LLM alignment safeguards when appended to a malicious instruction. Unlike existing suffix-based jailbreak approaches, QROA does not require access to the model's logit or any other internal information. It also eliminates reliance on human-crafted templates, operating solely through the standard query-response interface of LLMs. By framing the attack as an optimization bandit problem, QROA employs a surrogate model and token level optimization to efficiently explore suffix variations. Furthermore, we propose QROA-UNV, an extension that identifies universal adversarial suffixes for individual models, enabling one-query jailbreaks across a wide range of instructions. Testing on multiple models demonstrates Attack Success Rate (ASR) greater than 80\%. These findings highlight critical vulnerabilities, emphasize the need for advanced defenses, and contribute to the development of more robust safety evaluations for secure AI deployment. The code is made public on the following link: https://github.com/qroa/QROA
Authors: Zhenghao Zhou, Robert Frank, R. Thomas McCoy
Abstract: Large language models (LLMs) have shown the emergent capability of in-context learning (ICL). One line of research has claimed that ICL is functionally equivalent to gradient descent, a type of error-driven learning mechanism. In this paper, we introduce a new way of diagnosing whether ICL is functionally performing error-driven learning. Our approach is based on the inverse frequency effect (IFE) -- a phenomenon in which an agent's behavior is influenced to a greater degree when presented with improbable examples as compared to more likely ones. The IFE has previously been identified in psycholinguistics where humans exhibit the IFE in the context of structural priming (the tendency for people to produce sentence structures they have encountered recently). In that context, the IFE has been used as evidence that human structural priming must involve error-driven learning mechanisms. In our experiments, we simulated structural priming with ICL and found that LLMs indeed display the IFE, with the effect being stronger in larger models. We conclude that at least in the case we studied, ICL is indeed a type of error-driven learning, supporting the hypothesis that an error signal is implicitly computed in the forward pass during ICL. Our results suggest that both humans and LLMs make use of error-driven processing mechanisms in on-line processing.
Authors: Hourui Deng, Hongjie Zhang, Jie Ou, Chaosheng Feng
Abstract: Spatial reasoning in Large Language Models (LLMs) is the foundation for embodied intelligence. However, even in simple maze environments, LLMs still encounter challenges in long-term path-planning, primarily influenced by their spatial hallucination and context inconsistency hallucination by long-term reasoning. To address this challenge, this study proposes an innovative model, Spatial-to-Relational Transformation and Curriculum Q-Learning (S2RCQL). To address the spatial hallucination of LLMs, we propose the Spatial-to-Relational approach, which transforms spatial prompts into entity relations and paths representing entity relation chains. This approach fully taps the potential of LLMs in terms of sequential thinking. As a result, we design a path-planning algorithm based on Q-learning to mitigate the context inconsistency hallucination, which enhances the reasoning ability of LLMs. Using the Q-value of state-action as auxiliary information for prompts, we correct the hallucinations of LLMs, thereby guiding LLMs to learn the optimal path. Finally, we propose a reverse curriculum learning technique based on LLMs to further mitigate the context inconsistency hallucination. LLMs can rapidly accumulate successful experiences by reducing task difficulty and leveraging them to tackle more complex tasks. We performed comprehensive experiments based on Baidu's self-developed LLM: ERNIE-Bot 4.0. The results showed that our S2RCQL achieved a 23%--40% improvement in both success and optimality rates compared with advanced prompt engineering.
Authors: Makoto Fukushima, Shusuke Eshita, Hiroshige Fukuhara
Abstract: Color-word associations play a fundamental role in human cognition and design applications. Large Language Models (LLMs) have become widely available and have demonstrated intelligent behaviors in various benchmarks with natural conversation skills. However, their ability to replicate human color-word associations remains understudied. We compared multiple generations of LLMs (from GPT-3 to GPT-4o) against human color-word associations using data collected from over 10,000 Japanese participants, involving 17 colors and 80 words (10 word from eight categories) in Japanese. Our findings reveal a clear progression in LLM performance across generations, with GPT-4o achieving the highest accuracy in predicting the best voted word for each color and category. However, the highest median performance was approximately 50% even for GPT-4o with visual inputs (chance level of 10%). Moreover, we found performance variations across word categories and colors: while LLMs tended to excel in categories such as Rhythm and Landscape, they struggled with categories such as Emotions. Interestingly, color discrimination ability estimated from our color-word association data showed high correlation with human color discrimination patterns, consistent with previous studies. Thus, despite reasonable alignment in basic color discrimination, humans and LLMs still diverge systematically in the words they assign to those colors. Our study highlights both the advancements in LLM capabilities and their persistent limitations, raising the possibility of systematic differences in semantic memory structures between humans and LLMs in representing color-word associations.
Authors: Chengxing Xie, Bowen Li, Chang Gao, He Du, Wai Lam, Difan Zou, Kai Chen
Abstract: Large Language Models (LLMs) have demonstrated remarkable proficiency across a variety of complex tasks. One significant application of LLMs is in tackling software engineering challenges, particularly in resolving real-world tasks on GitHub by fixing code based on the issues reported by the users. However, many current approaches rely on proprietary LLMs, which limits reproducibility, accessibility, and transparency. The critical components of LLMs for addressing software engineering issues and how their capabilities can be effectively enhanced remain unclear. To address these challenges, we introduce SWE-Fixer, a novel open-source framework designed to effectively and efficiently resolve GitHub issues. SWE-Fixer comprises two essential modules: a code file retrieval module and a code editing module. The retrieval module employs BM25 along with a lightweight model to achieve coarse-to-fine file retrieval. Subsequently, the code editing module utilizes the other model to generate patches for the identified files. To mitigate the lack of publicly available datasets, we compile an extensive dataset that includes 110K GitHub issues along with their corresponding patches and train the two models of SWE-Fixer separately. We assess our approach on the SWE-Bench Lite and Verified benchmarks, achieving competitive performance among open-source models with scores of 22.0% and 30.2%. Furthermore, SWE-Fixer reaches state-of-the-art performance (24.7% on Lite and 32.8% on Verified) with PASS_TO_PASS (P2P) filtering. Additionally, our approach requires only two model calls per instance, making it significantly more efficient than existing methods. These results highlight the effectiveness of SWE-Fixer in real-world code-fixing scenarios. We will make our model, dataset, and code publicly available at https://github.com/InternLM/SWE-Fixer.
Authors: Huan Ma, Jingdong Chen, Joey Tianyi Zhou, Guangyu Wang, Changqing Zhang
Abstract: Over the past few years, Large Language Models (LLMs) have developed rapidly and are widely applied in various domains. However, LLMs face the issue of hallucinations, generating responses that may be unreliable when the models lack relevant knowledge. To be aware of potential hallucinations, uncertainty estimation methods have been introduced, and most of them have confirmed that reliability lies in critical tokens. However, probability-based methods perform poorly in identifying token reliability, limiting their practical utility. In this paper, we reveal that the probability-based method fails to estimate token reliability due to the loss of evidence strength information which is accumulated in the training stage. Therefore, we present Logits-induced token uncertainty (LogTokU), a framework for estimating decoupled token uncertainty in LLMs, enabling real-time uncertainty estimation without requiring multiple sampling processes. We employ evidence modeling to implement LogTokU and use the estimated uncertainty to guide downstream tasks. The experimental results demonstrate that LogTokU has significant effectiveness and promise.
Authors: Disen Lan, Weigao Sun, Jiaxi Hu, Jusen Du, Yu Cheng
Abstract: Transformers with linear recurrent modeling offer linear-time training and constant-memory inference. Despite their demonstrated efficiency and performance, pretraining such non-standard architectures from scratch remains costly and risky. The linearization of large language models (LLMs) transforms pretrained standard models into linear recurrent structures, enabling more efficient deployment. However, current linearization methods typically introduce additional feature map modules that require extensive fine-tuning and overlook the gating mechanisms used in state-of-the-art linear recurrent models. To address these issues, this paper presents Liger, short for Linearizing LLMs to gated recurrent structures. Liger is a novel approach for converting pretrained LLMs into gated linear recurrent models without adding extra parameters. It repurposes the pretrained key matrix weights to construct diverse gating mechanisms, facilitating the formation of various gated recurrent structures while avoiding the need to train additional components from scratch. Using lightweight fine-tuning with Low-Rank Adaptation (LoRA), Liger restores the performance of the linearized gated recurrent models to match that of the original LLMs. Additionally, we introduce Liger Attention, an intra-layer hybrid attention mechanism, which significantly recovers 93\% of the Transformer-based LLM at 0.02\% pre-training tokens during the linearization process, achieving competitive results across multiple benchmarks, as validated on models ranging from 1B to 8B parameters. Code is available at https://github.com/OpenSparseLLMs/Linearization.
Authors: Ben Hutchinson, Celeste Rodr\'iguez Louro, Glenys Collard, Ned Cooper
Abstract: In Australia, post-contact language varieties, including creoles and local varieties of international languages, emerged as a result of forced contact between Indigenous communities and English speakers. These contact varieties are widely used, yet are poorly supported by language technologies. This gap presents barriers to participation in civil and economic society for Indigenous communities using these varieties, and reproduces minoritisation of contemporary Indigenous sociolinguistic identities. This paper concerns three questions regarding this context. First, can speech technologies support speakers of Australian Aboriginal English, a local indigenised variety of English? Second, what risks are inherent in such a project? Third, what technology development practices are appropriate for this context, and how can researchers integrate meaningful community participation in order to mitigate risks? We argue that opportunities do exist -- as well as risks -- and demonstrate this through a case study exploring design practices in a real-world project aiming to improve speech technologies for Australian Aboriginal English. We discuss how we integrated culturally appropriate and participatory processes throughout the project. We call for increased support for languages used by Indigenous communities, including contact varieties, which provide practical economic and socio-cultural benefits, provided that participatory and culturally safe practices are enacted.
Authors: Bryan Wilie, Samuel Cahyawijaya, Junxian He, Pascale Fung
Abstract: Large language models (LLMs) trained on massive multilingual datasets hint at the formation of interlingual constructs--a shared subspace in the representation space. However, evidence regarding this phenomenon is mixed, leaving it unclear whether these models truly develop unified interlingual representations, or present a partially aligned constructs. We explore 31 diverse languages varying on their resource-levels, typologies, and geographical regions; and find that multilingual LLMs exhibit inconsistent cross-lingual alignments. To address this, we propose an interlingual representation framework identifying both the shared interlingual semantic subspace and fragmented components, existed due to representational limitations. We introduce Interlingual Local Overlap (ILO) score to quantify interlingual alignment by comparing the local neighborhood structures of high-dimensional representations. We utilize ILO to investigate the impact of single-language fine-tuning on the interlingual representations in multilingual LLMs. Our results indicate that training exclusively on a single language disrupts the alignment in early layers, while freezing these layers preserves the alignment of interlingual representations, leading to improved cross-lingual generalization. These results validate our framework and metric for evaluating interlingual representation, and further underscore that interlingual alignment is crucial for scalable multilingual learning.
Authors: Zhangcheng Qiang
Abstract: Hallucinations are often inevitable in downstream tasks using large language models (LLMs). To tackle the substantial challenge of addressing hallucinations for LLM-based ontology matching (OM) systems, we introduce a new benchmark dataset called OAEI-LLM-T. The dataset evolves from the TBox (i.e. schema-matching) datasets in the Ontology Alignment Evaluation Initiative (OAEI), capturing hallucinations of different LLMs performing OM tasks. These OM-specific hallucinations are carefully classified into two primary categories and six sub-categories. We showcase the usefulness of the dataset in constructing the LLM leaderboard and fine-tuning foundational LLMs for LLM-based OM systems.
Authors: Junichiro Niimi
Abstract: With the advance of large language models (LLMs), LLMs have been utilized for the various tasks. However, the issues of variability and reproducibility of results from each trial of LLMs have been largely overlooked in existing literature while actual human annotation uses majority voting to resolve disagreements among annotators. Therefore, this study introduces the straightforward ensemble strategy to a sentiment analysis using LLMs. As the results, we demonstrate that the ensemble of multiple inference using medium-sized LLMs produces more robust and accurate results than using a large model with a single attempt with reducing RMSE by 18.6%.
Authors: Roman Abramov, Felix Steinbauer, Gjergji Kasneci
Abstract: Transformers have achieved great success in numerous NLP tasks but continue to exhibit notable gaps in multi-step factual reasoning, especially when real-world knowledge is sparse. Recent advances in grokking have demonstrated that neural networks can transition from memorizing to perfectly generalizing once they detect underlying logical patterns - yet these studies have primarily used small, synthetic tasks. In this paper, for the first time, we extend grokking to real-world factual data and address the challenge of dataset sparsity by augmenting existing knowledge graphs with carefully designed synthetic data to raise the ratio $\phi_r$ of inferred facts to atomic facts above the threshold required for grokking. Surprisingly, we find that even factually incorrect synthetic data can strengthen emergent reasoning circuits rather than degrade accuracy, as it forces the model to rely on relational structure rather than memorization. When evaluated on multi-hop reasoning benchmarks, our approach achieves up to 95-100% accuracy on 2WikiMultiHopQA - substantially improving over strong baselines and matching or exceeding current state-of-the-art results. We further provide an in-depth analysis of how increasing $\phi_r$ drives the formation of generalizing circuits inside Transformers. Our findings suggest that grokking-based data augmentation can unlock implicit multi-hop reasoning capabilities, opening the door to more robust and interpretable factual reasoning in large-scale language models.
Authors: Andrew K. Lampinen, Arslan Chaudhry, Stephanie C. Y. Chan, Cody Wild, Diane Wan, Alex Ku, J\"org Bornschein, Razvan Pascanu, Murray Shanahan, James L. McClelland
Abstract: Large language models exhibit exciting capabilities, yet can show surprisingly narrow generalization from finetuning. E.g. they can fail to generalize to simple reversals of relations they are trained on, or fail to make simple logical deductions based on trained information. These failures to generalize from fine-tuning can hinder practical application of these models. On the other hand, language models' in-context learning shows different inductive biases, and can generalize better in some cases. Here, we explore these differences in generalization between in-context- and fine-tuning-based learning. To do so, we constructed several novel datasets to evaluate and improve models' abilities to generalize from finetuning data. The datasets are designed to create clean tests of generalization, by isolating the knowledge in the dataset from that in pretraining. We expose pretrained large models to controlled subsets of the information in these datasets -- either in context, or through fine-tuning -- and evaluate their performance on test sets that require various types of generalization. We find overall that in data-matched settings, in-context learning can generalize more flexibly than fine-tuning (though we also find some qualifications of prior findings, such as cases when fine-tuning can generalize to reversals embedded in a larger structure of knowledge). We build on these findings to propose a method to enable improved generalization from fine-tuning: adding in-context inferences to finetuning data. We show that this method improves generalization across various splits of our datasets and other benchmarks. Our results have implications for understanding the inductive biases of different modes of learning in language models, and practically improving their performance.
Authors: Yingquan Chen, Qianmu Li, Xiaocong Wu, Huifeng Li, Qing Chang
Abstract: Generating high-quality steganographic text is a fundamental challenge in the field of generative linguistic steganography. This challenge arises primarily from two aspects: firstly, the capabilities of existing models in text generation are limited; secondly, embedding algorithms fail to effectively mitigate the negative impacts of sensitive information's properties, such as semantic content or randomness. Specifically, to ensure that the recipient can accurately extract hidden information, embedding algorithms often have to consider selecting candidate words with relatively low probabilities. This phenomenon leads to a decrease in the number of high-probability candidate words and an increase in low-probability candidate words, thereby compromising the semantic coherence and logical fluency of the steganographic text and diminishing the overall quality of the generated steganographic material. To address this issue, this paper proposes a novel embedding algorithm, character-based diffusion embedding algorithm (CDEA). Unlike existing embedding algorithms that strive to eliminate the impact of sensitive information's properties on the generation process, CDEA leverages sensitive information's properties. It enhances the selection frequency of high-probability candidate words in the candidate pool based on general statistical properties at the character level and grouping methods based on power-law distributions, while reducing the selection frequency of low-probability candidate words in the candidate pool. Furthermore, to ensure the effective transformation of sensitive information in long sequences, we also introduce the XLNet model. Experimental results demonstrate that the combination of CDEA and XLNet significantly improves the quality of generated steganographic text, particularly in terms of perceptual-imperceptibility.
Authors: Tianyu Zong, Hongzhu Yi, Bingkang Shi, Yuanxiang Wang, Jungang Xu
Abstract: Unsupervised contrastive learning has become a hot research topic in natural language processing. Existing works usually aim at constraining the orientation distribution of the representations of positive and negative samples in the high-dimensional semantic space in contrastive learning, but the semantic representation tensor possesses both modulus and orientation features, and the existing works ignore the modulus feature of the representations and cause insufficient contrastive learning. % Therefore, we firstly propose a training objective that aims at modulus constraints on the semantic representation tensor, to strengthen the alignment between the positive samples in contrastive learning. Therefore, we first propose a training objective that is designed to impose modulus constraints on the semantic representation tensor, to strengthen the alignment between positive samples in contrastive learning. Then, the BERT-like model suffers from the phenomenon of sinking attention, leading to a lack of attention to CLS tokens that aggregate semantic information. In response, we propose a cross-attention structure among the twin-tower ensemble models to enhance the model's attention to CLS token and optimize the quality of CLS Pooling. Combining the above two motivations, we propose a new \textbf{J}oint \textbf{T}ensor representation modulus constraint and \textbf{C}ross-attention unsupervised contrastive learning \textbf{S}entence \textbf{E}mbedding representation framework JTCSE, which we evaluate in seven semantic text similarity computation tasks, and the experimental results show that JTCSE's twin-tower ensemble model and single-tower distillation model outperform the other baselines and become the current SOTA. In addition, we have conducted an extensive zero-shot downstream task evaluation, which shows that JTCSE outperforms other baselines overall on more than 130 tasks.
Authors: Chantal Pellegrini, Ege \"Ozsoy, Benjamin Busam, Nassir Navab, Matthias Keicher
Abstract: Conversational AI tools that can generate and discuss clinically correct radiology reports for a given medical image have the potential to transform radiology. Such a human-in-the-loop radiology assistant could facilitate a collaborative diagnostic process, thus saving time and improving the quality of reports. Towards this goal, we introduce RaDialog, the first thoroughly evaluated and publicly available large vision-language model for radiology report generation and interactive dialog. RaDialog effectively integrates visual image features and structured pathology findings with a large language model (LLM) while simultaneously adapting it to a specialized domain using parameter-efficient fine-tuning. To keep the conversational abilities of the underlying LLM, we propose a comprehensive, semi-automatically labeled, image-grounded instruct dataset for chest X-ray radiology tasks. By training with this dataset, our method achieves state-of-the-art clinical correctness in report generation and shows impressive abilities in interactive tasks such as correcting reports and answering questions, serving as a foundational step toward clinical dialog systems. Our code is available on github: https://github.com/ChantalMP/RaDialog.
Authors: Hung Manh Pham, Aaqib Saeed, Dong Ma
Abstract: The accurate interpretation of Electrocardiogram (ECG) signals is pivotal for diagnosing cardiovascular diseases. Integrating ECG signals with accompanying textual reports further holds immense potential to enhance clinical diagnostics by combining physiological data and qualitative insights. However, this integration faces significant challenges due to inherent modality disparities and the scarcity of labeled data for robust cross-modal learning. To address these obstacles, we propose D-BETA, a novel framework that pre-trains ECG and text data using a contrastive masked auto-encoder architecture. D-BETA uniquely combines the strengths of generative with boosted discriminative capabilities to achieve robust cross-modal representations. This is accomplished through masked modality modeling, specialized loss functions, and an improved negative sampling strategy tailored for cross-modal alignment. Extensive experiments on five public datasets across diverse downstream tasks demonstrate that D-BETA significantly outperforms existing methods, achieving an average AUC improvement of 15% in linear probing with only one percent of training data and 2% in zero-shot performance without requiring training data over state-of-the-art models. These results highlight the effectiveness of D-BETA, underscoring its potential to advance automated clinical diagnostics through multi-modal representations. Our sample code and checkpoint are made available at https://github.com/manhph2211/D-BETA.
Authors: Grace Luo, Trevor Darrell, Amir Bar
Abstract: Autoregressive vision-language models (VLMs) can handle many tasks within a single model, yet the representations that enable this capability remain opaque. We find that VLMs align conceptually equivalent inputs into a shared task vector, which is invariant to modality (text, image) and format (examples, instruction), and may simplify VLM processing. We measure this alignment via cross-modal transfer -- the ability of a task vector derived in one modality to trigger the correct generation in another -- on a range of tasks and model architectures. Although the task vector is highly compressed, we find that this single vector outperforms prompting the model with the full task information, unique to this cross-modal case. Furthermore, we show that task vectors can be transferred from a base language model to its fine-tuned vision-language counterpart, and that they can be derived solely from instructions without the need for examples. Taken together, our findings shed light on how VLMs internally process task information, and how they map different modalities into common semantic representations. Project page: https://vlm-cross-modal-reps.github.io.
Authors: Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Chunyu Wang, Xiyang Dai, Dongdong Chen, Chong Luo, Lili Qiu
Abstract: CLIP is a foundational multimodal model that aligns image and text features into a shared representation space via contrastive learning on large-scale image-text pairs. Its effectiveness primarily stems from the use of natural language as rich supervision. Motivated by the remarkable advancements in large language models (LLMs), this work explores how LLMs' superior text understanding and extensive open-world knowledge can enhance CLIP's capability, especially for processing longer and more complex image captions. We propose an efficient post-training strategy that integrates LLMs into pretrained CLIP. To address the challenge posed by the autoregressive nature of LLMs, we introduce a caption-to-caption contrastive fine-tuning framework, significantly enhancing the discriminative quality of LLM outputs. Extensive experiments demonstrate that our approach outperforms LoRA-based methods, achieving nearly fourfold faster training with superior performance. Furthermore, we validate substantial improvements over state-of-the-art models such as CLIP, EVA02, and SigLip2 across various zero-shot multimodal retrieval tasks, cross-lingual retrieval tasks, and multimodal language model pretraining.
Authors: Yingying Fang, Zihao Jin, Shaojie Guo, Jinda Liu, Zhiling Yue, Yijian Gao, Junzhi Ning, Zhi Li, Simon Walsh, Guang Yang
Abstract: Despite significant advancements in automated report generation, the opaqueness of text interpretability continues to cast doubt on the reliability of the content produced. This paper introduces a novel approach to identify specific image features in X-ray images that influence the outputs of report generation models. Specifically, we propose Cyclic Vision-Language Manipulator CVLM, a module to generate a manipulated X-ray from an original X-ray and its report from a designated report generator. The essence of CVLM is that cycling manipulated X-rays to the report generator produces altered reports aligned with the alterations pre-injected into the reports for X-ray generation, achieving the term "cyclic manipulation". This process allows direct comparison between original and manipulated X-rays, clarifying the critical image features driving changes in reports and enabling model users to assess the reliability of the generated texts. Empirical evaluations demonstrate that CVLM can identify more precise and reliable features compared to existing explanation methods, significantly enhancing the transparency and applicability of AI-generated reports.
Authors: Jiangang Hao, Wenju Cui, Patrick Kyllonen, Emily Kerzabi, Lei Liu, Michael Flor
Abstract: Collaborative problem solving (CPS) is widely recognized as a critical 21st-century skill. Assessing CPS depends heavily on coding the communication data using a construct-relevant framework, and this process has long been a major bottleneck to scaling up such assessments. Based on five datasets and two coding frameworks, we demonstrate that ChatGPT can code communication data to a satisfactory level, though performance varies across ChatGPT models, and depends on the coding framework and task characteristics. Interestingly, newer reasoning-focused models such as GPT-o1-mini and GPT-o3-mini do not necessarily yield better coding results. Additionally, we show that refining prompts based on feedback from miscoded cases can improve coding accuracy in some instances, though the effectiveness of this approach is not consistent across all tasks. These findings offer practical guidance for researchers and practitioners in developing scalable, efficient methods to analyze communication data in support of 21st-century skill assessment.
Authors: Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, Junxian He
Abstract: DeepSeek-R1 has shown that long chain-of-thought (CoT) reasoning can naturally emerge through a simple reinforcement learning (RL) framework with rule-based rewards, where the training may directly start from the base models-a paradigm referred to as zero RL training. Most recent efforts to reproduce zero RL training have primarily focused on the Qwen2.5 model series, which may not be representative as we find the base models already exhibit strong instruction-following and self-reflection abilities. In this work, we investigate zero RL training across 10 diverse base models, spanning different families and sizes including LLama3-8B, Mistral-7B/24B, DeepSeek-Math-7B, Qwen2.5-math-7B, and all Qwen2.5 models from 0.5B to 32B. Leveraging several key design strategies-such as adjusting format reward and controlling query difficulty-we achieve substantial improvements in both reasoning accuracy and response length across most settings. However, by carefully monitoring the training dynamics, we observe that different base models exhibit distinct patterns during training. For instance, the increased response length does not always correlate with the emergence of certain cognitive behaviors such as verification (i.e., the "aha moment"). Notably, we observe the "aha moment" for the first time in small models not from the Qwen family. We share the key designs that enable successful zero RL training, along with our findings and practices. To facilitate further research, we open-source the code, models, and analysis tools.
Authors: Quang P. M. Pham, Khoi T. N. Nguyen, Nhi H. Doan, Cuong A. Pham, Kentaro Inui, Dezhen Song
Abstract: Efficient path planning in robotics, particularly within large-scale, dynamic environments, remains a significant hurdle. While Large Language Models (LLMs) offer strong reasoning capabilities, their high computational cost and limited adaptability in dynamic scenarios hinder real-time deployment on edge devices. We present SmallPlan -- a novel framework leveraging LLMs as teacher models to train lightweight Small Language Models (SLMs) for high-level path planning tasks. In SmallPlan, the SLMs provide optimal action sequences to navigate across scene graphs that compactly represent full-scaled 3D scenes. The SLMs are trained in a simulation-powered, interleaved manner with LLM-guided supervised fine-tuning (SFT) and reinforcement learning (RL). This strategy not only enables SLMs to successfully complete navigation tasks but also makes them aware of important factors like travel distance and number of trials. Through experiments, we demonstrate that the fine-tuned SLMs perform competitively with larger models like GPT-4o on sequential path planning, without suffering from hallucination and overfitting. SmallPlan is resource-efficient, making it well-suited for edge-device deployment and advancing practical autonomous robotics.
Authors: Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, Yang Yue, Matthieu Lin, Shenzhi Wang, Qingyun Wu, Zilong Zheng, Gao Huang
Abstract: Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from outcome-based rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability by using a code executor to both validate proposed code reasoning tasks and verify answers, serving as an unified source of verifiable reward to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.
Authors: Fangming Cui, Yonggang Zhang, Xuan Wang, Xinmei Tian, Jun Yu
Abstract: Recent developments in prompt learning of large vision-language models have significantly improved performance in target-specific tasks. However, these prompt optimizing methods often struggle to tackle the target-unspecific or generalizable tasks effectively. It may be attributed to the fact that overfitting training causes the model to forget its general knowledge having strong promotion on target-unspecific tasks. To alleviate this issue, we propose a novel Features Matrix (FM) regularization approach designed to enhance these models on target-unspecific tasks. Our method extracts and leverages general knowledge, shaping a Features Matrix (FM). Specifically, the FM captures the semantics of diverse inputs from a deep and fine perspective, preserving essential general knowledge, which mitigates the risk of overfitting. Representative evaluations demonstrate that: 1) the FM is compatible with existing frameworks as a generic and flexible module, and 2) the FM significantly showcases its effectiveness in enhancing target-unspecific tasks, achieving state-of-the-art performance.