Mathematical reasoning is a cornerstone of artificial general intelligence and a primary benchmark for evaluating the capabilities of Large Language Models (LLMs). While state-of-the-art models show promise, they often falter when faced with complex problems that demand deep conceptual understanding and intricate, multi-step deliberation. To address this challenge, we introduce JT-Math-8B, a series of open-source models comprising base, instruct, and thinking versions, built upon a systematic, multi-stage optimization framework. Our pre-training corpus is a high-quality, 210B-token dataset curated through a dedicated data pipeline that uses model-based validation to ensure quality and diversity. The Instruct Model is optimized for direct, concise answers through Supervised Fine-Tuning (SFT) and a GRPO-based reinforcement learning (RL) method. The Thinking Model is trained for complex problem-solving using a Long Chain-of-Thought (Long CoT) approach, combining SFT with a novel, multi-stage RL curriculum that progressively increases task difficulty and context length up to 32K tokens. JT-Math-8B achieves state-of-the-art results among open-source models of similar size, surpassing prominent models like OpenAI's O1-mini and GPT-4o , and demonstrating superior performance on competition-level mathematics.

--The vulnerability of cities to natural disasters has increased with urbanization and climate change, making it more important to predict human mobility in the disaster scenarios for downstream tasks including location-based early disaster warning and pre-allocating rescue resources, etc. However, existing human mobility prediction models are mainly designed for normal scenarios, and fail to adapt to disaster scenarios due to the shift of human mobility patterns under disaster . T o address this issue, we introduce DisasterMobLLM, a mobility prediction framework for disaster scenarios that can be integrated into existing deep mobility prediction methods by leveraging LLMs to model the mobility intention and transferring the common knowledge of how different disasters affect mobility intentions between cities. This framework utilizes a RAG-Enhanced Intention Predictor to forecast the next intention, refines it with an LLM-based Intention Refiner, and then maps the intention to an exact location using an Intention-Modulated Location Predictor . Extensive experiments illustrate that DisasterMobLLM can achieve a 32.8% improvement in terms of Acc@1 and a 35.0% ITH the rapid urbanization [1] and climate change [2], cities across the world are becoming increasingly vulnerable to natural disasters (e.g., heavy rains), exposing more human lives and properties at risk. To tackle these challenges, a fundamental research problem is to predict human mobility during disaster scenarios, which can support a wide spectrum of downstream emergency response tasks including location-based early disaster warning [3]-[5], pre-allocating rescue resources [6], and planning humanitarian relief [7], etc. As a classic machine learning problem, human mobility prediction has been studied for decades; however, most existing work [8], [9] has focused on normal scenarios rather than disaster scenarios. As illustrated in Figure 1(a) and (b), we employ two representative algorithms trained in the normal scenario, i.e., DeepMove [8] and Flashback [10], to predict human mobility in normal scenarios and disaster scenarios, respectively. Their performance in disaster scenarios significantly decreases compared with normal scenarios, with an average relative performance gap of 46.4% and 24.5% in terms of accuracy and mean reciprocal rank, respectively.

Vector databases typically rely on approximate nearest neighbor (ANN) search to retrieve the top-k closest vectors to a query in embedding space. While effective, this approach often yields semantically redundant results, missing the diversity and contextual richness required by applications such as retrieval-augmented generation (RAG), multi-hop QA, and memory-augmented agents. We introduce a new retrieval paradigm: semantic compression, which aims to select a compact, representative set of vectors that captures the broader semantic structure around a query. We formalize this objective using principles from submodular optimization and information geometry, and show that it generalizes traditional top-k retrieval by prioritizing coverage and diversity. To operationalize this idea, we propose graph-augmented vector retrieval, which overlays semantic graphs (e.g., kNN or knowledge-based links) atop vector spaces to enable multi-hop, context-aware search. We theoretically analyze the limitations of proximity-based retrieval under high-dimensional concentration and highlight how graph structures can improve semantic coverage. Our work outlines a foundation for meaning-centric vector search systems, emphasizing hybrid indexing, diversity-aware querying, and structured semantic retrieval. We make our implementation publicly available to foster future research in this area.

In Table-to-Text (T2T) generation, existing approaches predominantly focus on providing objective descriptions of tabular data. However, generating text that incorporates subjectivity, where subjectivity refers to interpretations beyond raw numerical data, remains underexplored. To address this, we introduce a novel pipeline that leverages intermediate representations to generate both objective and subjective text from tables. Our three-stage pipeline consists of: 1) extraction of Resource Description Framework (RDF) triples, 2) aggregation of text into coherent narratives, and 3) infusion of subjectivity to enrich the generated text. By incorporating RDFs, our approach enhances factual accuracy while maintaining interpretability. Unlike large language models (LLMs) such as GPT-3.5, Mistral-7B, and Llama-2, our pipeline employs smaller, fine-tuned T5 models while achieving comparable performance to GPT-3.5 and outperforming Mistral-7B and Llama-2 in several metrics. We evaluate our approach through quantitative and qualitative analyses, demonstrating its effectiveness in balancing factual accuracy with subjective interpretation. To the best of our knowledge, this is the first work to propose a structured pipeline for T2T generation that integrates intermediate representations to enhance both factual correctness and subjectivity.

Although LLMs have attained significant success in high-resource languages, their capacity in low-resource linguistic environments like Kannada and Arabic is not yet fully understood. This work benchmarking the performance of multilingual and monolingual Large Language Models (LLMs) across Arabic, English, and Indic languages, with particular emphasis on the effects of model compression strategies such as pruning and quantization. Findings shows significant performance differences driven by linguistic diversity and resource availability on SOTA LLMS as BLOOMZ, AceGPT, Jais, LLaMA-2, XGLM, and AraGPT2. We find that multilingual versions of the model outperform their language-specific counterparts across the board, indicating substantial cross-lingual transfer benefits. Quantization (4-bit and 8-bit) is effective in maintaining model accuracy while promoting efficiency, but aggressive pruning significantly compromises performance, especially in bigger models. Our findings pinpoint key strategies to construct scalable and fair multilingual NLP solutions and underscore the need for interventions to address hallucination and generalization errors in the low-resource setting.

Imagine a humanoid that can safely and creatively dance with a human, adapting to its partner's proficiency, using haptic signaling as a primary form of communication. While today's AI systems excel at text or voice-based interaction with large language models, human communication extends far beyond text-it includes embodied movement, timing, and physical coordination. Modeling coupled interaction between two agents poses a formidable challenge: it is continuous, bidirectionally reactive, and shaped by individual variation. We present CoMPAS3D, the largest and most diverse motion capture dataset of improvised salsa dancing, designed as a challenging testbed for interactive, expressive humanoid AI. The dataset includes 3 hours of leader-follower salsa dances performed by 18 dancers spanning beginner, intermediate, and professional skill levels. For the first time, we provide fine-grained salsa expert annotations, covering over 2,800 move segments, including move types, combinations, execution errors and stylistic elements. We draw analogies between partner dance communication and natural language, evaluating CoMPAS3D on two benchmark tasks for synthetic humans that parallel key problems in spoken language and dialogue processing: leader or follower generation with proficiency levels (speaker or listener synthesis), and duet (conversation) generation. Towards a long-term goal of partner dance with humans, we release the dataset, annotations, and code, along with a multitask SalsaAgent model capable of performing all benchmark tasks, alongside additional baselines to encourage research in socially interactive embodied AI and creative, expressive humanoid motion generation.

Image-based product attribute prediction in e-commerce is a crucial task with numerous applications. The supervised fine-tuning of Vision Language Models (VLMs) faces significant scale challenges due to the cost of manual or API based annotation. In this paper, we investigate label-efficient semi-supervised fine-tuning strategies for compact VLMs (2B-3B parameters) that leverage unlabeled product listings through Direct Preference Optimization (DPO). Beginning with a small, API-based, annotated, and labeled set, we first employ PEFT to train low-rank adapter modules. T o update the adapter weights with unlabeled data, we generate multiple reasoning-and-answer chains per unlabeled sample and segregate these chains into preferred and dispreferred based on self-consistency. W e then fine-tune the model with DPO loss and use the updated model for the next iteration. By using PEFT fine-tuning with DPO, our method achieves efficient convergence with minimal compute overhead. On a dataset spanning twelve e-commerce verticals, DPO-based fine-tuning, which utilizes only unlabeled data, demonstrates a significant improvement over the supervised model. Moreover, experiments demonstrate that accuracy with DPO training improves with more unlabeled data, indicating that a large pool of unlabeled samples can be effectively leveraged to improve performance.

Due to the remarkable capabilities and growing impact of large language models (LLMs), they have been deeply integrated into many aspects of society. Thus, ensuring their alignment with human values and intentions has emerged as a critical challenge. This survey provides a comprehensive overview of practical alignment techniques, training protocols, and empirical findings in LLM alignment. We analyze the development of alignment methods across diverse paradigms, characterizing the fundamental trade-offs between core alignment objectives. Our analysis shows that while supervised fine-tuning enables basic instruction-following, preference-based methods offer more flexibility for aligning with nuanced human intent. We discuss state-of-the-art techniques, including Direct Preference Optimization (DPO), Constitutional AI, brain-inspired methods, and alignment uncertainty quantification (AUQ), highlighting their approaches to balancing quality and efficiency. We review existing evaluation frameworks and benchmarking datasets, emphasizing limitations such as reward misspecification, distributional robustness, and scalable oversight. We summarize strategies adopted by leading AI labs to illustrate the current state of practice. We conclude by outlining open problems in oversight, value pluralism, robustness, and continuous alignment. This survey aims to inform both researchers and practitioners navigating the evolving landscape of LLM alignment.

The intersection of AI and legal systems presents a growing need for tools that support legal education, particularly in under-resourced languages such as Romanian. In this work, we aim to evaluate the capabilities of Large Language Models (LLMs) and Vision-Language Models (VLMs) in understanding and reasoning about Romanian driving law through textual and visual question-answering tasks. To facilitate this, we introduce RoD-TAL, a novel multimodal dataset comprising Romanian driving test questions, text-based and image-based, alongside annotated legal references and human explanations. We implement and assess retrieval-augmented generation (RAG) pipelines, dense retrievers, and reasoning-optimized models across tasks including Information Retrieval (IR), Question Answering (QA), Visual IR, and Visual QA. Our experiments demonstrate that domain-specific fine-tuning significantly enhances retrieval performance. At the same time, chain-of-thought prompting and specialized reasoning models improve QA accuracy, surpassing the minimum grades required to pass driving exams. However, visual reasoning remains challenging, highlighting the potential and the limitations of applying LLMs and VLMs to legal education.

Understanding clients' thoughts and beliefs is fundamental in counseling, yet current evaluations of LLM therapists often fail to assess this ability. Existing evaluation methods rely on client simulators that clearly disclose internal states to the therapist, making it difficult to determine whether an LLM therapist can uncover unexpressed perspectives. To address this limitation, we introduce MindVoyager, a novel evaluation framework featuring a controllable and realistic client simulator which dynamically adapts itself based on the ongoing counseling session, offering a more realistic and challenging evaluation environment. We further introduce evaluation metrics that assess the exploration ability of LLM therapists by measuring their thorough understanding of client's beliefs and thoughts.