Recent progress in large language models (LLMs) has enabled understanding of both speech and non-speech audio, but exposing new safety risks emerging from complex audio inputs that are inadequately handled by current safeguards. We introduce SACRED-Bench (Speech-Audio Composition for RED-teaming) to evaluate the robustness of LLMs under complex audio-based attacks. Unlike existing perturbation-based methods that rely on noise optimization or white-box access, SACRED-Bench exploits speech-audio composition mechanisms. SACRED-Bench adopts three mechanisms: (a) speech overlap and multi-speaker dialogue, which embeds harmful prompts beneath or alongside benign speech; (b) speech-audio mixture, which imply unsafe intent via non-speech audio alongside benign speech or audio; and (c) diverse spoken instruction formats (open-ended QA, yes/no) that evade text-only filters. Experiments show that, even Gemini 2.5 Pro, the state-of-the-art proprietary LLM, still exhibits 66% attack success rate in SACRED-Bench test set, exposing vulnerabilities under cross-modal, speech-audio composition attacks. To bridge this gap, we propose SALMONN-Guard, a safeguard LLM that jointly inspects speech, audio, and text for safety judgments, reducing attack success down to 20%. Our results highlight the need for audio-aware defenses for the safety of multimodal LLMs. The benchmark and SALMONN-Guard checkpoints can be found at https://huggingface.co/datasets/tsinghua-ee/SACRED-Bench. Warning: this paper includes examples that may be offensive or harmful.

Abstract--The data-centric paradigm has emerged as a pivotal direction in artificial intelligence (AI), relying on high-quality training data. This shift is especially critical in the T ext-to-SQL task, where model performance is constrained by the scarcity, limited diversity, and structural simplicity of existing datasets. Our framework operates along six augmentation dimensions and integrates an end-to-end pipeline featuring SQL execution verification, natural language (NL) question generation, chain-of-thought (CoT) reasoning trace generation, and data classification. A modular Database Manager further ensures cross-database compatibility and scalability. This approach enables structure-aware example matching by modeling fine-grained alignments between NL questions and SQL queries. Our work establishes a scalable, data-centric foundation for advancing T ext-to-SQL systems and underscores the indispensable role of structured, high-fidelity data in modern AI development. Our code is available at https://github.com/T In recent years, the data-centric artificial intelligence (AI) paradigm has garnered increasing attention [1], [2]. Traditional algorithm-centric approaches primarily focus on expanding model architectures and optimizing learning algorithms. However, in many cutting-edge fields, the main bottleneck of development has gradually shifted from algorithmic complexity to the availability of high-quality data. Continuous optimization of algorithms is facing diminishing marginal returns, while vast amounts of data remain underutilized, containing immense potential value. Taking large language models (LLMs) as an example, their generalization ability and robustness highly depend on the breadth and quality of the training data. Similarly, in downstream tasks such as domain adaptation, high-quality data can serve both as reference material for generating answers and as guidance for solving problems [4].

Large language models (LLMs) have shown great promise in the medical domain, achieving strong performance on several benchmarks. However, they continue to underperform in real-world medical scenarios, which often demand stronger context-awareness, i.e., the ability to recognize missing or critical details (e.g., user identity, medical history, risk factors) and provide safe, helpful, and contextually appropriate responses. To address this issue, we propose Multifaceted Self-Refinement (MuSeR), a data-driven approach that enhances LLMs' context-awareness along three key facets (decision-making, communication, and safety) through self-evaluation and refinement. Specifically, we first design a attribute-conditioned query generator that simulates diverse real-world user contexts by varying attributes such as role, geographic region, intent, and degree of information ambiguity. An LLM then responds to these queries, self-evaluates its answers along three key facets, and refines its responses to better align with the requirements of each facet. Finally, the queries and refined responses are used for supervised fine-tuning to reinforce the model's context-awareness ability. Evaluation results on the latest HealthBench dataset demonstrate that our method significantly improves LLM performance across multiple aspects, with particularly notable gains in the context-awareness axis. Furthermore, by incorporating knowledge distillation with the proposed method, the performance of a smaller backbone LLM (e.g., Qwen3-32B) surpasses its teacher model, achieving a new SOTA across all open-source LLMs on HealthBench (63.8%) and its hard subset (43.1%). Code and dataset will be released at https://muser-llm.github.io.

Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generating human-like content. This has revolutionized various sectors such as healthcare, software development, and education. In education, LLMs offer potential for personalized and interactive learning experiences, especially in regions with limited teaching resources. However, adapting these models effectively to curriculum-specific content, such as the National Council of Educational Research and Training (NCERT) syllabus in India, presents unique challenges in terms of accuracy, alignment, and pedagogical relevance. In this paper, we present the framework "PustakAI"\footnote{Pustak means `book' in many Indian languages.} for the design and evaluation of a novel question-answering dataset "NCERT-QA" aligned with the NCERT curriculum for English and Science subjects of grades 6 to 8. We classify the curated QA pairs as Factoid, Inferential, and Others (evaluative and reasoning). We evaluate the dataset with various prompting techniques, such as meta-prompt, few-shot, and CoT-style prompting, using diverse evaluation metrics to understand which approach aligns more efficiently with the structure and demands of the curriculum. Along with the usability of the dataset, we analyze the strengths and limitations of current open-source LLMs (Gemma3:1b, Llama3.2:3b, and Nemotron-mini:4b) and high-end LLMs (Llama-4-Scout-17B and Deepseek-r1-70B) as AI-based learning tools in formal education systems.

The opioid crisis represents a significant moment in public health that reveals systemic shortcomings across regulatory systems, healthcare practices, corporate governance, and public policy. Analyzing how these interconnected systems simultaneously failed to protect public health requires innovative analytic approaches for exploring the vast amounts of data and documents disclosed in the UCSF-JHU Opioid Industry Documents Archive (OIDA). The complexity, multimodal nature, and specialized characteristics of these healthcare-related legal and corporate documents necessitate more advanced methods and models tailored to specific data types and detailed annotations, ensuring the precision and professionalism in the analysis. In this paper, we tackle this challenge by organizing the original dataset according to document attributes and constructing a benchmark with 400k training documents and 10k for testing. From each document, we extract rich multimodal information-including textual content, visual elements, and layout structures-to capture a comprehensive range of features. Using multiple AI models, we then generate a large-scale dataset comprising 360k training QA pairs and 10k testing QA pairs. Building on this foundation, we develop domain-specific multimodal Large Language Models (LLMs) and explore the impact of multimodal inputs on task performance. To further enhance response accuracy, we incorporate historical QA pairs as contextual grounding for answering current queries. Additionally, we incorporate page references within the answers and introduce an importance-based page classifier, further improving the precision and relevance of the information provided. Preliminary results indicate the improvements with our AI assistant in document information extraction and question-answering tasks. The dataset is available at: https://huggingface.co/datasets/opioidarchive/oida-qa

Computational humor is a frontier for creating advanced and engaging natural language processing (NLP) applications, such as sophisticated dialogue systems. While previous studies have benchmarked the humor capabilities of Large Language Models (LLMs), they have often relied on single-dimensional evaluations, such as judging whether something is simply ``funny.'' This paper argues that a multifaceted understanding of humor is necessary and addresses this gap by systematically evaluating LLMs through the lens of Oogiri, a form of Japanese improvisational comedy games. To achieve this, we expanded upon existing Oogiri datasets with data from new sources and then augmented the collection with Oogiri responses generated by LLMs. We then manually annotated this expanded collection with 5-point absolute ratings across six dimensions: Novelty, Clarity, Relevance, Intelligence, Empathy, and Overall Funniness. Using this dataset, we assessed the capabilities of state-of-the-art LLMs on two core tasks: their ability to generate creative Oogiri responses and their ability to evaluate the funniness of responses using a six-dimensional evaluation. Our results show that while LLMs can generate responses at a level between low- and mid-tier human performance, they exhibit a notable lack of Empathy. This deficit in Empathy helps explain their failure to replicate human humor assessment. Correlation analyses of human and model evaluation data further reveal a fundamental divergence in evaluation criteria: LLMs prioritize Novelty, whereas humans prioritize Empathy. We release our annotated corpus to the community to pave the way for the development of more emotionally intelligent and sophisticated conversational agents.

While existing social bot detectors perform well on benchmarks, their robustness across diverse real-world scenarios remains limited due to unclear ground truth and varied misleading cues. In particular, the impact of shortcut learning, where models rely on spurious correlations instead of capturing causal task-relevant features, has received limited attention. To address this gap, we conduct an in-depth study to assess how detectors are influenced by potential shortcuts based on textual features, which are most susceptible to manipulation by social bots. We design a series of shortcut scenarios by constructing spurious associations between user labels and superficial textual cues to evaluate model robustness. Results show that shifts in irrelevant feature distributions significantly degrade social bot detector performance, with an average relative accuracy drop of 32\% in the baseline models. To tackle this challenge, we propose mitigation strategies based on large language models, leveraging counterfactual data augmentation. These methods mitigate the problem from data and model perspectives across three levels, including data distribution at both the individual user text and overall dataset levels, as well as the model's ability to extract causal information. Our strategies achieve an average relative performance improvement of 56\% under shortcut scenarios.

Visual grounding is the task of localising image regions from natural language queries and is critical for reasoning capable Graphical User Interface agents. Many existing methods rely on massive, noisy synthetic datasets. This work introduces an efficient training pipeline that combines model-based data filtering with parameter-efficient fine-tuning. From 4.8M synthetic examples, 12K clean and diverse instances are curated by first identifying challenging cases, removing misaligned and then selecting a diverse set of multimodal instances. On this data, a 3B-parameter Vision-Language Model is trained under three regimes: supervised fine-tuning, chain-of-thought-augmented fine-tuning, and reinforcement learning via Group Relative Policy Optimization. Models trained with the filtered data and lightweight training strategies match or surpass larger baselines on benchmarks such as ScreenSpot, Multimodal-Mind2Web, and AndroidControl. These results demonstrate that principled data curation and robust adaptation can rival large-scale training, enabling compact yet capable multimodal reasoning agents.

Efficient retrieval of external knowledge bases and web pages is crucial for enhancing the reasoning abilities of LLMs. Previous works on training LLMs to leverage external retrievers for solving complex problems have predominantly employed end-to-end reinforcement learning. However, these approaches neglect supervision over the reasoning process, making it difficult to guarantee logical coherence and rigor. To address these limitations, we propose Thinker, a hierarchical thinking model for deep search through multi-turn interaction, making the reasoning process supervisable and verifiable. It decomposes complex problems into independently solvable sub-problems, each dually represented in both natural language and an equivalent logical function to support knowledge base and web searches. Concurrently, dependencies between sub-problems are passed as parameters via these logical functions, enhancing the logical coherence of the problem-solving process. To avoid unnecessary external searches, we perform knowledge boundary determination to check if a sub-problem is within the LLM's intrinsic knowledge, allowing it to answer directly. Experimental results indicate that with as few as several hundred training samples, the performance of Thinker is competitive with established baselines. Furthermore, when scaled to the full training set, Thinker significantly outperforms these methods across various datasets and model sizes. The source code is available at https://github.com/OpenSPG/KAG-Thinker.

Large language model (LLM) queries are predominantly processed by frontier models in centralized cloud infrastructure. Rapidly growing demand strains this paradigm, and cloud providers struggle to scale infrastructure at pace. Two advances enable us to rethink this paradigm: small LMs (<=20B active parameters) now achieve competitive performance to frontier models on many tasks, and local accelerators (e.g., Apple M4 Max) run these models at interactive latencies. This raises the question: can local inference viably redistribute demand from centralized infrastructure? Answering this requires measuring whether local LMs can accurately answer real-world queries and whether they can do so efficiently enough to be practical on power-constrained devices (i.e., laptops). We propose intelligence per watt (IPW), task accuracy divided by unit of power, as a metric for assessing capability and efficiency of local inference across model-accelerator pairs. We conduct a large-scale empirical study across 20+ state-of-the-art local LMs, 8 accelerators, and a representative subset of LLM traffic: 1M real-world single-turn chat and reasoning queries. For each query, we measure accuracy, energy, latency, and power. Our analysis reveals $3$ findings. First, local LMs can accurately answer 88.7% of single-turn chat and reasoning queries with accuracy varying by domain. Second, from 2023-2025, IPW improved 5.3x and local query coverage rose from 23.2% to 71.3%. Third, local accelerators achieve at least 1.4x lower IPW than cloud accelerators running identical models, revealing significant headroom for optimization. These findings demonstrate that local inference can meaningfully redistribute demand from centralized infrastructure, with IPW serving as the critical metric for tracking this transition. We release our IPW profiling harness for systematic intelligence-per-watt benchmarking.