Despite the success of Large Multimodal Models (LMMs) in recent years, prompt design for LMMs in Multiple-Choice Question Answering (MCQA) remains poorly understood. We show that even minor variations in prompt phrasing and structure can lead to accuracy deviations of up to 15% for certain prompts and models. This variability poses a challenge for transparent and fair LMM evaluation, as models often report their best-case performance using carefully selected prompts. To address this, we introduce Promptception, a systematic framework for evaluating prompt sensitivity in LMMs. It consists of 61 prompt types, spanning 15 categories and 6 supercategories, each targeting specific aspects of prompt formulation, and is used to evaluate 10 LMMs ranging from lightweight open-source models to GPT-4o and Gemini 1.5 Pro, across 3 MCQA benchmarks: MMStar, MMMU-Pro, MVBench. Our findings reveal that proprietary models exhibit greater sensitivity to prompt phrasing, reflecting tighter alignment with instruction semantics, while open-source models are steadier but struggle with nuanced and complex phrasing. Based on this analysis, we propose Prompting Principles tailored to proprietary and open-source LMMs, enabling more robust and fair model evaluation.

In deployment and application, large language models (LLMs) typically undergo safety alignment to prevent illegal and unethical outputs. However, the continuous advancement of jailbreak attack techniques, designed to bypass safety mechanisms with adversarial prompts, has placed increasing pressure on the security defenses of LLMs. Strengthening resistance to jailbreak attacks requires an in-depth understanding of the security mechanisms and vulnerabilities of LLMs. However, the vast number of parameters and complex structure of LLMs make analyzing security weaknesses from an internal perspective a challenging task. This paper presents NeuroBreak, a top-down jailbreak analysis system designed to analyze neuron-level safety mechanisms and mitigate vulnerabilities. We carefully design system requirements through collaboration with three experts in the field of AI security. The system provides a comprehensive analysis of various jailbreak attack methods. By incorporating layer-wise representation probing analysis, NeuroBreak offers a novel perspective on the model's decision-making process throughout its generation steps. Furthermore, the system supports the analysis of critical neurons from both semantic and functional perspectives, facilitating a deeper exploration of security mechanisms. We conduct quantitative evaluations and case studies to verify the effectiveness of our system, offering mechanistic insights for developing next-generation defense strategies against evolving jailbreak attacks.

The effectiveness of large language models (LLMs) to fact-check misinformation remains uncertain, despite their growing use. To this end, we present CANDY, a benchmark designed to systematically evaluate the capabilities and limitations of LLMs in fact-checking Chinese misinformation. Specifically, we curate a carefully annotated dataset of ~20k instances. Our analysis shows that current LLMs exhibit limitations in generating accurate fact-checking conclusions, even when enhanced with chain-of-thought reasoning and few-shot prompting. To understand these limitations, we develop a taxonomy to categorize flawed LLM-generated explanations for their conclusions and identify factual fabrication as the most common failure mode. Although LLMs alone are unreliable for fact-checking, our findings indicate their considerable potential to augment human performance when deployed as assistive tools in scenarios. Our dataset and code can be accessed at https://github.com/SCUNLP/CANDY

Recent significant advancements in Large Language Models (LLMs) have greatly propelled the development of Role-Playing Conversational Agents (RPCAs). These systems aim to create immersive user experiences through consistent persona adoption. However, current RPCA research faces dual limitations. First, existing work predominantly focuses on the textual modality, entirely overlooking critical paralinguistic features including intonation, prosody, and rhythm in speech, which are essential for conveying character emotions and shaping vivid identities. Second, the speech-based role-playing domain suffers from a long-standing lack of standardized evaluation benchmarks. Most current spoken dialogue datasets target only fundamental capability assessments, featuring thinly sketched or ill-defined character profiles. Consequently, they fail to effectively quantify model performance on core competencies like long-term persona consistency. To address this critical gap, we introduce VoxRole, the first comprehensive benchmark specifically designed for the evaluation of speech-based RPCAs. The benchmark comprises 13335 multi-turn dialogues, totaling 65.6 hours of speech from 1228 unique characters across 261 movies. To construct this resource, we propose a novel two-stage automated pipeline that first aligns movie audio with scripts and subsequently employs an LLM to systematically build multi-dimensional profiles for each character. Leveraging VoxRole, we conduct a multi-dimensional evaluation of contemporary spoken dialogue models, revealing crucial insights into their respective strengths and limitations in maintaining persona consistency.

Recent advancements in large language models (LLMs) have revolutionized natural language processing through their remarkable capabilities in understanding and executing diverse tasks. While supervised fine-tuning, particularly in Retrieval-Augmented Generation (RAG) scenarios, effectively enhances task-specific performance, it often leads to catastrophic forgetting, where models lose their previously acquired knowledge and general capabilities. Existing solutions either require access to general instruction data or face limitations in preserving the model's original distribution. To overcome these limitations, we propose SelfAug, a self-distribution alignment method that aligns input sequence logits to preserve the model's semantic distribution, thereby mitigating catastrophic forgetting and improving downstream performance. Extensive experiments demonstrate that SelfAug achieves a superior balance between downstream learning and general capability retention. Our comprehensive empirical analysis reveals a direct correlation between distribution shifts and the severity of catastrophic forgetting in RAG scenarios, highlighting how the absence of RAG capabilities in general instruction tuning leads to significant distribution shifts during fine-tuning. Our findings not only advance the understanding of catastrophic forgetting in RAG contexts but also provide a practical solution applicable across diverse fine-tuning scenarios. Our code is publicly available at https://github.com/USTC-StarTeam/SelfAug.

Smartphones have become indispensable in people's daily lives, permeating nearly every aspect of modern society. With the continuous advancement of large language models (LLMs), numerous LLM-based mobile agents have emerged. These agents are capable of accurately parsing diverse user queries and automatically assisting users in completing complex or repetitive operations. However, current agents 1) heavily rely on the comprehension ability of LLMs, which can lead to errors caused by misoperations or omitted steps during tasks, 2) lack interaction with the external environment, often terminating tasks when an app cannot fulfill user queries, and 3) lack memory capabilities, requiring each instruction to reconstruct the interface and being unable to learn from and correct previous mistakes. To alleviate the above issues, we propose MobileRAG, a mobile agents framework enhanced by Retrieval-Augmented Generation (RAG), which includes InterRAG, LocalRAG, and MemRAG. It leverages RAG to more quickly and accurately identify user queries and accomplish complex and long-sequence mobile tasks. Additionally, to more comprehensively assess the performance of MobileRAG, we introduce MobileRAG-Eval, a more challenging benchmark characterized by numerous complex, real-world mobile tasks that require external knowledge assistance. Extensive experimental results on MobileRAG-Eval demonstrate that MobileRAG can easily handle real-world mobile tasks, achieving 10.3% improvement over state-of-the-art methods with fewer operational steps. Our code is publicly available at: https://github.

The development of Long-CoT reasoning has advanced LLM performance across various tasks, including language understanding, complex problem solving, and code generation. This paradigm enables models to generate intermediate reasoning steps, thereby improving both accuracy and interpretability. However, despite these advancements, a comprehensive understanding of how CoT-based reasoning affects the trustworthiness of language models remains underdeveloped. In this paper, we survey recent work on reasoning models and CoT techniques, focusing on five core dimensions of trustworthy reasoning: truthfulness, safety, robustness, fairness, and privacy. For each aspect, we provide a clear and structured overview of recent studies in chronological order, along with detailed analyses of their methodologies, findings, and limitations. Future research directions are also appended at the end for reference and discussion. Overall, while reasoning techniques hold promise for enhancing model trustworthiness through hallucination mitigation, harmful content detection, and robustness improvement, cutting-edge reasoning models themselves often suffer from comparable or even greater vulnerabilities in safety, robustness, and privacy. By synthesizing these insights, we hope this work serves as a valuable and timely resource for the AI safety community to stay informed on the latest progress in reasoning trustworthiness. A full list of related papers can be found at \href{https://github.com/ybwang119/Awesome-reasoning-safety}{https://github.com/ybwang119/Awesome-reasoning-safety}.

Accurate prediction of communication link quality metrics is essential for vehicle-to-infrastructure (V2I) systems, enabling smooth handovers, efficient beam management, and reliable low-latency communication. The increasing availability of sensor data from modern vehicles motivates the use of multimodal large language models (MLLMs) because of their adaptability across tasks and reasoning capabilities. However, MLLMs inherently lack three-dimensional spatial understanding. To overcome this limitation, a lightweight, plug-and-play bird's-eye view (BEV) injection connector is proposed. In this framework, a BEV of the environment is constructed by collecting sensing data from neighboring vehicles. This BEV representation is then fused with the ego vehicle's input to provide spatial context for the large language model. To support realistic multimodal learning, a co-simulation environment combining CARLA simulator and MATLAB-based ray tracing is developed to generate RGB, LiDAR, GPS, and wireless signal data across varied scenarios. Instructions and ground-truth responses are programmatically extracted from the ray-tracing outputs. Extensive experiments are conducted across three V2I link prediction tasks: line-of-sight (LoS) versus non-line-of-sight (NLoS) classification, link availability, and blockage prediction. Simulation results show that the proposed BEV injection framework consistently improved performance across all tasks. The results indicate that, compared to an ego-only baseline, the proposed approach improves the macro-average of the accuracy metrics by up to 13.9%. The results also show that this performance gain increases by up to 32.7% under challenging rainy and nighttime conditions, confirming the robustness of the framework in adverse settings.

The Observational Medical Outcomes Partnership (OMOP) common data model (CDM) provides a standardized representation of heterogeneous health data to support large-scale, multi-institutional research. One critical step in data standardization using OMOP CDM is the mapping of source medical terms to OMOP standard concepts, a procedure that is resource-intensive and error-prone. While large language models (LLMs) have the potential to facilitate this process, their tendency toward hallucination makes them unsuitable for clinical deployment without training and expert validation. Here, we developed a zero-training, hallucination-preventive mapping system based on the Model Context Protocol (MCP), a standardized and secure framework allowing LLMs to interact with external resources and tools. The system enables explainable mapping and significantly improves efficiency and accuracy with minimal effort. It provides real-time vocabulary lookups and structured reasoning outputs suitable for immediate use in both exploratory and production environments.

Large language models (LLMs) have ushered in a new era for document-level machine translation (\textit{doc}-mt), yet their whole-document outputs challenge existing evaluation methods that assume sentence-by-sentence alignment. We introduce \textit{\textbf{Align-then-Slide}}, a complete evaluation framework for ultra-long doc-mt. In the Align stage, we automatically infer sentence-level source-target correspondences and rebuild the target to match the source sentence number, resolving omissions and many-to-one/one-to-many mappings. In the n-Chunk Sliding Evaluate stage, we calculate averaged metric scores under 1-, 2-, 3- and 4-chunk for multi-granularity assessment. Experiments on the WMT benchmark show a Pearson correlation of 0.929 between our method with expert MQM rankings. On a newly curated real-world test set, our method again aligns closely with human judgments. Furthermore, preference data produced by Align-then-Slide enables effective CPO training and its direct use as a reward model for GRPO, both yielding translations preferred over a vanilla SFT baseline. The results validate our framework as an accurate, robust, and actionable evaluation tool for doc-mt systems.