Large Language Models (LLMs) should answer factual questions truthfully, grounded in objective knowledge, regardless of user context such as self-disclosed personal information, or system personalization. In this paper, we present the first systematic evaluation of LLM robustness to inquiry personas, i.e. user profiles that convey attributes like identity, expertise, or belief. While prior work has primarily focused on adversarial inputs or distractors for robustness testing, we evaluate plausible, human-centered inquiry persona cues that users disclose in real-world interactions. We find that such cues can meaningfully alter QA accuracy and trigger failure modes such as refusals, hallucinated limitations, and role confusion. These effects highlight how model sensitivity to user framing can compromise factual reliability, and position inquiry persona testing as an effective tool for robustness evaluation.

Analyzing large-scale scientific datasets presents substantial challenges due to their sheer volume, structural complexity, and the need for specialized domain knowledge. Automation tools, such as PandasAI, typically require full data ingestion and lack context of the full data structure, making them impractical as intelligent data analysis assistants for datasets at the terabyte scale. To overcome these limitations, we propose InferA, a multi-agent system that leverages large language models to enable scalable and efficient scientific data analysis. At the core of the architecture is a supervisor agent that orchestrates a team of specialized agents responsible for distinct phases of the data retrieval and analysis. The system engages interactively with users to elicit their analytical intent and confirm query objectives, ensuring alignment between user goals and system actions. To demonstrate the framework's usability, we evaluate the system using ensemble runs from the HACC cosmology simulation which comprises several terabytes.

Critical thinking represents a fundamental competency in today's education landscape. Developing critical thinking skills through timely assessment and feedback is crucial; however, there has not been extensive work in the learning analytics community on defining, measuring, and supporting critical thinking. In this paper, we investigate the feasibility of measuring core "subskills" that underlie critical thinking. We ground our work in an authentic task where students operationalize critical thinking: student-written argumentative essays. We developed a coding rubric based on an established skills progression and completed human coding for a corpus of student essays. We then evaluated three distinct approaches to automated scoring: zero-shot prompting, few-shot prompting, and supervised fine-tuning, implemented across three large language models (GPT-5, GPT-5-mini, and ModernBERT). GPT-5 with few-shot prompting achieved the strongest results and demonstrated particular strength on subskills with separable, frequent categories, while lower performance was observed for subskills that required detection of subtle distinctions or rare categories. Our results underscore critical trade-offs in automated critical thinking assessment: proprietary models offer superior reliability at higher cost, while open-source alternatives provide practical accuracy with reduced sensitivity to minority categories. Our work represents an initial step toward scalable assessment of higher-order reasoning skills across authentic educational contexts.

Recently, several multi-turn dialogue benchmarks have been proposed to evaluate the conversational abilities of large language models (LLMs). As LLMs are increasingly recognized as a key technology for advancing intelligent education, owing to their ability to deeply understand instructional contexts and provide personalized guidance, the construction of dedicated teacher-student dialogue benchmarks has become particularly important. To this end, we present EduDial, a comprehensive multi-turn teacher-student dialogue dataset. EduDial covers 345 core knowledge points and consists of 34,250 dialogue sessions generated through interactions between teacher and student agents. Its design is guided by Bloom's taxonomy of educational objectives and incorporates ten questioning strategies, including situational questioning, zone of proximal development (ZPD) questioning, and metacognitive questioning-thus better capturing authentic classroom interactions. Furthermore, we design differentiated teaching strategies for students at different cognitive levels, thereby providing more targeted teaching guidance. Building on EduDial, we further develop EduDial-LLM 32B via training and propose an 11-dimensional evaluation framework that systematically measures the teaching abilities of LLMs, encompassing both overall teaching quality and content quality. Experiments on 17 mainstream LLMs reveal that most models struggle in student-centered teaching scenarios, whereas our EduDial-LLM achieves significant gains, consistently outperforming all baselines across all metrics. The code is available at https://github.com/Mind-Lab-ECNU/EduDial/tree/main.

Robotic manipulation policies often struggle to generalize to novel objects, limiting their real-world utility. In contrast, cognitive science suggests that children develop generalizable dexterous manipulation skills by mastering a small set of simple toys and then applying that knowledge to more complex items. Inspired by this, we study if similar generalization capabilities can also be achieved by robots. Our results indicate robots can learn generalizable grasping using randomly assembled objects that are composed from just four shape primitives: spheres, cuboids, cylinders, and rings. We show that training on these "toys" enables robust generalization to real-world objects, yielding strong zero-shot performance. Crucially, we find the key to this generalization is an object-centric visual representation induced by our proposed detection pooling mechanism. Evaluated in both simulation and on physical robots, our model achieves a 67% real-world grasping success rate on the YCB dataset, outperforming state-of-the-art approaches that rely on substantially more in-domain data. We further study how zero-shot generalization performance scales by varying the number and diversity of training toys and the demonstrations per toy. We believe this work offers a promising path to scalable and generalizable learning in robotic manipulation. Demonstration videos, code, checkpoints and our dataset are available on our project page: https://lego-grasp.github.io/ .

Large Language Models (LLMs) are increasingly being deployed as the reasoning engines for agentic AI systems, yet they exhibit a critical flaw: a rigid adherence to explicit rules that leads to decisions misaligned with human common sense and intent. This "rule-rigidity" is a significant barrier to building trustworthy autonomous agents. While prior work has shown that supervised fine-tuning (SFT) with human explanations can mitigate this issue, SFT is computationally expensive and inaccessible to many practitioners. To address this gap, we introduce the Rule-Intent Distinction (RID) Framework, a novel, low-compute meta-prompting technique designed to elicit human-aligned exception handling in LLMs in a zero-shot manner. The RID framework provides the model with a structured cognitive schema for deconstructing tasks, classifying rules, weighing conflicting outcomes, and justifying its final decision. We evaluated the RID framework against baseline and Chain-of-Thought (CoT) prompting on a custom benchmark of 20 scenarios requiring nuanced judgment across diverse domains. Our human-verified results demonstrate that the RID framework significantly improves performance, achieving a 95% Human Alignment Score (HAS), compared to 80% for the baseline and 75% for CoT. Furthermore, it consistently produces higher-quality, intent-driven reasoning. This work presents a practical, accessible, and effective method for steering LLMs from literal instruction-following to liberal, goal-oriented reasoning, paving the way for more reliable and pragmatic AI agents.

Large language models (LLMs) are now widely deployed in user-facing applications, reaching hundreds of millions worldwide. As they become integrated into everyday tasks, growing reliance on their outputs raises significant concerns. In particular, users may unknowingly be exposed to model-inherent biases that systematically disadvantage or stereotype certain groups. However, existing bias benchmarks continue to rely on templated prompts or restrictive multiple-choice questions that are suggestive, simplistic, and fail to capture the complexity of real-world user interactions. In this work, we address this gap by introducing a counterfactual bias evaluation framework that automatically generates realistic, open-ended questions over sensitive attributes such as sex, race, or religion. By iteratively mutating and selecting bias-inducing questions, our approach systematically explores areas where models are most susceptible to biased behavior. Beyond detecting harmful biases, we also capture distinct response dimensions that are increasingly relevant in user interactions, such as asymmetric refusals and explicit acknowledgment of bias. Leveraging our framework, we construct CAB, a human-verified benchmark spanning diverse topics, designed to enable cross-model comparisons. Using CAB, we analyze a range of LLMs across multiple bias dimensions, revealing nuanced insights into how different models manifest bias. For instance, while GPT-5 outperforms other models, it nonetheless exhibits persistent biases in specific scenarios. These findings underscore the need for continual improvements to ensure fair model behavior.

Large Audio-Language Models and Multi-Modal Large Language Models have demonstrated strong capabilities in tasks such as Audio Question Answering (AQA), Audio Captioning, and Automatic Speech Recognition (ASR). However, there is growing evidence that these models can hallucinate about the content of the audio. To address this issue, we probe the models' internal states and propose Adaptive Vector Steering (AVS), a method that better grounds generation in audio content. We also identify a strong correlation between output correctness and internal representations. Experiments show consistent performance gains across two models and two benchmarks. On the Audio Hallucination QA dataset, our method boosts the F1-score of Gemma from 0.550 to 0.619 and Qwen from 0.626 to 0.632. Furthermore, our method increases the accuracy of Qwen on MMAU from 0.548 to 0.592, marking an 8% relative increase. To the best of our knowledge, this is the first work to apply vector steering to mitigate hallucination in audio.

Pseudo-Alignment is a pervasive challenge in many large language models for time series (LLM4TS) models, often causing them to underperform compared to linear models or randomly initialised backbones. However, there is limited discussion in the community for the reasons that pseudo-alignment occurs. In this work, we conduct a thorough investigation into the root causes of pseudo-alignment in LLM4TS and build a connection of pseudo-alignment to the cone effect in LLM. We demonstrate that pseudo-alignment arises from the interplay of cone effect within pretrained LLM components and the intrinsically low-dimensional manifold of time-series data. In addition, we also introduce \textit{\textbf{TimeSUP}}, a novel technique designed to mitigate this issue and improve forecast performance in existing LLM4TS approaches. TimeSUP addresses this by increasing the time series manifold to more closely match the intrinsic dimension of language embeddings, allowing the model to distinguish temporal signals clearly while still capturing shared structures across modalities. As a result, representations for time and language tokens remain distinct yet exhibit high cosine similarity, signifying that the model preserves each modality unique features while learning their commonalities in a unified embedding space. Empirically, TimeSUP consistently outperforms state-of-the-art LLM4TS methods and other lightweight baselines on long-term forecasting performance. Furthermore, it can be seamlessly integrated into four existing LLM4TS pipelines and delivers significant improvements in forecasting performance.

Vision Language Models (VLMs) are pivotal for advancing perception in intelligent agents. Yet, evaluation of VLMs remains limited to predominantly English-centric benchmarks in which the image-text pairs comprise short texts. To evaluate VLM fine-grained abilities, in four languages under long-text settings, we introduce a novel multilingual benchmark VLURes featuring eight vision-and-language tasks, and a pioneering unrelatedness task, to probe the fine-grained Visual and Linguistic Understanding capabilities of VLMs across English, Japanese, and low-resource languages, Swahili, and Urdu. Our datasets, curated from web resources in the target language, encompass ten diverse image categories and rich textual context, introducing valuable vision-language resources for Swahili and Urdu. By prompting VLMs to generate responses and rationales, evaluated automatically and by native speakers, we uncover performance disparities across languages and tasks critical to intelligent agents, such as object recognition, scene understanding, and relationship understanding. We conducted evaluations of ten VLMs with VLURes. The best performing model, GPT-4o, achieves an overall accuracy of 90.8% and lags human performance by 6.7%, though the gap is larger for open-source models. The gap highlights VLURes' critical role in developing intelligent agents to tackle multi-modal visual reasoning.