Reinforcement learning (RL) agents can learn to solve complex tasks from visual inputs, but generalizing these learned skills to new environments remains a major challenge in RL application, especially robotics. While data augmentation can improve generalization, it often compromises sample efficiency and training stability. This paper introduces DeGuV, an RL framework that enhances both generalization and sample efficiency. In specific, we leverage a learnable masker network that produces a mask from the depth input, preserving only critical visual information while discarding irrelevant pixels. Through this, we ensure that our RL agents focus on essential features, improving robustness under data augmentation. In addition, we incorporate contrastive learning and stabilize Q-value estimation under augmentation to further enhance sample efficiency and training stability. We evaluate our proposed method on the RL-ViGen benchmark using the Franka Emika robot and demonstrate its effectiveness in zero-shot sim-to-real transfer. Our results show that DeGuV outperforms state-of-the-art methods in both generalization and sample efficiency while also improving interpretability by highlighting the most relevant regions in the visual input

The development of autonomous agents for graphical user interfaces (GUIs) presents major challenges in artificial intelligence. While recent advances in native agent models have shown promise by unifying perception, reasoning, action, and memory through end-to-end learning, open problems remain in data scalability, multi-turn reinforcement learning (RL), the limitations of GUI-only operation, and environment stability. In this technical report, we present UI-TARS-2, a native GUI-centered agent model that addresses these challenges through a systematic training methodology: a data flywheel for scalable data generation, a stabilized multi-turn RL framework, a hybrid GUI environment that integrates file systems and terminals, and a unified sandbox platform for large-scale rollouts. Empirical evaluation demonstrates that UI-TARS-2 achieves significant improvements over its predecessor UI-TARS-1.5. On GUI benchmarks, it reaches 88.2 on Online-Mind2Web, 47.5 on OSWorld, 50.6 on WindowsAgentArena, and 73.3 on AndroidWorld, outperforming strong baselines such as Claude and OpenAI agents. In game environments, it attains a mean normalized score of 59.8 across a 15-game suite-roughly 60% of human-level performance-and remains competitive with frontier proprietary models (e.g., OpenAI o3) on LMGame-Bench. Additionally, the model can generalize to long-horizon information-seeking tasks and software engineering benchmarks, highlighting its robustness across diverse agent tasks. Detailed analyses of training dynamics further provide insights into achieving stability and efficiency in large-scale agent RL. These results underscore UI-TARS-2's potential to advance the state of GUI agents and exhibit strong generalization to real-world interactive scenarios.

Large language models (LLMs) have advanced virtual educators and learners, bridging NLP with AI4Education. Existing work often lacks scalability and fails to leverage diverse, large-scale course content, with limited frameworks for assessing pedagogic quality. To this end, we propose WikiHowAgent, a multi-agent workflow leveraging LLMs to simulate interactive teaching-learning conversations. It integrates teacher and learner agents, an interaction manager, and an evaluator to facilitate procedural learning and assess pedagogic quality. We introduce a dataset of 114,296 teacher-learner conversations grounded in 14,287 tutorials across 17 domains and 727 topics. Our evaluation protocol combines computational and rubric-based metrics with human judgment alignment. Results demonstrate the workflow's effectiveness in diverse setups, offering insights into LLM capabilities across domains. Our datasets and implementations are fully open-sourced.

There is growing interest in understanding how people interact with large language models (LLMs) and whether such models elicit dependency or even addictive behaviour. Validated tools to assess the extent to which individuals may become dependent on LLMs are scarce and primarily build on classic behavioral addiction symptoms, adapted to the context of LLM use. We view this as a conceptual limitation, as the LLM-human relationship is more nuanced and warrants a fresh and distinct perspective. To address this gap, we developed and validated a new 12-item questionnaire to measure LLM dependency, referred to as LLM-D12. The scale was based on the authors' prior theoretical work, with items developed accordingly and responses collected from 526 participants in the UK. Exploratory and confirmatory factor analyses, performed on separate halves of the total sample using a split-sample approach, supported a two-factor structure: Instrumental Dependency (six items) and Relationship Dependency (six items). Instrumental Dependency reflects the extent to which individuals rely on LLMs to support or collaborate in decision-making and cognitive tasks. Relationship Dependency captures the tendency to perceive LLMs as socially meaningful, sentient, or companion-like entities. The two-factor structure demonstrated excellent internal consistency and clear discriminant validity. External validation confirmed both the conceptual foundation and the distinction between the two subscales. The psychometric properties and structure of our LLM-D12 scale were interpreted in light of the emerging view that dependency on LLMs does not necessarily indicate dysfunction but may still reflect reliance levels that could become problematic in certain contexts.

Despite the great potential of large language models(LLMs) in machine comprehension, it is still disturbing to fully count on them in real-world scenarios. This is probably because there is no rational explanation for whether the comprehension process of LLMs is aligned with that of experts. In this paper, we propose SCOP to carefully examine how LLMs perform during the comprehension process from a cognitive view. Specifically, it is equipped with a systematical definition of five requisite skills during the comprehension process, a strict framework to construct testing data for these skills, and a detailed analysis of advanced open-sourced and closed-sourced LLMs using the testing data. With SCOP, we find that it is still challenging for LLMs to perform an expert-level comprehension process. Even so, we notice that LLMs share some similarities with experts, e.g., performing better at comprehending local information than global information. Further analysis reveals that LLMs can be somewhat unreliable -- they might reach correct answers through flawed comprehension processes. Based on SCOP, we suggest that one direction for improving LLMs is to focus more on the comprehension process, ensuring all comprehension skills are thoroughly developed during training.

The limited ability to reason across occupational data from different sources is a long-standing bottleneck for data-driven labour market analytics. Previous research has relied on hand-crafted ontologies that allow such reasoning but are computationally expensive and require careful maintenance by human experts. The rise of language processing machine learning models offers a scalable alternative by learning shared semantic spaces that bridge diverse occupational vocabularies without extensive human curation. We present an embedding-based alignment process that links any free-form German job title to two established ontologies - the German Klassifikation der Berufe and the International Standard Classification of Education. Using publicly available data from the German Federal Employment Agency, we construct a dataset to fine-tune a Sentence-BERT model to learn the structure imposed by the ontologies. The enriched pairs (job title, embedding) define a similarity graph structure that we can use for efficient approximate nearest-neighbour search, allowing us to frame the classification process as a semantic search problem. This allows for greater flexibility, e.g., adding more classes. We discuss design decisions, open challenges, and outline ongoing work on extending the graph with other ontologies and multilingual titles.

Robust evaluation in the presence of linguistic variation is key to understanding the generalization capabilities of Natural Language to SQL (NL2SQL) models, yet existing benchmarks rarely address this factor in a systematic or controlled manner. We propose a novel schema-aligned paraphrasing framework that leverages SQL-to-NL (SQL2NL) to automatically generate semantically equivalent, lexically diverse queries while maintaining alignment with the original schema and intent. This enables the first targeted evaluation of NL2SQL robustness to linguistic variation in isolation-distinct from prior work that primarily investigates ambiguity or schema perturbations. Our analysis reveals that state-of-the-art models are far more brittle than standard benchmarks suggest. For example, LLaMa3.3-70B exhibits a 10.23% drop in execution accuracy (from 77.11% to 66.9%) on paraphrased Spider queries, while LLaMa3.1-8B suffers an even larger drop of nearly 20% (from 62.9% to 42.5%). Smaller models (e.g., GPT-4o mini) are disproportionately affected. We also find that robustness degradation varies significantly with query complexity, dataset, and domain -- highlighting the need for evaluation frameworks that explicitly measure linguistic generalization to ensure reliable performance in real-world settings.

Large Language Models (LLMs) effectiveness is usually evaluated by means of benchmarks such as MMLU, ARC-C, or Hel-laSwag, where questions are presented in their original wording, thus in a fixed, standardized format. However, real-world applications involve linguistic variability, requiring models to maintain their effectiveness across diverse rewordings of the same question or query. In this study, we systematically assess the robustness of LLMs to paraphrased benchmark questions and investigate whether benchmark-based evaluations provide a reliable measure of model capabilities. We systematically generate various paraphrases of all the questions across six different common benchmarks, and measure the resulting variations in effectiveness of 34 state-of-the-art LLMs, of different size and effectiveness. Our findings reveal that while LLM rankings remain relatively stable across paraphrased inputs, absolute effectiveness scores change, and decline significantly. This suggests that LLMs struggle with linguistic variability, raising concerns about their generalization abilities and evaluation methodologies. Furthermore, the observed performance drop challenges the reliability of benchmark-based evaluations, indicating that high benchmark scores may not fully capture a model's robustness to real-world input variations. We discuss the implications of these findings for LLM evaluation methodologies, emphasizing the need for robustness-aware benchmarks that better reflect practical deployment scenarios.

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.

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}.