Graphical User Interface (GUI) agents offer cross-platform solutions for automating complex digital tasks, with significant potential to transform productivity workflows. However, their performance is often constrained by the scarcity of high-quality trajectory data. To address this limitation, we propose training Vision Language Models (VLMs) on data-rich, reasoning-intensive tasks during a dedicated mid-training stage, and then examine how incorporating these tasks facilitates generalization to GUI planning scenarios. Specifically, we explore a range of tasks with readily available instruction-tuning data, including GUI perception, multimodal reasoning, and textual reasoning. Through extensive experiments across 11 mid-training tasks, we demonstrate that: (1) Task generalization proves highly effective, yielding substantial improvements across most settings. For instance, multimodal mathematical reasoning enhances performance on AndroidWorld by an absolute 6.3%. Remarkably, text-only mathematical data significantly boosts GUI web agent performance, achieving a 5.6% improvement on WebArena and 5.4% improvement on AndroidWorld, underscoring notable cross-modal generalization from text-based to visual domains; (2) Contrary to prior assumptions, GUI perception data - previously considered closely aligned with GUI agent tasks and widely utilized for training - has a comparatively limited impact on final performance; (3) Building on these insights, we identify the most effective mid-training tasks and curate optimized mixture datasets, resulting in absolute performance gains of 8.0% on WebArena and 12.2% on AndroidWorld. Our work provides valuable insights into cross-domain knowledge transfer for GUI agents and offers a practical approach to addressing data scarcity challenges in this emerging field. The code, data and models will be available at https://github.com/hkust-nlp/GUIMid.

One of the major challenges in machine learning is maintaining the accuracy of the deployed model (e.g., a classifier) in a non-stationary environment. The non-stationary environment results in distribution shifts and, consequently, a degradation in accuracy. Continuous learning of the deployed model with new data could be one remedy. However, the question arises as to how we should update the model with new training data so that it retains its accuracy on the old data while adapting to the new data. In this work, we propose a task-conditioned ensemble of models to maintain the performance of the existing model. The method involves an ensemble of expert models based on task membership information. The in-domain models-based on the local outlier concept (different from the expert models) provide task membership information dynamically at run-time to each probe sample. To evaluate the proposed method, we experiment with three setups: the first represents distribution shift between tasks (LivDet-Iris-2017), the second represents distribution shift both between and within tasks (LivDet-Iris-2020), and the third represents disjoint distribution between tasks (Split MNIST). The experiments highlight the benefits of the proposed method. The source code is available at https://github.com/iPRoBe-lab/Continuous_Learning_FE_DM.

Large reasoning models exhibit remarkable reasoning capabilities via long, elaborate reasoning trajectories. Supervised fine-tuning on such reasoning traces, also known as distillation, can be a cost-effective way to boost reasoning capabilities of student models. However, empirical observations reveal that these reasoning trajectories are often suboptimal, switching excessively between different lines of thought, resulting in under-thinking, over-thinking, and even degenerate responses. We introduce Retro-Search, an MCTS-inspired search algorithm, for distilling higher quality reasoning paths from large reasoning models. Retro-Search retrospectively revises reasoning paths to discover better, yet shorter traces, which can then lead to student models with enhanced reasoning capabilities with shorter, thus faster inference. Our approach can enable two use cases: self-improvement, where models are fine-tuned on their own Retro-Search-ed thought traces, and weak-to-strong improvement, where a weaker model revises stronger model's thought traces via Retro-Search. For self-improving, R1-distill-7B, fine-tuned on its own Retro-Search-ed traces, reduces the average reasoning length by 31.2% while improving performance by 7.7% across seven math benchmarks. For weak-to-strong improvement, we retrospectively revise R1-671B's traces from the OpenThoughts dataset using R1-distill-32B as the Retro-Search-er, a model 20x smaller. Qwen2.5-32B, fine-tuned on this refined data, achieves performance comparable to R1-distill-32B, yielding an 11.3% reduction in reasoning length and a 2.4% performance improvement compared to fine-tuning on the original OpenThoughts data. Our work counters recently emergent viewpoints that question the relevance of search algorithms in the era of large reasoning models, by demonstrating that there are still opportunities for algorithmic advancements, even for frontier models.

By the time you read this article, there's a good chance it will have already been scanned by an artificially intelligent machine. If asked about the artist David Salle, large language models such as ChatGPT or Gemini may repurpose some of the words below to come up with their answer. The bigger the data set, the more convincing the response – and Salle has been written about exhaustively since he first rose to art world stardom in the 1980s. The question is whether AI can ever say anything new about the artist and his work, or if it's for ever condemned to generate more of the same. A similar question lingers beneath the surface of the paintings that Salle has been making since 2023, a new series of which he has just unveiled at Thaddaeus Ropac in London.

In today's globalized world, there are increasing opportunities for individuals to communicate using a common non-native language (lingua franca). Non-native speakers often have opportunities to listen to foreign languages, but may not comprehend them as fully as native speakers do. To aid real-time comprehension, live transcription of subtitles is frequently used in everyday life (e.g., during Zoom conversations, watching YouTube videos, or on social networking sites). However, simultaneously reading subtitles while listening can increase cognitive load. In this study, we propose Dynamik, a system that reduces cognitive load during reading by decreasing the size of less important words and enlarging important ones, thereby enhancing sentence contrast. Our results indicate that Dynamik can reduce certain aspects of cognitive load, specifically, participants' perceived performance and effort among individuals with low proficiency in English, as well as enhance the users' sense of comprehension, especially among people with low English ability. We further discuss our methods' applicability to other languages and potential improvements and further research directions.

We investigate reinforcement learning (RL) algorithms in the context of fine-tuning large language models (LLMs) with verifiable rewards. Our focus is on mathematical reasoning tasks, which have recently received significant attention following the release of models such as OpenAI's O1 Model (Jaech et al., 2024) and DeepSeek-R1 (DeepSeek-AI et al., 2025). The dominant approach in LLM post-training has been Proximal Policy Optimization (PPO) (Schulman et al., 2017; Bai et al., 2022; Ouyang et al., 2022). However, PPO requires an additional critic network beyond the vanilla Reinforce algorithm (Williams and Peng, 1991), introducing both computational overhead and algorithmic complexity. Meanwhile, the deterministic transition nature of LLM also simplifies the problem with a relatively lower variance, many of PPO's sophisticated components may be unnecessary in this setting. This observation has inspired growing interest in designing simpler yet effective RL algorithms for post-training LLMs. Several recent works revisit Reinforce-style approaches, including ReMax (Li et al., 2023), RLOO (Ahma-dian et al., 2024; Kool et al., 2019), GRPO (Shao et al., 2024), and Reinforce++ (Hu, 2025). In parallel, other methods explore different directions beyond policy gradients. Reward-ranked fine-tuning (RAFT) (Anthony et al., 2017; Dong et al., 2023) iteratively generates n responses per prompt, filter out those with incorrect answers, and fine-tune the LLM on the remaining accepted samples.

Bipartite ranking is a fundamental supervised learning problem, with the goal of learning a ranking over instances with maximal area under the ROC curve (AUC) against a single binary target label. However, one may often observe multiple binary target labels, e.g., from distinct human annotators. How can one synthesize such labels into a single coherent ranking? In this work, we formally analyze two approaches to this problem -- loss aggregation and label aggregation -- by characterizing their Bayes-optimal solutions. Based on this, we show that while both methods can yield Pareto-optimal solutions, loss aggregation can exhibit label dictatorship: one can inadvertently (and undesirably) favor one label over others. This suggests that label aggregation can be preferable to loss aggregation, which we empirically verify.

Extracting key concepts and their relationships from course information and materials facilitates the provision of visualizations and recommendations for learners who need to select the right courses to take from a large number of courses. However, identifying and extracting themes manually is labor-intensive and time-consuming. Previous machine learning-based methods to extract relevant concepts from courses heavily rely on detailed course materials, which necessitates labor-intensive preparation of course materials. This paper investigates the potential of LLMs such as GPT in automatically generating course concepts and their relations. Specifically, we design a suite of prompts and provide GPT with the course information with different levels of detail, thereby generating high-quality course concepts and identifying their relations. Furthermore, we comprehensively evaluate the quality of the generated concepts and relationships through extensive experiments. Our results demonstrate the viability of LLMs as a tool for supporting educational content selection and delivery.

Language models (LMs) are increasingly being integrated into a wide range of applications, yet the modern evaluation paradigm does not sufficiently reflect how they are actually being used. Current evaluations rely on benchmarks that often lack direct applicability to the real-world contexts in which LMs are being deployed. To address this gap, we propose Dimensional and Contextual Evaluation (DICE), an approach that evaluates LMs on granular, context-dependent dimensions. In this position paper, we begin by examining the insufficiency of existing LM benchmarks, highlighting their limited applicability to real-world use cases. Next, we propose a set of granular evaluation parameters that capture dimensions of LM behavior that are more meaningful to stakeholders across a variety of application domains. Specifically, we introduce the concept of context-agnostic parameters - such as robustness, coherence, and epistemic honesty - and context-specific parameters that must be tailored to the specific contextual constraints and demands of stakeholders choosing to deploy LMs into a particular setting. We then discuss potential approaches to operationalize this evaluation framework, finishing with the opportunities and challenges DICE presents to the LM evaluation landscape. Ultimately, this work serves as a practical and approachable starting point for context-specific and stakeholder-relevant evaluation of LMs.

Robots with wheeled, quadrupedal, or humanoid forms are increasingly integrated into built environments. However, unlike human social learning, they lack a critical pathway for intrinsic cognitive development, namely, learning from human feedback during interaction. To understand human ubiquitous observation, supervision, and shared control in dynamic and uncertain environments, this study presents a brain-computer interface (BCI) framework that enables classification of Electroencephalogram (EEG) signals to detect cognitively demanding and safety-critical events. As a timely and motivating co-robotic engineering application, we simulate a human-in-the-loop scenario to flag risky events in semi-autonomous robotic driving-representative of long-tail cases that pose persistent bottlenecks to the safety performance of smart mobility systems and robotic vehicles. Drawing on recent advances in few-shot learning, we propose a dual-attention Siamese convolutional network paired with Dynamic Time Warping Barycenter Averaging approach to generate robust EEG-encoded signal representations. Inverse source localization reveals activation in Broadman areas 4 and 9, indicating perception-action coupling during task-relevant mental imagery. The model achieves 80% classification accuracy under data-scarce conditions and exhibits a nearly 100% increase in the utility of salient features compared to state-of-the-art methods, as measured through integrated gradient attribution. Beyond performance, this study contributes to our understanding of the cognitive architecture required for BCI agents-particularly the role of attention and memory mechanisms-in categorizing diverse mental states and supporting both inter- and intra-subject adaptation. Overall, this research advances the development of cognitive robotics and socially guided learning for service robots in complex built environments.