Distillation has emerged as a practical and effective approach to enhance the reasoning capabilities of open-source language models. In this work, we conduct a large-scale empirical study on reasoning data distillation by collecting verified outputs from three state-of-the-art teacher models-AM-Thinking-v1, Qwen3-235B-A22B, and DeepSeek-R1-on a shared corpus of 1.89 million queries. We construct three parallel datasets and analyze their distributions, revealing that AM-Thinking-v1-distilled data exhibits greater token length diversity and lower perplexity. Student models trained on each dataset are evaluated on reasoning benchmarks including AIME2024, AIME2025, MATH500, and LiveCodeBench. The model distilled from AM-Thinking-v1 consistently achieves the best performance (e.g., 84.3 on AIME2024, 72.2 on AIME2025, 98.4 on MATH500, and 65.9 on LiveCodeBench) and demonstrates adaptive output behavior-producing longer responses for harder tasks and shorter ones for simpler tasks. These findings highlight the value of high-quality, verified reasoning traces. We release the AM-Thinking-v1 and Qwen3-235B-A22B distilled datasets to support future research on open and high-performing reasoning-oriented language models. The datasets are publicly available on Hugging Face\footnote{Datasets are available on Hugging Face: \href{https://huggingface.co/datasets/a-m-team/AM-Thinking-v1-Distilled}{AM-Thinking-v1-Distilled}, \href{https://huggingface.co/datasets/a-m-team/AM-Qwen3-Distilled}{AM-Qwen3-Distilled}.}.

While integrating external tools into large language models (LLMs) enhances their ability to access real-time information and domain-specific services, existing approaches focus narrowly on functional tool selection following user instructions, overlooking the context-aware personalization in tool selection. This oversight leads to suboptimal user satisfaction and inefficient tool utilization, particularly when overlapping toolsets require nuanced selection based on contextual factors. To bridge this gap, we introduce ToolSpectrum, a benchmark designed to evaluate LLMs' capabilities in personalized tool utilization. Specifically, we formalize two key dimensions of personalization, user profile and environmental factors, and analyze their individual and synergistic impacts on tool utilization. Through extensive experiments on ToolSpectrum, we demonstrate that personalized tool utilization significantly improves user experience across diverse scenarios. However, even state-of-the-art LLMs exhibit the limited ability to reason jointly about user profiles and environmental factors, often prioritizing one dimension at the expense of the other. Our findings underscore the necessity of context-aware personalization in tool-augmented LLMs and reveal critical limitations for current models. Our data and code are available at https://github.com/Chengziha0/ToolSpectrum.

Large Language Models (LLMs) increasingly incorporate multilingual capabilities, fueling the demand to transfer them into target language-specific models. However, most approaches, which blend the source model's embedding by replacing the source vocabulary with the target language-specific vocabulary, may constrain expressive capacity in the target language since the source model is predominantly trained on English data. In this paper, we propose Semantic Aware Linear Transfer (SALT), a novel cross-lingual transfer technique that recycles embeddings from target language Pre-trained Language Models (PLMs) to transmit the deep representational strengths of PLM-derived embedding to LLMs. SALT derives unique regression lines based on the similarity in the overlap of the source and target vocabularies, to handle each non-overlapping token's embedding space. Our extensive experiments show that SALT significantly outperforms other transfer methods and achieves lower loss with accelerating faster convergence during language adaptation. Notably, SALT obtains remarkable performance in cross-lingual understanding setups compared to other methods. Furthermore, we highlight the scalable use of PLMs to enhance the functionality of contemporary LLMs by conducting experiments with varying architectures.

Much scholarship considers how U.S. federal agencies govern artificial intelligence (AI) through rulemaking and their own internal use policies. But agencies have an overlooked AI governance role: setting discretionary grant policy when directing billions of dollars in federal financial assistance. These dollars enable state and local entities to study, create, and use AI. This funding not only goes to dedicated AI programs, but also to grantees using AI in the course of meeting their routine grant objectives. As discretionary grantmakers, agencies guide and restrict what grant winners do -- a hidden lever for AI governance. Agencies pull this lever by setting program objectives, judging criteria, and restrictions for AI use. Using a novel dataset of over 40,000 non-defense federal grant notices of funding opportunity (NOFOs) posted to the U.S. federal grants website between 2009 and 2024, we analyze how agencies regulate the use of AI by grantees. We select records mentioning AI and review their stated goals and requirements. We find agencies promoting AI in notice narratives, shaping adoption in ways other records of grant policy might fail to capture. Of the grant opportunities that mention AI, we find only a handful of AI-specific judging criteria or restrictions. This silence holds even when agencies fund AI uses in contexts affecting people's rights and which, under an analogous federal procurement regime, would result in extra oversight. These findings recast grant notices as a site of AI policymaking -- albeit one that is developing out of step with other regulatory efforts and incomplete in its consideration of transparency, accountability, and privacy protections. The paper concludes by drawing lessons from AI procurement scholarship, while identifying distinct challenges in grantmaking that invite further study.

Effective social intelligence simulation requires language agents to dynamically adjust reasoning depth, a capability notably absent in current studies. Existing methods either lack this kind of reasoning capability or enforce Long Chain-of-Thought reasoning uniformly across all scenarios, resulting in excessive token usage and inflexible social simulation. To address this, we propose an $\textbf{A}$daptive $\textbf{M}$ode $\textbf{L}$earning ($\textbf{AML}$) framework in this paper, aiming to improve the adaptive thinking ability of language agents in dynamic social interactions. To this end, we first identify hierarchical thinking modes ranging from intuitive response to deep deliberation based on the cognitive control theory. We then develop the $\textbf{A}$daptive $\textbf{M}$ode $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{AMPO}$) algorithm to optimize the context-aware mode switching and reasoning. Our framework advances existing research in three key aspects: (1) Multi-granular thinking mode design, (2) Context-aware mode switching across social interaction, and (3) Token-efficient reasoning via depth-adaptive processing. Extensive experiments on social intelligence benchmarks verify that AML achieves 15.6% higher task performance than GPT-4o. Notably, our AMPO outperforms GRPO by 7.0% with 32.8% shorter reasoning chains, demonstrating the advantage of adaptive thinking mode selection and optimization mechanism in AMPO over GRPO's fixed-depth solution.

The rapid advancement of multimodal large language models (MLLMs) has profoundly impacted the document domain, creating a wide array of application scenarios. This progress highlights the need for a comprehensive benchmark to evaluate these models' capabilities across various document-specific tasks. However, existing benchmarks often fail to locate specific model weaknesses or guide systematic improvements. To bridge this gap, we introduce a General Document Intelligence Benchmark (GDI-Bench), featuring 2.3k images across 9 key scenarios and 19 document-specific tasks. By decoupling visual complexity and reasoning complexity, the GDI-Bench structures graded tasks that allow performance assessment by difficulty, aiding in model weakness identification and optimization guidance. We evaluate various open-source and closed-source models on GDI-Bench, conducting decoupled analyses in the visual and reasoning domains, revealing their strengths and weaknesses. To address the diverse tasks and domains in the GDI-Bench, we propose a GDI-Model that mitigates catastrophic forgetting during the supervised fine-tuning (SFT) process through an intelligence-preserving training strategy, thereby reinforcing the inherent weaknesses of the base model. Our model achieves state-of-the-art performance on previous benchmarks and the GDI-Bench. Both our benchmark and models are or will be open-sourced on https://huggingface.co/GDIBench.

Embedded Computing Systems Faculty of Informatics, TU Wien Vienna, Austria daniel.mueller-gritschneder@tuwien.ac.at Abstract --Selecting data points for model training is critical in machine learning. Effective selection methods can reduce the labeling effort, optimize on-device training for embedded systems with limited data storage, and enhance the model performance. This paper introduces a novel algorithm that uses Grad-CAM to make online decisions about retaining or discarding data points. Optimized for embedded devices, the algorithm computes a unique DRIP Score to quantify the importance of each data point. This enables dynamic decision-making on whether a data point should be stored for potential retraining or discarded without compromising model performance. Experimental evaluations on four benchmark datasets demonstrate that our approach can match or even surpass the accuracy of models trained on the entire dataset, while achieving storage savings of up to 39%. T o our knowledge, this is the first algorithm to make online decisions about data point retention without requiring access to the entire dataset. In the rapidly evolving domain of machine learning, the quantity of available data have reached unprecedented levels. While large datasets have traditionally been the bedrock of robust machine learning models, the sheer magnitude of data now available poses both opportunities and challenges. One of the primary challenges is efficient data management, especially but not only in scenarios with constrained computational and storage resources [1].

This paper elaborates on the concept of moral exercises as a means to help AI actors cultivate virtues that enable effective human oversight of AI systems. We explore the conceptual framework and significance of moral exercises, situating them within the contexts of philosophical discourse, ancient practices, and contemporary AI ethics scholarship. We outline the core pillars of the moral exercises methodology -- eliciting an engaged personal disposition, fostering relational understanding, and cultivating technomoral wisdom -- and emphasize their relevance to key activities and competencies essential for human oversight of AI systems. Our argument is supported by findings from three pilot studies involving a company, a multidisciplinary team of AI researchers, and higher education students. These studies allow us to explore both the potential and the limitations of moral exercises. Based on the collected data, we offer insights into how moral exercises can foster a responsible AI culture within organizations, and suggest directions for future research.

Complex narrative contexts often challenge language models' ability to follow instructions, and existing benchmarks fail to capture these difficulties. To address this, we propose Concise-SAE, a training-free framework that improves instruction following by identifying and editing instruction-relevant neurons using only natural language instructions, without requiring labelled data. To thoroughly evaluate our method, we introduce FreeInstruct, a diverse and realistic benchmark of 1,212 examples that highlights the challenges of instruction following in narrative-rich settings. While initially motivated by complex narratives, Concise-SAE demonstrates state-of-the-art instruction adherence across varied tasks without compromising generation quality.

Students often do not fully understand the code they have written. This sometimes does not become evident until later in their education, which can mean it is harder to fix their incorrect knowledge or misunderstandings. In addition, being able to fully understand code is increasingly important in a world where students have access to generative artificial intelligence (GenAI) tools, such as GitHub Copilot. One effective solution is to utilise code comprehension questions, where a marker asks questions about a submission to gauge understanding, this can also have the side effect of helping to detect plagiarism. However, this approach is time consuming and can be difficult and/or expensive to scale. This paper introduces AutoMCQ, which uses GenAI for the automatic generation of multiple-choice code comprehension questions. This is integrated with the CodeRunner automated assessment platform.