We present Nanbeige4-3B, a family of small-scale but high-performing language models. Pretrained on 23T high-quality tokens and finetuned on over 30 million diverse instructions, we extend the boundary of the scaling law for small language models. In pre-training, we design a Fine-Grained Warmup-Stable-Decay (FG-WSD) training scheduler, which progressively refines data mixtures across stages to boost model performance. In post-training, to improve the quality of the SFT data, we design a joint mechanism that integrates deliberative generation refinement and chain-of-thought reconstruction, yielding substantial gains on complex tasks. Following SFT, we employ our flagship reasoning model to distill Nanbeige4-3B through our proposed Dual Preference Distillation (DPD) method, which leads to further performance gains. Finally, a multi-stage reinforcement learning phase was applied, leveraging verifiable rewards and preference modeling to strengthen abilities on both reasoning and human alignment. Extensive evaluations show that Nanbeige4-3B not only significantly outperforms models of comparable parameter scale but also rivals much larger models across a wide range of benchmarks. The model checkpoints are available at https://huggingface.co/Nanbeige.

Sentence simplification aims to modify a sentence to make it easier to read and understand while preserving the meaning. Different applications require distinct simplification policies, such as replacing only complex words at the lexical level or rewriting the entire sentence while trading off details for simplicity. However, achieving such policy-driven control remains an open challenge. In this work, we introduce a simple yet powerful approach that leverages Large Language Model-as-a-Judge (LLM-as-a-Judge) to automatically construct policy-aligned training data, completely removing the need for costly human annotation or parallel corpora. Our method enables building simplification systems that adapt to diverse simplification policies. Sentence simplification could benefit users with reading difficulties, such as second-language (L2) learners and people with reading impairments (e.g., dyslexic individuals), by making text easier to read and understand (Alva-Manchego et al., 2020b). It involves a series of edits, such as lexical paraphrasing, sentence splitting, and removing irrelevant details (Xu et al., 2015). The preferred edit policy, i.e., permissible or appropriate edits in given texts, varies significantly depending on the target audience. In L2 education, one of the major application areas for simplification, previous work in both NLP and language education research has shown that the desired type and degree of simplification edits change depending on learner proficiency and readability levels (Agrawal et al., 2021; Zhong et al., 2020). Specifically, low-to intermediate-level learners benefit from a combination of lexical paraphrasing, structural modifications, and selective deletions to reduce cognitive load. In contrast, advanced learners benefit from lexical paraphrasing, which supports vocabulary acquisition (Chen, 2019), but they gain comparatively less from added cohesion or deletion (Hosoda, 2016; Zhong et al., 2020). Motivated by these findings, we introduce two distinct edit policies. As illustrated in Table 1, overall-rewriting simplification often combines lexical paraphrasing, structural modifications, and deletions to improve readability for intermediate-level language learners. In contrast, lexical-paraphrasing (Paetzold & Specia, 2016; Li et al., 2025) adheres to the original sentence closely while supporting more efficient vocabulary acquisition for advanced learners.

What if users could meet their future selves today? AI-generated future selves simulate meaningful encounters with a digital twin decades in the future. As AI systems advance, combining cloned voices, age-progressed facial rendering, and autobiographical narratives, a central question emerges: Does the modality of these future selves alter their psychological and affective impact? How might a text-based chatbot, a voice-only system, or a photorealistic avatar shape present-day decisions and our feeling of connection to the future? We report a randomized controlled study (N=92) evaluating three modalities of AI-generated future selves (text, voice, avatar) against a neutral control condition. We also report a systematic model evaluation between Claude 4 and three other Large Language Models (LLMs), assessing Claude 4 across psychological and interaction dimensions and establishing conversational AI quality as a critical determinant of intervention effectiveness. All personalized modalities strengthened Future Self-Continuity (FSC), emotional well-being, and motivation compared to control, with avatar producing the largest vividness gains, yet with no significant differences between formats. Interaction quality metrics, particularly persuasiveness, realism, and user engagement, emerged as robust predictors of psychological and affective outcomes, indicating that how compelling the interaction feels matters more than the form it takes. Content analysis found thematic patterns: text emphasized career planning, while voice and avatar facilitated personal reflection. Claude 4 outperformed ChatGPT 3.5, Llama 4, and Qwen 3 in enhancing psychological, affective, and FSC outcomes.

Conventional wisdom in the age of LLMs dictates that solving IQ-test-like visual puzzles from the ARC-AGI-1 benchmark requires capabilities derived from massive pretraining. To counter this, we introduce CompressARC, a 76K parameter model without any pretraining that solves 20% of evaluation puzzles by minimizing the description length (MDL) of the target puzzle purely during inference time. The MDL endows CompressARC with extreme generalization abilities typically unheard of in deep learning. To our knowledge, CompressARC is the only deep learning method for ARC-AGI where training happens only on a single sample: the target inference puzzle itself, with the final solution information removed. Moreover, CompressARC does not train on the pre-provided ARC-AGI "training set". Under these extremely data-limited conditions, we do not ordinarily expect any puzzles to be solvable at all. Yet CompressARC still solves a diverse distribution of creative ARC-AGI puzzles, suggesting MDL to be an alternative feasible way to produce intelligence, besides conventional pretraining.

Assessment of medication history-taking has traditionally relied on human observation, limiting scalability and detailed performance data. While Generative AI (GenAI) platforms enable extensive data collection and learning analytics provide powerful methods for analyzing educational traces, these approaches remain largely underexplored in pharmacy clinical training. This study addresses this gap by applying learning analytics to understand how students develop clinical communication competencies with GenAI-powered virtual patients -- a crucial endeavor given the diversity of student cohorts, varying language backgrounds, and the limited opportunities for individualized feedback in traditional training settings. We analyzed 323 students' interaction logs across Australian and Malaysian institutions, comprising 50,871 coded utterances from 1,487 student-GenAI dialogues. Combining Epistemic Network Analysis to model inquiry co-occurrences with Sequential Pattern Mining to capture temporal sequences, we found that high performers demonstrated strategic deployment of information recognition behaviors. Specifically, high performers centered inquiry on recognizing clinically relevant information, integrating rapport-building and structural organization, while low performers remained in routine question-verification loops. Demographic factors including first-language background, prior pharmacy work experience, and institutional context, also shaped distinct inquiry patterns. These findings reveal inquiry patterns that may indicate clinical reasoning development in GenAI-assisted contexts, providing methodological insights for health professions education assessment and informing adaptive GenAI system design that supports diverse learning pathways.

While large language models (LLMs) have introduced novel paradigms in science and education, their adoption in higher education is constrained by inherent limitations. These include a tendency to produce inaccuracies and high computational requirements, which compromise the strict demands for accurate and reliable knowledge essential in higher education. Small language models (MiniLMs), by contrast, offer distinct advantages in professional education due to their lightweight nature and precise retrieval capabilities. This research takes "Atmospheric Physics" as an example. We established a specialized corpus and image repository by gathering over 550,000 full-text PDFs from over 130 international well-respected journals in Earth and environmental science. From this collection, we extracted over 100 million high-quality sentence-level corpus and more than 3 million high-resolution academic images. Using MiniLMs, these resources were organized into a high-dimensional vector library for precise retrieval and efficient utilization of extensive educational content. Consequently, we systematically redesigned the courses, textbooks, and teaching strategies for "Atmospheric Physics" based on MiniLMs. The course is designed as a "interdisciplinary-frontier" system, breaking down traditional boundaries between atmospheric science, space science, hydrology, and remote sensing. Teaching materials are transformed from static, lagging text formats into a dynamic digital resource library powered by MiniLM. For teaching methods, we have designed a question-based learning pathway. This paradigm promotes a shift from passive knowledge transfer to active cognitive development. Consequently, this MiniLM-driven "Atmospheric Physics" course demonstrates a specific avenue for "AI for education".

With the rapid advancement of conversational and diffusion-based AI, there is a growing adoption of AI in educational services, ranging from grading and assessment tools to personalized learning systems that provide targeted support for students. However, this adaptability has yet to fully extend to the domain of children's speech, where existing models often fail due to their reliance on datasets designed for clear, articulate adult speech. Children, particularly those in early developmental stages or with speech and language pathologies, present unique challenges that current AI models and datasets are ill-equipped to handle. To address this, we introduce KidSpeak, a multi-task speech-enhanced Foundation Model capable of both generative and discriminative tasks specifically tailored to children's speech patterns. Our framework employs a two-stage training process that incorporates phonetic knowledge into the speech encoder, achieving an average accuracy of 87% across four separate tasks. Furthermore, recognizing the limitations of scalable human annotation and existing speech alignment tools, we propose the Flexible and Automatic Speech Aligner (F ASA) and leverage the method to construct high quality datasets for training and evaluation. This novel alignment tool significantly improves the quality of aligned children's speech from noisy data, enhancing data quality by 13.6 compared to human annotations, as demonstrated on the CHILDES dataset. To the best of our knowledge, KidSpeak and F ASA represent the first comprehensive solution designed for speech and language therapy in children, offering both a multi-purpose speech LLM and a robust alignment tool.

Long-horizon, dexterous manipulation tasks such as furniture assembly, food preparation, and manufacturing have been a holy grail in robotics. Recent large robot action models (T eam et al., 2025; Black et al., 2024; Kim et al., 2024) have made substantial breakthroughs towards these goals by imitating expert demonstrations of diverse qualities. We provide a more comprehensive review of related work in Section 6, but highlight here a key trend: while supervised learning from demonstration, also known as behavior cloning (BC), has been applied across domains for decades (Pomerleau, 1988), its recent success in robotic manipulation has coincided with the adoption of what we term generative control policies (GCPs): robotic control policies that use generative modeling architectures, such as diffusion models, flow models, and autoregressive transformers, as parameterizations of the mapping from observation to action. Given the seemingly transformative nature of GCPs for robot learning, there has been much speculation about the origin of their superior performance relative to policies trained with a regression loss, henceforth regression control policies (RCPs). GCPs, by modeling conditional distributions over actions, are uniquely suited to the multi-task pretraining paradigm popular in today's large robotic models.

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone for aligning large language models (LLMs) with human values. However, these methods typically assume that preference data is clean and that all labels are equally reliable. In practice, large-scale preference datasets contain substantial noise due to annotator mistakes, inconsistent instructions, varying expertise, and even adversarial or low-effort feedback. This mismatch between recorded labels and ground-truth preferences can misguide training and degrade model performance. To address this issue, we introduce Robust Enhanced Policy Optimization (RE-PO), which uses an expectation-maximization procedure to infer the posterior correctness of each label and then adaptively reweight data points in the training loss to mitigate label noise. We further generalize this idea by establishing a theoretical link between arbitrary preference losses and their underlying probabilistic models, enabling a systematic transformation of existing alignment algorithms into robust counterparts and elevating RE-PO from a single method to a general framework for robust preference alignment. Theoretically, we prove that, under a perfectly calibrated model, RE-PO recovers the true noise level of the dataset. Empirically, we show that RE-PO consistently improves four state-of-the-art alignment methods (DPO, IPO, SimPO, and CPO); when applied to Mistral and Llama 3 models, the RE-PO-enhanced variants increase AlpacaEval 2 win rates by up to 7.0 percent over their respective baselines.

Orthogonal Gradient Descent (OGD) has emerged as a powerful method for continual learning. However, its Euclidean projections do not leverage the underlying information-geometric structure of the problem, which can lead to suboptimal convergence in learning tasks. To address this, we propose incorporating the natural gradient into OGD and present \textbf{ONG (Orthogonal Natural Gradient Descent)}. ONG preconditions each new task-specific gradient with an efficient EKFAC approximation of the inverse Fisher information matrix, yielding updates that follow the steepest descent direction under a Riemannian metric. To preserve performance on previously learned tasks, ONG projects these natural gradients onto the orthogonal complement of prior tasks' natural gradients. We provide an initial theoretical justification for this procedure, introduce the Orthogonal Natural Gradient Descent (ONG) algorithm, and present preliminary results on the Permuted and Rotated MNIST benchmarks. Our preliminary results, however, indicate that a naive combination of natural gradients and orthogonal projections has potential issues. This finding has motivated continued future work focused on robustly reconciling these geometric perspectives to develop a continual learning method, establishing a more rigorous theoretical foundation with formal convergence guarantees, and extending empirical validation to large-scale continual learning benchmarks.