This study evaluates the integration of AI-powered robots in early childhood education, focusing on their impact on emotional self-regulation, engagement, and collaborative skills. A ten-week experimental design involving two groups of children assessed the robot's effectiveness through progress assessments, parental surveys, and teacher feedback. Results demonstrated that early exposure to the robot significantly enhanced emotional recognition, while sustained interaction further improved collaborative and social engagement. Parental and teacher feedback highlighted high acceptance levels, emphasizing the robot's ease of integration and positive influence on classroom dynamics. This research underscores the transformative potential of AI and robotics in education. The findings advocate for the broader adoption of AI-powered interventions, carefully examining equitable access, ethical considerations, and sustainable implementation. This work sets a foundation for exploring long-term impacts and expanding applications of AI in inclusive and impactful educational settings.

Climate-Eval is a comprehensive benchmark designed to evaluate natural language processing models across a broad range of tasks related to climate change. Climate-Eval aggregates existing datasets along with a newly developed news classification dataset, created specifically for this release. This results in a benchmark of 25 tasks based on 13 datasets, covering key aspects of climate discourse, including text classification, question answering, and information extraction. Our benchmark provides a standardized evaluation suite for systematically assessing the performance of large language models (LLMs) on these tasks. Additionally, we conduct an extensive evaluation of open-source LLMs (ranging from 2B to 70B parameters) in both zero-shot and few-shot settings, analyzing their strengths and limitations in the domain of climate change.

Recently, significant progress has been made in developing reasoning-capable Large Language Models (LLMs) through long Chain-of-Thought (CoT) techniques. However, this long-CoT reasoning process imposes substantial memory overhead due to the large Key-Value (KV) Cache memory overhead. Post-training KV Cache quantization has emerged as a promising compression technique and has been extensively studied in short-context scenarios. However, directly applying existing methods to long-CoT LLMs causes significant performance degradation due to the following two reasons: (1) Large cumulative error: Existing methods fail to adequately leverage available memory, and they directly quantize the KV Cache during each decoding step, leading to large cumulative quantization error. (2) Short-context calibration: Due to Rotary Positional Embedding (RoPE), the use of short-context data during calibration fails to account for the distribution of less frequent channels in the Key Cache, resulting in performance loss. We propose Progressive Mixed-Precision KV Cache Quantization (PM-KVQ) for long-CoT LLMs to address the above issues in two folds: (1) To reduce cumulative error, we design a progressive quantization strategy to gradually lower the bit-width of KV Cache in each block. Then, we propose block-wise memory allocation to assign a higher bit-width to more sensitive transformer blocks. (2) To increase the calibration length without additional overhead, we propose a new calibration strategy with positional interpolation that leverages short calibration data with positional interpolation to approximate the data distribution of long-context data. Extensive experiments on 7B-70B long-CoT LLMs show that PM-KVQ improves reasoning benchmark performance by up to 8% over SOTA baselines under the same memory budget. Our code is available at https://github.com/thu-nics/PM-KVQ.

We present MSA-MathEval, our submission to the BEA 2025 Shared Task on evaluating AI tutor responses across four instructional dimensions: Mistake Identification, Mistake Location, Providing Guidance, and Actionability. Our approach uses a unified training pipeline to fine-tune a single instruction-tuned language model across all tracks, without any task-specific architectural changes. To improve prediction reliability, we introduce a disagreement-aware ensemble inference strategy that enhances coverage of minority labels. Our system achieves strong performance across all tracks, ranking 1st in Providing Guidance, 3rd in Actionability, and 4th in both Mistake Identification and Mistake Location. These results demonstrate the effectiveness of scalable instruction tuning and disagreement-driven modeling for robust, multi-dimensional evaluation of LLMs as educational tutors.

Recent advances in cross-prompt automated essay scoring (AES) typically train models jointly on all source prompts, often requiring additional access to unlabeled target prompt essays simultaneously. However, using all sources is suboptimal in our pilot study, and re-accessing source datasets during adaptation raises privacy concerns. We propose a source-free adaptation approach that selectively merges individually trained source models' parameters instead of datasets. In particular, we simulate joint training through linear combinations of task vectors -- the parameter updates from fine-tuning. To optimize the combination's coefficients, we propose Prior-encoded Information Maximization (PIM), an unsupervised objective which promotes the model's score discriminability regularized by priors pre-computed from the sources. We employ Bayesian optimization as an efficient optimizer of PIM. Experimental results with LLMs on in-dataset and cross-dataset adaptation show that our method (1) consistently outperforms training jointly on all sources, (2) maintains superior robustness compared to other merging methods, (3) excels under severe distribution shifts where recent leading cross-prompt methods struggle, all while retaining computational efficiency.

Cross-capability transfer is a key challenge in large language model (LLM) research, with applications in multi-task integration, model compression, and continual learning. Recent works like FuseLLM and FuseChat have demonstrated the potential of transferring multiple model capabilities to lightweight models, enhancing adaptability and efficiency, which motivates our investigation into more efficient cross-capability transfer methods. However, existing approaches primarily focus on small, homogeneous models, limiting their applicability. For large, heterogeneous models, knowledge distillation with full-parameter fine-tuning often overlooks the student model's intrinsic capacity and risks catastrophic forgetting, while PEFT methods struggle to effectively absorb knowledge from source LLMs. To address these issues, we introduce GraftLLM, a novel method that stores source model capabilities in a target model with SkillPack format. This approach preserves general capabilities, reduces parameter conflicts, and supports forget-free continual learning and model fusion. We employ a module-aware adaptive compression strategy to compress parameter updates, ensuring efficient storage while maintaining task-specific knowledge. The resulting SkillPack serves as a compact and transferable knowledge carrier, ideal for heterogeneous model fusion and continual learning. Experiments across various scenarios demonstrate that GraftLLM outperforms existing techniques in knowledge transfer, knowledge fusion, and forget-free learning, providing a scalable and efficient solution for cross-capability transfer. The code is publicly available at: https://github.com/duguodong7/GraftLLM.

--Federated Learning (FL) facilitates the fine-tuning of Foundation Models (FMs) using distributed data sources, with Low-Rank Adaptation (LoRA) gaining popularity due to its low communication costs and strong performance. While recent work acknowledges the benefits of heterogeneous LoRA in FL and introduces flexible algorithms to support its implementation, our theoretical analysis reveals a critical gap: existing methods lack formal convergence guarantees due to parameter truncation and biased gradient updates. Specifically, adapting client-specific LoRA ranks necessitates truncating global parameters, which introduces inherent truncation errors and leads to subsequent inaccurate gradient updates that accumulate over training rounds, ultimately degrading performance. T o address the above issues, we propose FedHL, a simple yet effective Federated Learning framework tailored for Heterogeneous LoRA. By leveraging the full-rank global model as a calibrated aggregation basis, FedHL eliminates the direct truncation bias from initial alignment with client-specific ranks. Furthermore, we derive the theoretically optimal aggregation weights by minimizing the gradient drift term in the convergence upper bound. Our analysis shows that FedHL guarantees O (1 / T) convergence rate, and experiments on multiple real-world datasets demonstrate a 1-3% improvement over several state-of-the-art methods. Zihao Peng, Jiandian Zeng, and Guo Li are with the Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China (e-mail: { pzh cs, liguo }@mail.bnu.edu.cn; Boyuan Li is with the School of Computer Science and Artificial Intelligence, Zhengzhou University, Zhengzhou 450001, China (e-mail: l202311841010602@gs.zzu.edu.cn). Shengbo Chen is with the School of Software, Nanchang University, Nanchang 330000, China (e-mail: ccb02kingdom@gmail.com).

Error correction is an important capability when applying large language models (LLMs) to facilitate user typing on mobile devices. In this paper, we use LLMs to synthesize a high-quality dataset of error correction pairs to evaluate and improve LLMs for mobile applications. We first prompt LLMs with error correction domain knowledge to build a scalable and reliable addition to the existing data synthesis pipeline. We then adapt the synthetic data distribution to match the mobile application domain by reweighting the samples. The reweighting model is learnt by predicting (a handful of) live A/B test metrics when deploying LLMs in production, given the LLM performance on offline evaluation data and scores from a small privacy-preserving on-device language model. Finally, we present best practices for mixing our synthetic data with other data sources to improve model performance on error correction in both offline evaluation and production live A/B testing.

Recent advances in large reasoning models (LRMs) show strong performance in structured domains such as mathematics and programming; however, they often lack pedagogical coherence and realistic teaching behaviors. To bridge this gap, we introduce Pedagogy-R1, a framework that adapts LRMs for classroom use through three innovations: (1) a distillation-based pipeline that filters and refines model outputs for instruction-tuning, (2) the Well-balanced Educational Benchmark (WBEB), which evaluates performance across subject knowledge, pedagogical knowledge, tracing, essay scoring, and teacher decision-making, and (3) a Chain-of-Pedagogy (CoP) prompting strategy for generating and eliciting teacher-style reasoning. Our mixed-method evaluation combines quantitative metrics with qualitative analysis, providing the first systematic assessment of LRMs' pedagogical strengths and limitations.

To tackle the huge computational demand of large foundation models, activation-aware compression techniques without retraining have been introduced. However, since these rely on calibration data, domain shift may arise for unknown downstream tasks. With a computationally efficient calibration, activation-aware pruning can be executed for every prompt adaptively, yet achieving reduced complexity at inference. We formulate it as a mixture of micro-experts, called $μ$-MoE. Several experiments demonstrate that $μ$-MoE can dynamically adapt to task/prompt-dependent structured sparsity on the fly.