Despite the impressive performance of large language models (LLMs) in general domains, they often underperform in specialized domains. Existing approaches typically rely on data synthesis methods and yield promising results by using unlabeled data to capture domain-specific features. However, these methods either incur high computational costs or suffer from performance limitations, while also demonstrating insufficient generalization across different tasks. To address these challenges, we propose AQuilt, a framework for constructing instruction-tuning data for any specialized domains from corresponding unlabeled data, including Answer, Question, Unlabeled data, Inspection, Logic, and Task type. By incorporating logic and inspection, we encourage reasoning processes and self-inspection to enhance model performance. Moreover, customizable task instructions enable high-quality data generation for any task. As a result, we construct a dataset of 703k examples to train a powerful data synthesis model. Experiments show that AQuilt is comparable to DeepSeek-V3 while utilizing just 17% of the production cost. Further analysis demonstrates that our generated data exhibits higher relevance to downstream tasks. Source code, models, and scripts are available at https://github.com/Krueske/AQuilt.

Training deep neural networks typically requires large datasets, which may not be available in privacy-sensitive or resource-constrained domains. Data-free knowledge distillation (DFKD) [1] has emerged as a promising approach where a student model learns from synthetic data generated via a pretrained teacher network. However, existing DFKD methods often fail to ensure structural alignment between synthetic and real data. We propose C2G-KD, a method leveraging PCA-derived constraints to guide a conditional generator in producing class-specific synthetic samples without direct access to real data. Central to this approach is the use of PCA [2] to impose topological constraints from minimal real samples, ensuring that generated images structurally align with class manifolds.

Eye-tracking data reveals valuable insights into users' cognitive states but is difficult to analyze due to its structured, non-linguistic nature. While large language models (LLMs) excel at reasoning over text, they struggle with temporal and numerical data. This paper presents a multimodal human-AI collaborative framework designed to enhance cognitive pattern extraction from eye-tracking signals. The framework includes: (1) a multi-stage pipeline using horizontal and vertical segmentation alongside LLM reasoning to uncover latent gaze patterns; (2) an Expert-Model Co-Scoring Module that integrates expert judgment with LLM output to generate trust scores for behavioral interpretations; and (3) a hybrid anomaly detection module combining LSTM-based temporal modeling with LLM-driven semantic analysis. Our results across several LLMs and prompt strategies show improvements in consistency, interpretability, and performance, with up to 50% accuracy in difficulty prediction tasks. This approach offers a scalable, interpretable solution for cognitive modeling and has broad potential in adaptive learning, human-computer interaction, and educational analytics.

While large language models (LLMs) leverage both knowledge and reasoning during inference, the capacity to distinguish between them plays a pivotal role in model analysis, interpretability, and development. Inspired by dual-system cognitive theory, we propose a cognition attribution framework to decouple the contribution of knowledge and reasoning. In particular, the cognition of LLMs is decomposed into two distinct yet complementary phases: knowledge retrieval (Phase 1) and reasoning adjustment (Phase 2). To separate these phases, LLMs are prompted to generate answers under two different cognitive modes, fast thinking and slow thinking, respectively. The performance under different cognitive modes is analyzed to quantify the contribution of knowledge and reasoning. This architecture is employed to 15 LLMs across 3 datasets. Results reveal: (1) reasoning adjustment is domain-specific, benefiting reasoning-intensive domains (e.g., mathematics, physics, and chemistry) and potentially imparing knowledge-intensive domains. (2) Parameter scaling improves both knowledge and reasoning, with knowledge improvements being more pronounced. Additionally, parameter scaling make LLMs reasoning significantly more prudent, while moderately more intelligent. (3) Knowledge primarily resides in lower network layers, while reasoning operates in higher layers. Our framework not only helps understand LLMs from a "decoupling" perspective, but also provides new insights into existing research, including scaling laws, hierarchical knowledge editing, and limitations of small-model reasoning.

Recent work has shown that language models can self-improve by maximizing their own confidence in their predictions, without relying on external verifiers or reward signals. In this work, we study the test-time scaling of language models for mathematical reasoning tasks, where the model's own confidence is used to select the most promising attempts. Surprisingly, we find that we can achieve significant performance gains by continuing only the most promising attempt, selected by the model's prefix-confidence. We systematically evaluate prefix-confidence scaling on five mathematical reasoning datasets: the school-level GSM8K and MATH500, and the competition-level AMC23, AIME24, and AIME25. We find that prefix-confidence scaling with prefixes of only 32 tokens achieves a better accuracy-compute trade-off than majority voting. Moreover, prefix-confidence scaling appears less susceptible than BoN to length biases. Finally, we also evaluate test-time training with prefix-confidence and find that, while outperforming the base model, it does not improve over prefix-confidence scaling.

While AI systems demonstrate exponentially improving capabilities, the pace of AI research itself remains linearly bounded by human cognitive capacity, creating an increasingly severe development bottleneck. We present ASI-Arch, the first demonstration of Artificial Superintelligence for AI research (ASI4AI) in the critical domain of neural architecture discovery--a fully autonomous system that shatters this fundamental constraint by enabling AI to conduct its own architectural innovation. Moving beyond traditional Neural Architecture Search (NAS), which is fundamentally limited to exploring human-defined spaces, we introduce a paradigm shift from automated optimization to automated innovation. ASI-Arch can conduct end-to-end scientific research in the domain of architecture discovery, autonomously hypothesizing novel architectural concepts, implementing them as executable code, training and empirically validating their performance through rigorous experimentation and past experience. ASI-Arch conducted 1,773 autonomous experiments over 20,000 GPU hours, culminating in the discovery of 106 innovative, state-of-the-art (SOTA) linear attention architectures. Like AlphaGo's Move 37 that revealed unexpected strategic insights invisible to human players, our AI-discovered architectures demonstrate emergent design principles that systematically surpass human-designed baselines and illuminate previously unknown pathways for architectural innovation. Crucially, we establish the first empirical scaling law for scientific discovery itself--demonstrating that architectural breakthroughs can be scaled computationally, transforming research progress from a human-limited to a computation-scalable process. We provide comprehensive analysis of the emergent design patterns and autonomous research capabilities that enabled these breakthroughs, establishing a blueprint for self-accelerating AI systems.

Due to practical constraints such as partial observability and limited communication, Centralized Training with Decentralized Execution (CTDE) has become the dominant paradigm in cooperative Multi-Agent Reinforcement Learning (MARL). However, existing CTDE methods often underutilize centralized training or lack theoretical guarantees. We propose Multi-Agent Guided Policy Optimization (MAGPO), a novel framework that better leverages centralized training by integrating centralized guidance with decentralized execution. MAGPO uses an auto-regressive joint policy for scalable, coordinated exploration and explicitly aligns it with decentralized policies to ensure deployability under partial observability. We provide theoretical guarantees of monotonic policy improvement and empirically evaluate MAGPO on 43 tasks across 6 diverse environments. Results show that MAGPO consistently outperforms strong CTDE baselines and matches or surpasses fully centralized approaches, offering a principled and practical solution for decentralized multi-agent learning. Our code and experimental data can be found in https://github.com/liyheng/MAGPO.

Accurately assessing post quality requires complex relational reasoning to capture nuanced topic-post relationships. However, existing studies face three major limitations: (1) treating the task as unimodal categorization, which fails to leverage multimodal cues and fine-grained quality distinctions; (2) introducing noise during deep multimodal fusion, leading to misleading signals; and (3) lacking the ability to capture complex semantic relationships like relevance and comprehensiveness. To address these issues, we propose the Multimodal Fine-grained Topic-post Relational Reasoning (MFTRR) framework, which mimics human cognitive processes. MFTRR reframes post-quality assessment as a ranking task and incorporates multimodal data to better capture quality variations. It consists of two key modules: (1) the Local-Global Semantic Correlation Reasoning Module, which models fine-grained semantic interactions between posts and topics at both local and global levels, enhanced by a maximum information fusion mechanism to suppress noise; and (2) the Multi-Level Evidential Relational Reasoning Module, which explores macro- and micro-level relational cues to strengthen evidence-based reasoning. We evaluate MFTRR on three newly constructed multimodal topic-post datasets and the public Lazada-Home dataset. Experimental results demonstrate that MFTRR significantly outperforms state-of-the-art baselines, achieving up to 9.52% NDCG@3 improvement over the best unimodal method on the Art History dataset.

Large Language Model (LLM) agents are increasingly utilized in AI-aided education to support tutoring and learning. Effective communication strategies among LLM agents improve collaborative problem-solving efficiency and facilitate cost-effective adoption in education. However, little research has systematically evaluated the impact of different communication strategies on agents' problem-solving. Our study examines four communication modes, \textit{teacher-student interaction}, \textit{peer-to-peer collaboration}, \textit{reciprocal peer teaching}, and \textit{critical debate}, in a dual-agent, chat-based mathematical problem-solving environment using the OpenAI GPT-4o model. Evaluated on the MATH dataset, our results show that dual-agent setups outperform single agents, with \textit{peer-to-peer collaboration} achieving the highest accuracy. Dialogue acts like statements, acknowledgment, and hints play a key role in collaborative problem-solving. While multi-agent frameworks enhance computational tasks, effective communication strategies are essential for tackling complex problems in AI education.

Argument mining (AM) is an interdisciplinary research field that integrates insights from logic, philosophy, linguistics, rhetoric, law, psychology, and computer science. It involves the automatic identification and extraction of argumentative components, such as premises and claims, and the detection of relationships between them, such as support, attack, or neutrality. Recently, the field has advanced significantly, especially with the advent of large language models (LLMs), which have enhanced the efficiency of analyzing and extracting argument semantics compared to traditional methods and other deep learning models. There are many benchmarks for testing and verifying the quality of LLM, but there is still a lack of research and results on the operation of these models in publicly available argument classification databases. This paper presents a study of a selection of LLM's, using diverse datasets such as Args.me and UKP. The models tested include versions of GPT, Llama, and DeepSeek, along with reasoning-enhanced variants incorporating the Chain-of-Thoughts algorithm. The results indicate that ChatGPT-4o outperforms the others in the argument classification benchmarks. In case of models incorporated with reasoning capabilities, the Deepseek-R1 shows its superiority. However, despite their superiority, GPT-4o and Deepseek-R1 still make errors. The most common errors are discussed for all models. To our knowledge, the presented work is the first broader analysis of the mentioned datasets using LLM and prompt algorithms. The work also shows some weaknesses of known prompt algorithms in argument analysis, while indicating directions for their improvement. The added value of the work is the in-depth analysis of the available argument datasets and the demonstration of their shortcomings.