Abstract--The continual learning capability of large language models (LLMs) is crucial for advancing artificial general intelligence. However, continual fine-tuning LLMs across various domains often suffers from catastrophic forgetting, characterized by: 1) significant forgetting of their general capabilities, and 2) sharp performance declines in previously learned tasks. T o simultaneously address both issues in a simple yet stable manner, we propose General Sample Replay (GeRe), a framework that use usual pretraining texts for efficient anti-forgetting. Beyond revisiting the most prevalent replay-based practices under GeRe, we further leverage neural states to introduce a enhanced activation states constrained optimization method using threshold-based margin (TM) loss, which maintains activation state consistency during replay learning. We are the first to validate that a small, fixed set of pre-collected general replay samples is sufficient to resolve both concerns--retaining general capabilities while promoting overall performance across sequential tasks. Indeed, the former can inherently facilitate the latter. Through controlled experiments, we systematically compare TM with different replay strategies under the GeRe framework, including vanilla label fitting, logit imitation via KL divergence and feature imitation via L1/L2 losses. Results demonstrate that TM consistently improves performance and exhibits better robustness. Our work paves the way for efficient replay of LLMs for the future. Our code and data are available at https://github.com/Qznan/GeRe. The finetuned model forgets its original world knowledge or basic instruction-following skills [1], [2]. Y unan Zhang, Shuoran Jiang, Y ang Fan, Mengchen Zhao, Xiangping Wu, Qingcai Chen are with the Department of Computer Science and T echnology, Harbin Institute of T echnology, Shenzhen, China. Qingcai Chen and Xiangping Wu are the corresponding authors.

Studies on the use o f AI tools in CS course s focus on prescriptive uses of AI tools that preemptively assign a use case to the tool. To our knowledge, this is the first user study that surveys CS students on how they use AI tools for their coding tasks by personal choice. We surveyed 26 CS students and practitioners with various programming experiences using AI tools for coding tasks (i.e., writ e, debug, etc.). When asked about the most common resources they have used to write a 300 - line program from scratch, blog entries ( e.g., Stack Exchange, etc.) were their top choice, followed by AI coding assistants (e.g., GitHub Copilot, etc.). When asked about resources to debug code, AI chatbots (e.g., ChatGPT, etc.) were the most common choice, followed by blog entries. AI coding assistants are used more for writing code, while AI chatbots are used for debugging tasks. Respondents with all programming experience prefer online resources for coding tasks - whether AI - powered or not - rather than direct human help from peers an d instructors.

While AI agents show potential in scientific ideation, most existing frameworks rely on single-agent refinement, limiting creativity due to bounded knowledge and perspective. Inspired by real-world research dynamics, this paper investigates whether structured multi-agent discussions can surpass solitary ideation. We propose a cooperative multi-agent framework for generating research proposals and systematically compare configurations including group size, leader-led versus leaderless structures, and team compositions varying in interdisciplinarity and seniority. To assess idea quality, we employ a comprehensive protocol with agent-based scoring and human review across dimensions such as novelty, strategic vision, and integration depth. Our results show that multi-agent discussions substantially outperform solitary baselines. A designated leader acts as a catalyst, transforming discussion into more integrated and visionary proposals. Notably, we find that cognitive diversity is a primary driver of quality, yet expertise is a non-negotiable prerequisite, as teams lacking a foundation of senior knowledge fail to surpass even a single competent agent. These findings offer actionable insights for designing collaborative AI ideation systems and shed light on how team structure influences creative outcomes.

Fostering students' abilities for knowledge integration and transfer in complex problem-solving scenarios is a core objective of modern education, and interdisciplinary STEM is a key pathway to achieve this, yet it requires expert guidance that is difficult to scale. While LLMs offer potential in this regard, their true capability for guided instruction remains unclear due to the lack of an effective evaluation benchmark. To address this, we introduce SID, the first benchmark designed to systematically evaluate the higher-order guidance capabilities of LLMs in multi-turn, interdisciplinary Socratic dialogues. Our contributions include a large-scale dataset of 10,000 dialogue turns across 48 complex STEM projects, a novel annotation schema for capturing deep pedagogical features, and a new suite of evaluation metrics (e.g., X-SRG). Baseline experiments confirm that even state-of-the-art LLMs struggle to execute effective guided dialogues that lead students to achieve knowledge integration and transfer. This highlights the critical value of our benchmark in driving the development of more pedagogically-aware LLMs.

Lately, Personalized Federated Learning (PFL) has emerged as a prevalent paradigm to deliver personalized models by collaboratively training while simultaneously adapting to each client's local applications. Existing PFL methods typically face a significant challenge due to the ubiquitous data heterogeneity (i.e., non-IID data) across clients, which severely hinders convergence and degrades performance. We identify that the root issue lies in the long-standing reliance on gradient-based updates, which are inherently sensitive to non-IID data. To fundamentally address this issue and bridge the research gap, in this paper, we propose a Heterogeneity-invariant Personalized Federated learning scheme, named FedHiP, through analytical (i.e., closed-form) solutions to avoid gradient-based updates. Specifically, we exploit the trend of self-supervised pre-training, leveraging a foundation model as a frozen backbone for gradient-free feature extraction. Following the feature extractor, we further develop an analytic classifier for gradient-free training. To support both collective generalization and individual personalization, our FedHiP scheme incorporates three phases: analytic local training, analytic global aggregation, and analytic local personalization. The closed-form solutions of our FedHiP scheme enable its ideal property of heterogeneity invariance, meaning that each personalized model remains identical regardless of how non-IID the data are distributed across all other clients. Extensive experiments on benchmark datasets validate the superiority of our FedHiP scheme, outperforming the state-of-the-art baselines by at least 5.79%-20.97% in accuracy.

This paper addresses the critical need for scalable and high-quality educational assessment tools within the Malaysian education system. It highlights the potential of Generative AI (GenAI) while acknowledging the significant challenges of ensuring factual accuracy and curriculum alignment, especially for low-resource languages like Bahasa Melayu. This research introduces and compares four incremental pipelines for generating Form 1 Mathematics multiple-choice questions (MCQs) in Bahasa Melayu using OpenAI's GPT-4o. The methods range from non-grounded prompting (structured and basic) to Retrieval-Augmented Generation (RAG) approaches (one using the LangChain framework, one implemented manually). The system is grounded in official curriculum documents, including teacher-prepared notes and the yearly teaching plan (RPT). A dual-pronged automated evaluation framework is employed to assess the generated questions. Curriculum alignment is measured using Semantic Textual Similarity (STS) against the RPT, while contextual validity is verified through a novel RAG-based Question-Answering (RAG-QA) method. The results demonstrate that RAG-based pipelines significantly outperform non-grounded prompting methods, producing questions with higher curriculum alignment and factual validity. The study further analyzes the trade-offs between the ease of implementation of framework-based RAG and the fine-grained control offered by a manual pipeline. This work presents a validated methodology for generating curriculum-specific educational content in a low-resource language, introduces a symbiotic RAG-QA evaluation technique, and provides actionable insights for the development and deployment of practical EdTech solutions in Malaysia and similar regions.

High-quality, multi-turn instructional dialogues between novices and experts are essential for developing AI systems that support teaching, learning, and decision-making. These dialogues often involve scaffolding -- the process by which an expert supports a novice's thinking through questions, feedback, and step-by-step guidance. However, such data are scarce due to privacy concerns in recording and the vulnerability inherent in help-seeking. We present SimInstruct, a scalable, expert-in-the-loop tool for collecting scaffolding dialogues. Using teaching development coaching as an example domain, SimInstruct simulates novice instructors via LLMs, varying their teaching challenges and LLM's persona traits, while human experts provide multi-turn feedback, reasoning, and instructional support. This design enables the creation of realistic, pedagogically rich dialogues without requiring real novice participants. Our results reveal that persona traits, such as extroversion and introversion, meaningfully influence how experts engage. Compared to real mentoring recordings, SimInstruct dialogues demonstrate comparable pedagogical relevance and cognitive depth. Experts also reported the process as engaging and reflective, improving both data quality and their own professional insight. We further fine-tuned a LLaMA model to be an expert model using the augmented dataset, which outperformed GPT-4o in instructional quality. Our analysis highlights GPT-4o's limitations in weak reflective questioning, overuse of generic praise, a condescending tone, and a tendency to overwhelm novices with excessive suggestions.

Previous work has demonstrated that AI methods for analysing scientific literature benefit significantly from annotating sentences in papers according to their rhetorical roles, such as research gaps, results, limitations, extensions of existing methodologies, and others. Such representations also have the potential to support the development of a new generation of systems capable of producing high-quality literature reviews. However, achieving this goal requires the definition of a relevant annotation schema and effective strategies for large-scale annotation of the literature. This paper addresses these challenges by 1) introducing a novel annotation schema specifically designed to support literature review generation and 2) conducting a comprehensive evaluation of a wide range of state-of-the-art large language models (LLMs) in classifying rhetorical roles according to this schema. To this end, we also present Sci-Sentence, a novel multidisciplinary benchmark comprising 700 sentences manually annotated by domain experts and 2,240 sentences automatically labelled using LLMs. We evaluate 37 LLMs on this benchmark, spanning diverse model families and sizes, using both zero-shot learning and fine-tuning approaches. The experiments yield several novel insights that advance the state of the art in this challenging domain. First, the current generation of LLMs performs remarkably well on this task when fine-tuned on high-quality data, achieving performance levels above 96\% F1. Second, while large proprietary models like GPT-4o achieve the best results, some lightweight open-source alternatives also demonstrate excellent performance. Finally, enriching the training data with semi-synthetic examples generated by LLMs proves beneficial, enabling small encoders to achieve robust results and significantly enhancing the performance of several open decoder models.

Interactive multimodal agents must convert raw visual observations into coherent sequences of language-conditioned actions -- a capability that current vision-language models (VLMs) still lack. Earlier reinforcement-learning (RL) efforts could, in principle, endow VLMs with such skills, but they have seldom tested whether the learned behaviours generalize beyond their training simulators, and they depend either on brittle hyperparameter tuning or on dense-reward environments with low state variability. We introduce Vision-Language Decoupled Actor-Critic (VL-DAC), a lightweight, hyperparameter-free RL algorithm. VL-DAC applies PPO updates to action tokens while learning value only at the environment-step level: an arrangement, to our knowledge, not previously explored for large VLMs or LLMs. This simple decoupling removes unstable weighting terms and yields faster, more reliable convergence. Training a single VLM with VL-DAC in one inexpensive simulator at a time (MiniWorld, Gym-Cards, ALFWorld, or WebShop) already produces policies that generalize widely: +50\% relative on BALROG (game-centric agentic control), +5\% relative on the hardest part of VSI-Bench (spatial planning), and +2\% on VisualWebBench (web navigation), all without degrading general image understanding accuracy. These results provide the first evidence that a simple RL algorithm can train VLMs entirely in cheap synthetic worlds while delivering measurable gains on real-image agentic, spatial-reasoning, and web-navigation benchmarks.

This paper presents a comprehensive comparative survey of TensorFlow and PyTorch, the two leading deep learning frameworks, focusing on their usability, performance, and deployment trade-offs. We review each framework's programming paradigm and developer experience, contrasting TensorFlow's graph-based (now optionally eager) approach with PyTorch's dynamic, Pythonic style. We then compare model training speeds and inference performance across multiple tasks and data regimes, drawing on recent benchmarks and studies. Deployment flexibility is examined in depth - from TensorFlow's mature ecosystem (TensorFlow Lite for mobile/embedded, TensorFlow Serving, and JavaScript support) to PyTorch's newer production tools (TorchScript compilation, ONNX export, and TorchServe). We also survey ecosystem and community support, including library integrations, industry adoption, and research trends (e.g., PyTorch's dominance in recent research publications versus TensorFlow's broader tooling in enterprise). Applications in computer vision, natural language processing, and other domains are discussed to illustrate how each framework is used in practice. Finally, we outline future directions and open challenges in deep learning framework design, such as unifying eager and graph execution, improving cross-framework interoperability, and integrating compiler optimizations (XLA, JIT) for improved speed. Our findings indicate that while both frameworks are highly capable for state-of-the-art deep learning, they exhibit distinct trade-offs: PyTorch offers simplicity and flexibility favored in research, whereas TensorFlow provides a fuller production-ready ecosystem - understanding these trade-offs is key for practitioners selecting the appropriate tool. We include charts, code snippets, and more than 20 references to academic papers and official documentation to support this comparative analysis