In this work, we compile $\textbf{$\texttt{DroidCollection}$}$, the most extensive open data suite for training and evaluating machine-generated code detectors, comprising over a million code samples, seven programming languages, outputs from 43 coding models, and over three real-world coding domains. Alongside fully AI-generated samples, our collection includes human-AI co-authored code, as well as adversarial samples explicitly crafted to evade detection. Subsequently, we develop $\textbf{$\texttt{DroidDetect}$}$, a suite of encoder-only detectors trained using a multi-task objective over $\texttt{DroidCollection}$. Our experiments show that existing detectors' performance fails to generalise to diverse coding domains and programming languages outside of their narrow training data. Additionally, we demonstrate that while most detectors are easily compromised by humanising the output distributions using superficial prompting and alignment approaches, this problem can be easily amended by training on a small amount of adversarial data. Finally, we demonstrate the effectiveness of metric learning and uncertainty-based resampling as means to enhance detector training on possibly noisy distributions.

Large language models (LLMs) are beginning to reshape how chemists plan and run reactions in organic synthesis. Trained on millions of reported transformations, these text-based models can propose synthetic routes, forecast reaction outcomes and even instruct robots that execute experiments without human supervision. Here we survey the milestones that turned LLMs from speculative tools into practical lab partners. We show how coupling LLMs with graph neural networks, quantum calculations and real-time spectroscopy shrinks discovery cycles and supports greener, data-driven chemistry. We discuss limitations, including biased datasets, opaque reasoning and the need for safety gates that prevent unintentional hazards. Finally, we outline community initiatives open benchmarks, federated learning and explainable interfaces that aim to democratize access while keeping humans firmly in control. These advances chart a path towards rapid, reliable and inclusive molecular innovation powered by artificial intelligence and automation.

The learning rate is a crucial hyperparameter in deep learning, with its ideal value depending on the problem and potentially changing during training. In this paper, we investigate the practical utility of adaptive learning rate mechanisms that adjust step sizes dynamically in response to the loss landscape. We revisit a cumulative path-based adaptation scheme proposed in 2017, which adjusts the learning rate based on the discrepancy between the observed path length, computed as a time-discounted sum of normalized gradient steps, and the expected length of a random walk. While the original approach offers a compelling intuition, we show that its adaptation mechanism for Adam is conceptually inconsistent due to the optimizer's internal preconditioning. We propose a corrected variant that better reflects Adam's update dynamics. To assess the practical value of online learning rate adaptation, we benchmark SGD and Adam, with and without cumulative adaptation, and compare them to a recent alternative method. Our results aim to clarify when and why such adaptive strategies offer practical benefits.

Recent studies have shown that deep neural networks (DNNs) are vulnerable to backdoor attacks, where a designed trigger is injected into the dataset, causing erroneous predictions when activated. In this paper, we propose a novel defense mechanism, Non-target label Training and Mutual Learning (NT-ML), which can successfully restore the poisoned model under advanced backdoor attacks. NT aims to reduce the harm of poisoned data by retraining the model with the outputs of the standard training. At this stage, a teacher model with high accuracy on clean data and a student model with higher confidence in correct prediction on poisoned data are obtained. Then, the teacher and student can learn the strengths from each other through ML to obtain a purified student model. Extensive experiments show that NT-ML can effectively defend against 6 backdoor attacks with a small number of clean samples, and outperforms 5 state-of-the-art backdoor defenses.

There has been rapid development in generative AI tools across the education sector, which in turn is leading to increased adoption by teachers. However, this raises concerns regarding the safety and age-appropriateness of the AI-generated content that is being created for use in classrooms. This paper explores Oak National Academy's approach to addressing these concerns within the development of the UK Government's first publicly available generative AI tool - our AI-powered lesson planning assistant (Aila). Aila is intended to support teachers planning national curriculum-aligned lessons that are appropriate for pupils aged 5-16 years. To mitigate safety risks associated with AI-generated content we have implemented four key safety guardrails: (1) prompt engineering to ensure AI outputs are generated within pedagogically sound and curriculum-aligned parameters; (2) input threat detection to mitigate attacks; (3) an Independent Asynchronous Content Moderation Agent (IACMA) to assess outputs against predefined safety categories; and (4) taking a human-in-the-loop approach, to encourage teachers to review generated content before it is used in the classroom. Through our on-going evaluation of these safety guardrails we have identified several challenges and opportunities to take into account when implementing and testing safety guardrails. This paper highlights ways to build more effective safety guardrails in generative AI education tools including the on-going iteration and refinement of guardrails, as well as enabling cross-sector collaboration through sharing both open-source code/datasets and learnings.

Semi-supervised continual learning (SSCL) seeks to leverage both labeled and unlabeled data in a sequential learning setup, aiming to reduce annotation costs while managing continual data arrival. SSCL introduces complex challenges, including ensuring effective unlabeled learning (UL), while balancing memory stability (MS) and learning plasticity (LP). Previous SSCL efforts have typically focused on isolated aspects of the three, while this work presents USP, a divide-and-conquer framework designed to synergistically enhance these three aspects: (1) Feature Space Reservation (FSR) strategy for LP, which constructs reserved feature locations for future classes by shaping old classes into an equiangular tight frame; (2) Divide-and-Conquer Pseudo-labeling (DCP) approach for UL, which assigns reliable pseudo-labels across both high- and low-confidence unlabeled data; and (3) Class-mean-anchored Unlabeled Distillation (CUD) for MS, which reuses DCP's outputs to anchor unlabeled data to stable class means for distillation to prevent forgetting. Comprehensive evaluations show USP outperforms prior SSCL methods, with gains up to 5.94% in the last accuracy, validating its effectiveness. The code is available at https://github.com/NJUyued/USP4SSCL.

Traditional parameter-efficient fine-tuning (PEFT) methods such as LoRA are tightly coupled with the base model architecture, which constrains their applicability across heterogeneous pretrained large language models (LLMs). To address this limitation, we introduce Cross-LoRA, a data-free framework for transferring LoRA modules between diverse base models without requiring additional training data. Cross-LoRA consists of two key components: (a) LoRA-Align, which performs subspace alignment between source and target base models through rank-truncated singular value decomposition (SVD) and Frobenius-optimal linear transformation, ensuring compatibility under dimension mismatch; and (b) LoRA-Shift, which applies the aligned subspaces to project source LoRA weight updates into the target model parameter space. Both components are data-free, training-free, and enable lightweight adaptation on a commodity GPU in 20 minutes. Experiments on ARCs, OBQA and HellaSwag show that Cross-LoRA achieves relative gains of up to 5.26% over base models . Across other commonsense reasoning benchmarks, Cross-LoRA maintains performance comparable to that of directly trained LoRA adapters.

Generative AI is no longer a peripheral tool in higher education. It is rapidly evolving into a general-purpose infrastructure that reshapes how knowledge is generated, mediated, and validated. This paper presents findings from a controlled experiment evaluating a Socratic AI Tutor, a large language model designed to scaffold student research question development through structured dialogue grounded in constructivist theory. Conducted with 65 pre-service teacher students in Germany, the study compares interaction with the Socratic Tutor to engagement with an uninstructed AI chatbot. Students using the Socratic Tutor reported significantly greater support for critical, independent, and reflective thinking, suggesting that dialogic AI can stimulate metacognitive engagement and challenging recent narratives of de-skilling due to generative AI usage. These findings serve as a proof of concept for a broader pedagogical shift: the use of multi-agent systems (MAS) composed of specialised AI agents. To conceptualise this, we introduce the notion of orchestrated MAS, modular, pedagogically aligned agent constellations, curated by educators, that support diverse learning trajectories through differentiated roles and coordinated interaction. To anchor this shift, we propose an adapted offer-and-use model, in which students appropriate instructional offers from these agents. Beyond technical feasibility, we examine system-level implications for higher education institutions and students, including funding necessities, changes to faculty roles, curriculars, competencies and assessment practices. We conclude with a comparative cost-effectiveness analysis highlighting the scalability of such systems. In sum, this study contributes both empirical evidence and a conceptual roadmap for hybrid learning ecosystems that embed human-AI co-agency and pedagogical alignment.

Reinforcement Learning with Verifiable Rewards (RLVR) improves instruction following capabilities of large language models (LLMs), but suffers from training inefficiency due to inadequate difficulty assessment. Moreover, RLVR is prone to over-optimization, where LLMs exploit verification shortcuts without aligning to the actual intent of user instructions. We introduce Instruction Following Decorator (IFDecorator}, a framework that wraps RLVR training into a robust and sample-efficient pipeline. It consists of three components: (1) a cooperative-adversarial data flywheel that co-evolves instructions and hybrid verifications, generating progressively more challenging instruction-verification pairs; (2) IntentCheck, a bypass module enforcing intent alignment; and (3) trip wires, a diagnostic mechanism that detects reward hacking via trap instructions, which trigger and capture shortcut exploitation behaviors. Our Qwen2.5-32B-Instruct-IFDecorator achieves 87.43% accuracy on IFEval, outperforming larger proprietary models such as GPT-4o. Additionally, we demonstrate substantial improvements on FollowBench while preserving general capabilities. Our trip wires show significant reductions in reward hacking rates. We will release models, code, and data for future research.

Nonverbal communication (NVC) plays an integral role in human language, but studying NVC in general is challenging because of its broad scope and high variance in interpretation among individuals and cultures. However, mime -- the theatrical technique of suggesting intent using only gesture, expression, and movement -- is a subset of NVC that consists of explicit and embodied actions with much lower human interpretation variance. We argue that a solid understanding of mimed actions is a crucial prerequisite for vision-language models capable of interpreting and commanding more subtle aspects of NVC. Hence, we propose Mime Identification Multimodal Evaluation (MIME), a novel video-based question answering benchmark comprising of 86 mimed actions. Constructed with motion capture data, MIME consists of variations of each action with perturbations applied to the character, background, and viewpoint for evaluating recognition robustness. We find that both open-weight and API-based vision-language models perform significantly worse than humans on MIME, motivating the need for increased research for instilling more robust understanding of human gestures.