There is a growing need to evaluate Large Language Models (LLMs) on complex, high-impact, real-world tasks to assess their true readiness as reasoning agents. To address this gap, we introduce AgentCaster, a contamination-free framework employing multimodal LLMs end-to-end for the challenging, long-horizon task of tornado forecasting. Within AgentCaster, models interpret heterogeneous spatiotemporal data from a high-resolution convection-allowing forecast archive. We assess model performance over a 40-day period featuring diverse historical data, spanning several major tornado outbreaks and including over 500 tornado reports. Each day, models query interactively from a pool of 3,625 forecast maps and 40,125 forecast soundings for a forecast horizon of 12-36 hours. Probabilistic tornado-risk polygon predictions are verified against ground truths derived from geometric comparisons across disjoint risk bands in projected coordinate space. To quantify accuracy, we propose domain-specific TornadoBench and TornadoHallucination metrics, with TornadoBench highly challenging for both LLMs and domain expert human forecasters. Notably, human experts significantly outperform state-of-the-art models, which demonstrate a strong tendency to hallucinate and overpredict risk intensity, struggle with precise geographic placement, and exhibit poor spatiotemporal reasoning in complex, dynamically evolving systems. AgentCaster aims to advance research on improving LLM agents for challenging reasoning tasks in critical domains.

Large Language Models (LLMs) are increasingly deployed in multi-agent systems, where effective inter-model communication is crucial. Existing communication protocols either rely on natural language, incurring high inference costs and information loss, or on hidden states, which suffer from information concentration bias and inefficiency. To address these limitations, we propose KVComm, a novel communication framework that enables efficient communication between LLMs through selective sharing of KV pairs. KVComm leverages the rich information encoded in the KV pairs while avoiding the pitfalls of hidden states. We introduce a KV layer-wise selection strategy based on attention importance scores with a Gaussian prior to identify the most informative KV pairs for communication. Extensive experiments across diverse tasks and model pairs demonstrate that KVComm achieves comparable performance to the upper-bound method, which directly merges inputs to one model without any communication, while transmitting as few as 30% of layers' KV pairs. Our study highlights the potential of KV pairs as an effective medium for inter-LLM communication, paving the way for scalable and efficient multi-agent systems. Large Language Models (LLMs) have catalyzed a paradigm shift from isolated model capabilities towards collaborative multi-agent systems (Guo et al., 2024; Tran et al., 2025). CAMEL (Li et al., 2023), AutoGen (Wu et al., 2024), and ChatDev (Qian et al., 2023) have demonstrated the potential of LLMs to collaborate effectively in multi-agent systems, achieving impressive results in various tasks. These systems leverage the strengths of individual LLMs and enable them to work together to solve complex problems that are beyond the capabilities of a single model (Y ang et al., 2024a).

We start with reviewing normative actions of international organizations and concerns expressed about the current technical landscape. Then we proceed with proposing a framework that comprises five key dimensions relating to the main challenges relating to AI in higher education institutions, followed by five key strategic actions that the main stakeholders need to take in order to address the current developments . W e map these actions to the main stakeholders of higher education and propose a deployment plan . This defines a framework along the dimensions: C hallenges, Actions, Stakeholders, Deployment CASD . Examples of AI specific actions at the institutional and individu al course level are also provided and discussed.

Transformer models have become the dominant backbone for sequence modeling, leveraging self-attention to produce contextualized token representations. These are typically aggregated into fixed-size vectors via pooling operations for downstream tasks. While much of the literature has focused on attention mechanisms, the role of pooling remains underexplored despite its critical impact on model behavior. In this paper, we introduce a theoretical framework that rigorously characterizes the expressivity of Transformer-based models equipped with widely used pooling methods by deriving closed-form bounds on their representational capacity and the ability to distinguish similar inputs. Our analysis extends to different variations of attention formulations, demonstrating that these bounds hold across diverse architectural variants. We empirically evaluate pooling strategies across tasks requiring both global and local contextual understanding, spanning three major modalities: computer vision, natural language processing, and time-series analysis. Results reveal consistent trends in how pooling choices affect accuracy, sensitivity, and optimization behavior. Our findings unify theoretical and empirical perspectives, providing practical guidance for selecting or designing pooling mechanisms suited to specific tasks. This work positions pooling as a key architectural component in Transformer models and lays the foundation for more principled model design beyond attention alone.

Early detection of Alzheimer's Dementia (AD) and Mild Cognitive Impairment (MCI) is critical for timely intervention, yet current diagnostic approaches remain resource-intensive and invasive. Speech, encompassing both acoustic and linguistic dimensions, offers a promising non-invasive biomarker for cognitive decline. In this study, we present a machine learning framework for the PROCESS Challenge, leveraging both audio embeddings and linguistic features derived from spontaneous speech recordings. Audio representations were extracted using Whisper embeddings from the Cookie Theft description task, while linguistic features-spanning pronoun usage, syntactic complexity, filler words, and clause structure-were obtained from transcriptions across Semantic Fluency, Phonemic Fluency, and Cookie Theft picture description. Classification models aimed to distinguish between Healthy Controls (HC), MCI, and AD participants, while regression models predicted Mini-Mental State Examination (MMSE) scores. Results demonstrated that voted ensemble models trained on concatenated linguistic features achieved the best classification performance (F1 = 0.497), while Whisper embedding-based ensemble regressors yielded the lowest MMSE prediction error (RMSE = 2.843). Comparative evaluation within the PROCESS Challenge placed our models among the top submissions in regression task, and mid-range for classification, highlighting the complementary strengths of linguistic and audio embeddings. These findings reinforce the potential of multimodal speech-based approaches for scalable, non-invasive cognitive assessment and underline the importance of integrating task-specific linguistic and acoustic markers in dementia detection.

Deep learning (DL) schedulers are pivotal in optimizing resource allocation in GPU clusters, but operate with a critical limitation: they are largely blind to the semantic context of the jobs they manage. This forces them to rely on limited metadata, leading to high profiling overhead, unreliable duration estimation, inadequate failure handling, and poor observability. To this end, we propose SchedMate, a framework that bridges this semantic gap by systematically extracting deep insights from overlooked, unstructured data sources: source code, runtime logs, and historical jobs. SchedMate enhances existing schedulers non-intrusively through three LLM-based components. Our implementation integrates seamlessly with existing deep learning schedulers. Evaluations on a 128-GPU physical cluster and extensive simulations on production traces show SchedMate reduces average job completion times by up to 1.91x, substantially enhancing the scheduling performance, demonstrating the critical role of semantic-awareness in modern DL scheduling.

Peptide drugs incorporating non-standard amino acids (NSAAs) offer improved binding affinity and improved pharmacological properties. However, existing peptide design methods are limited to standard amino acids, leaving NSAA-aware design largely unexplored. We introduce NS-Pep, a unified framework for co-designing peptide sequences and structures with NSAAs. The main challenge is that NSAAs are extremely underrepresented-even the most frequent one, SEP, accounts for less than 0.4% of residues-resulting in a severe long-tailed distribution. To improve generalization to rare amino acids, we propose Residue Frequency-Guided Modification (RFGM), which mitigates over-penalization through frequency-aware logit calibration, supported by both theoretical and empirical analysis. Furthermore, we identify that insufficient side-chain modeling limits geometric representation of NSAAs. To address this, we introduce Progressive Side-chain Perception (PSP) for coarse-to-fine torsion and location prediction, and Interaction-Aware Weighting (IAW) to emphasize pocket-proximal residues. Moreover, NS-Pep generalizes naturally to the peptide folding task with NSAAs, addressing a major limitation of current tools. Experiments show that NS-Pep improves sequence recovery rate and binding affinity by 6.23% and 5.12%, respectively, and outperforms AlphaFold3 by 17.76% in peptide folding success rate.

The reconstruction of a frequency with minimal delay from a sinusoidal signal is a common task in several fields; for example Ring Laser Gyroscopes, since their output signal is a beat frequency. While conventional methods require several seconds of data, we present a neural network approach capable of reconstructing frequencies of several hundred Hertz within approximately 10 milliseconds. This enables rapid trigger generation. The method outperforms standard Fourier-based techniques, improving frequency estimation precision by a factor of 2 in the operational range of GINGERINO, our Ring Laser Gyroscope.\\ In addition to fast frequency estimation, we introduce an automated classification framework to identify physical disturbances in the signal, such as laser instabilities and seismic events, achieving accuracy rates between 99\% and 100\% on independent test datasets for the seismic class. These results mark a step forward in integrating artificial intelligence into signal analysis for geophysical applications.

A significant portion of real-world data is inherently represented as textual graphs, and integrating these graphs into large language models (LLMs) is promising to enable complex graph-based question answering. However, a key challenge in LLM-based textual graph QA systems lies in graph retrieval, i.e., how to retrieve relevant content from large graphs that is sufficiently informative while remaining compact for the LLM context. Existing retrievers suffer from poor performance since they either rely on shallow embedding similarity or employ interactive retrieving policies that demand excessive data labeling and training cost. To address these issues, we present Graph-$S^3$, an agentic textual graph reasoning framework that employs an LLM-based retriever trained with synthetic stepwise supervision. Instead of rewarding the agent based on the final answers, which may lead to sparse and unstable training signals, we propose to closely evaluate each step of the retriever based on offline-extracted golden subgraphs. Our main techniques include a data synthesis pipeline to extract the golden subgraphs for reward generation and a two-stage training scheme to learn the interactive graph exploration policy based on the synthesized rewards. Based on extensive experiments on three common datasets in comparison with seven strong baselines, our approach achieves an average improvement of 8.1\% in accuracy and 9.7\% in F$_1$ score. The advantage is even higher in more complicated multi-hop reasoning tasks. Our code will be open-sourced.

Deep neural networks (NNs) for computer vision are vulnerable to adversarial attacks, i.e., miniscule malicious changes to inputs may induce unintuitive outputs. One key approach to verify and mitigate such robustness issues is to falsify expected output behavior. This allows, e.g., to locally proof security, or to (re)train NNs on obtained adversarial input examples. Due to the black-box nature of NNs, current attacks only falsify a class of the final output, such as flipping from $\texttt{stop_sign}$ to $\neg\texttt{stop_sign}$. In this short position paper we generalize this to search for generally illogical behavior, as considered in NN verification: falsify constraints (concept-based properties) involving further human-interpretable concepts, like $\texttt{red}\wedge\texttt{octogonal}\rightarrow\texttt{stop_sign}$. For this, an easy implementation of concept-based properties on already trained NNs is proposed using techniques from explainable artificial intelligence. Further, we sketch the theoretical proof that attacks on concept-based properties are expected to have a reduced search space compared to simple class falsification, whilst arguably be more aligned with intuitive robustness targets. As an outlook to this work in progress we hypothesize that this approach has potential to efficiently and simultaneously improve logical compliance and robustness.