Engineering sustainable and equitable healthcare requires medical language models that do not change clinically correct diagnoses when presented with non-decisive cultural information. We introduce a counterfactual benchmark that expands 150 MedQA test items into 1650 variants by inserting culture-related (i) identifier tokens, (ii) contextual cues, or (iii) their combination for three groups (Indigenous Canadian, Middle-Eastern Muslim, Southeast Asian), plus a length-matched neutral control, where a clinician verified that the gold answer remains invariant in all variants. We evaluate GPT-5.2, Llama-3.1-8B, DeepSeek-R1, and MedGemma (4B/27B) under option-only and brief-explanation prompting. Across models, cultural cues significantly affect accuracy (Cochran's Q, $p<10^-14$), with the largest degradation when identifier and context co-occur (up to 3-7 percentage points under option-only prompting), while neutral edits produce smaller, non-systematic changes. A human-validated rubric ($κ=0.76$) applied via an LLM-as-judge shows that more than half of culturally grounded explanations end in an incorrect answer, linking culture-referential reasoning to diagnostic failure. We release prompts and augmentations to support evaluation and mitigation of culturally induced diagnostic errors.

Heat-pipe microreactors (HPMRs) are compact and transportable nuclear power systems exhibiting inherent safety, well-suited for deployment in remote regions where access is limited and reliance on costly fossil fuels is prevalent. In prior work, we developed a design optimization framework that incorporates techno-economic considerations through surrogate modeling and reinforcement learning (RL)-based optimization, focusing solely on minimizing the levelized cost of electricity (LCOE) by using a bottom-up cost estimation approach. In this study, we extend that framework to a multi-objective optimization that uses the Pareto Envelope Augmented with Reinforcement Learning (PEARL) algorithm. The objectives include minimizing both the rod-integrated peaking factor ($F_{Δh}$) and LCOE -- subject to safety and operational constraints. We evaluate three cost scenarios: (1) a high-cost axial and drum reflectors, (2) a low-cost axial reflector, and (3) low-cost axial and drum reflectors. Our findings indicate that reducing the solid moderator radius, pin pitch, and drum coating angle -- all while increasing the fuel height -- effectively lowers $F_{Δh}$. Across all three scenarios, four key strategies consistently emerged for optimizing LCOE: (1) minimizing the axial reflector contribution when costly, (2) reducing control drum reliance, (3) substituting expensive tri-structural isotropic (TRISO) fuel with axial reflector material priced at the level of graphite, and (4) maximizing fuel burnup. While PEARL demonstrates promise in navigating trade-offs across diverse design scenarios, discrepancies between surrogate model predictions and full-order simulations remain. Further improvements are anticipated through constraint relaxation and surrogate development, constituting an ongoing area of investigation.

Contrastive Language-Image Pre-training (CLIP) has become a cornerstone in vision-language representation learning, powering diverse downstream tasks and serving as the default vision backbone in multimodal large language models (MLLMs). Despite its success, CLIP's dense and opaque latent representations pose significant interpretability challenges. A common assumption is that interpretability and performance are in tension: enforcing sparsity during training degrades accuracy, motivating recent post-hoc approaches such as Sparse Autoencoders (SAEs). However, these post-hoc approaches often suffer from degraded downstream performance and loss of CLIP's inherent multimodal capabilities, with most learned features remaining unimodal. We propose a simple yet effective approach that integrates sparsity directly into CLIP training, yielding representations that are both interpretable and performant. Compared to SAEs, our Sparse CLIP representations preserve strong downstream task performance, achieve superior interpretability, and retain multimodal capabilities. We show that multimodal sparse features enable straightforward semantic concept alignment and reveal training dynamics of how cross-modal knowledge emerges. Finally, as a proof of concept, we train a vision-language model on sparse CLIP representations that enables interpretable, vision-based steering capabilities. Our findings challenge conventional wisdom that interpretability requires sacrificing accuracy and demonstrate that interpretability and performance can be co-optimized, offering a promising design principle for future models.

We address the challenge of training Vision Transformers (ViTs) when labeled data is scarce but unlabeled data is abundant. We propose Semi-Supervised Masked Autoencoder (SSMAE), a framework that jointly optimizes masked image reconstruction and classification using both unlabeled and labeled samples with dynamically selected pseudo-labels. SSMAE introduces a validation-driven gating mechanism that activates pseudo-labeling only after the model achieves reliable, high-confidence predictions that are consistent across both weakly and strongly augmented views of the same image, reducing confirmation bias. On CIFAR-10 and CIFAR-100, SSMAE consistently outperforms supervised ViT and fine-tuned MAE, with the largest gains in low-label regimes (+9.24% over ViT on CIFAR-10 with 10% labels). Our results demonstrate that when pseudo-labels are introduced is as important as how they are generated for data-efficient transformer training. Codes are available at https://github.com/atik666/ssmae.

Open-vocabulary semantic segmentation aims to assign labels to every pixel in an image based on text labels. Existing approaches typically utilize vision-language models (VLMs), such as CLIP, for dense prediction. However, VLMs, pre-trained on image-text pairs, are biased toward salient, object-centric regions and exhibit two critical limitations when adapted to segmentation: (i) Foreground Bias, which tends to ignore background regions, and (ii) Limited Spatial Localization, resulting in blurred object boundaries. To address these limitations, we introduce DiSa, a novel saliency-aware foreground-background disentangled framework. By explicitly incorporating saliency cues in our designed Saliency-aware Disentanglement Module (SDM), DiSa separately models foreground and background ensemble features in a divide-and-conquer manner. Additionally, we propose a Hierarchical Refinement Module (HRM) that leverages pixel-wise spatial contexts and enables channel-wise feature refinement through multi-level updates. Extensive experiments on six benchmarks demonstrate that DiSa consistently outperforms state-of-the-art methods.

The rise of chronic diseases related to diet, such as obesity and diabetes, emphasizes the need for accurate monitoring of food intake. While AI-driven dietary assessment has made strides in recent years, the ill-posed nature of recovering size (portion) information from monocular images for accurate estimation of ``how much did you eat?'' is a pressing challenge. Some 3D reconstruction methods have achieved impressive geometric reconstruction but fail to recover the crucial real-world scale of the reconstructed object, limiting its usage in precision nutrition. In this paper, we bridge the gap between 3D computer vision and digital health by proposing a method that recovers a true-to-scale 3D reconstructed object from a monocular image. Our approach leverages rich visual features extracted from models trained on large-scale datasets to estimate the scale of the reconstructed object. This learned scale enables us to convert single-view 3D reconstructions into true-to-life, physically meaningful models. Extensive experiments and ablation studies on two publicly available datasets show that our method consistently outperforms existing techniques, achieving nearly a 30% reduction in mean absolute volume-estimation error, showcasing its potential to enhance the domain of precision nutrition. Code: https://gitlab.com/viper-purdue/size-matters

Metastatic castration-resistant prostate cancer (mCRPC) is a highly aggressive disease with poor prognosis and heterogeneous treatment response. In this work, we developed and externally validated a visit-level 180-day mortality risk model using longitudinal data from two Phase III cohorts (n=526 and n=640). Only visits with observable 180-day outcomes were labeled; right-censored cases were excluded from analysis. We compared five candidate architectures: Long Short-Term Memory, Gated Recurrent Unit (GRU), Cox Proportional Hazards, Random Survival Forest (RSF), and Logistic Regression. For each dataset, we selected the smallest risk-threshold that achieved an 85% sensitivity floor. The GRU and RSF models showed high discrimination capabilities initially (C-index: 87% for both). In external validation, the GRU obtained a higher calibration (slope: 0.93; intercept: 0.07) and achieved an PR-AUC of 0.87. Clinical impact analysis showed a median time-in-warning of 151.0 days for true positives (59.0 days for false positives) and 18.3 alerts per 100 patient-visits. Given late-stage frailty or cachexia and hemodynamic instability, permutation importance ranked BMI and systolic blood pressure as the strongest associations. These results suggest that longitudinal routine clinical markers can estimate short-horizon mortality risk in mCRPC and support proactive care planning over a multi-month window.

Standard Bayesian Optimization (BO) assumes uniform smoothness across the search space an assumption violated in multi-regime problems such as molecular conformation search through distinct energy basins or drug discovery across heterogeneous molecular scaffolds. A single GP either oversmooths sharp transitions or hallucinates noise in smooth regions, yielding miscalibrated uncertainty. We propose RAMBO, a Dirichlet Process Mixture of Gaussian Processes that automatically discovers latent regimes during optimization, each modeled by an independent GP with locally-optimized hyperparameters. We derive collapsed Gibbs sampling that analytically marginalizes latent functions for efficient inference, and introduce adaptive concentration parameter scheduling for coarse-to-fine regime discovery. Our acquisition functions decompose uncertainty into intra-regime and inter-regime components. Experiments on synthetic benchmarks and real-world applications, including molecular conformer optimization, virtual screening for drug discovery, and fusion reactor design, demonstrate consistent improvements over state-of-the-art baselines on multi-regime objectives.

Despite the ubiquity of tabular data in high-stakes domains, traditional deep learning architectures often struggle to match the performance of gradient-boosted decision trees while maintaining scientific interpretability. Standard neural networks typically treat features as independent entities, failing to exploit the inherent manifold structural dependencies that define tabular distributions. We propose Structural Compositional Function Networks (StructuralCFN), a novel architecture that imposes a Relation-Aware Inductive Bias via a differentiable structural prior. StructuralCFN explicitly models each feature as a mathematical composition of its counterparts through Differentiable Adaptive Gating, which automatically discovers the optimal activation physics (e.g., attention-style filtering vs. inhibitory polarity) for each relationship. Our framework enables Structured Knowledge Integration, allowing domain-specific relational priors to be injected directly into the architecture to guide discovery. We evaluate StructuralCFN across a rigorous 10-fold cross-validation suite on 18 benchmarks, demonstrating statistically significant improvements (p < 0.05) on scientific and clinical datasets (e.g., Blood Transfusion, Ozone, WDBC). Furthermore, StructuralCFN provides Intrinsic Symbolic Interpretability: it recovers the governing "laws" of the data manifold as human-readable mathematical expressions while maintaining a compact parameter footprint (300--2,500 parameters) that is over an order of magnitude (10x--20x) smaller than standard deep baselines.

Health influencers play a growing role in shaping public beliefs, yet their content is often conveyed through conversational narratives and rhetorical strategies rather than explicit factual claims. As a result, claim-centric verification methods struggle to capture the pragmatic meaning of influencer discourse. In this paper, we propose TAIGR (Takeaway Argumentation Inference with Grounded References), a structured framework designed to analyze influencer discourse, which operates in three stages: (1) identifying the core influencer recommendation--takeaway; (2) constructing an argumentation graph that captures influencer justification for the takeaway; (3) performing factor graph-based probabilistic inference to validate the takeaway. We evaluate TAIGR on a content validation task over influencer video transcripts on health, showing that accurate validation requires modeling the discourse's pragmatic and argumentative structure rather than treating transcripts as flat collections of claims.

The Linear Representation Hypothesis asserts that the embeddings learned by neural networks can be understood as linear combinations of features corresponding to high-level concepts. Based on this ansatz, sparse autoencoders (SAEs) have recently become a popular method for decomposing embeddings into a sparse combination of linear directions, which have been shown empirically to often correspond to human-interpretable semantics. However, recent attempts to apply SAEs to multimodal embedding spaces (such as the popular CLIP embeddings for image/text data) have found that SAEs often learn "split dictionaries", where most of the learned sparse features are essentially unimodal, active only for data of a single modality. In this work, we study how to effectively adapt SAEs for the setting of multimodal embeddings while ensuring multimodal alignment. We first argue that the existence of a split dictionary decomposition on an aligned embedding space implies the existence of a non-split dictionary with improved modality alignment. Then, we propose a new SAE-based approach to multimodal embedding decomposition using cross-modal random masking and group-sparse regularization. We apply our method to popular embeddings for image/text (CLIP) and audio/text (CLAP) data and show that, compared to standard SAEs, our approach learns a more multimodal dictionary while reducing the number of dead neurons and improving feature semanticity. We finally demonstrate how this improvement in alignment of concepts between modalities can enable improvements in the interpretability and control of cross-modal tasks.

Natural-language planning often involves vague predicates (e.g., suitable substitute, stable enough) whose satisfaction is inherently graded. Existing category-theoretic planners provide compositional structure and pullback-based hard-constraint verification, but treat applicability as crisp, forcing thresholding that collapses meaningful distinctions and cannot track quality degradation across multi-step plans. We propose Fuzzy Category-theoretic Planning (FCP), which annotates each action (morphism) with a degree in [0,1], composes plan quality via a t-norm Lukasiewicz, and retains crisp executability checks via pullback verification. FCP grounds graded applicability from language using an LLM with k-sample median aggregation and supports meeting-in-the-middle search using residuum-based backward requirements. We evaluate on (i) public PDDL3 preference/oversubscription benchmarks and (ii) RecipeNLG-Subs, a missing-substitute recipe-planning benchmark built from RecipeNLG with substitution candidates from Recipe1MSubs and FoodKG. FCP improves success and reduces hard-constraint violations on RecipeNLG-Subs compared to LLM-only and ReAct-style baselines, while remaining competitive with classical PDDL3 planners.

Inference-time planning with large language models frequently breaks under partial observability: when task-critical preconditions are not specified at query time, models tend to hallucinate missing facts or produce plans that violate hard constraints. We introduce \textbf{Self-Querying Bidirectional Categorical Planning (SQ-BCP)}, which explicitly represents precondition status (\texttt{Sat}/\texttt{Viol}/\texttt{Unk}) and resolves unknowns via (i) targeted self-queries to an oracle/user or (ii) \emph{bridging} hypotheses that establish the missing condition through an additional action. SQ-BCP performs bidirectional search and invokes a pullback-based verifier as a categorical certificate of goal compatibility, while using distance-based scores only for ranking and pruning. We prove that when the verifier succeeds and hard constraints pass deterministic checks, accepted plans are compatible with goal requirements; under bounded branching and finite resolution depth, SQ-BCP finds an accepting plan when one exists. Across WikiHow and RecipeNLG tasks with withheld preconditions, SQ-BCP reduces resource-violation rates to \textbf{14.9\%} and \textbf{5.8\%} (vs.\ \textbf{26.0\%} and \textbf{15.7\%} for the best baseline), while maintaining competitive reference quality.

The rapid progress of large language models has enabled the generation of text that closely resembles human writing, creating challenges for authenticity verification in education, publishing, and digital security. Detecting AI-generated text has therefore become a crucial technical and ethical issue. This paper presents a comprehensive study of AI-generated text detection based on large-scale corpora and novel training strategies. We introduce a 1-billion-token corpus of human-authored texts spanning multiple genres and a 1.9-billion-token corpus of AI-generated texts produced by prompting a variety of LLMs across diverse domains. Using these resources, we develop and evaluate numerous detection models and propose two novel training paradigms: Per LLM and Per LLM family fine-tuning. Across a 100-million-token benchmark covering 21 large language models, our best fine-tuned detector achieves up to $99.6\%$ token-level accuracy, substantially outperforming existing open-source baselines.

Industrial image anomaly detection is a challenging problem owing to extreme class imbalance and the scarcity of labeled defective samples, particularly in few-shot settings. We propose BayPrAnoMeta, a Bayesian generalization of Proto-MAML for few-shot industrial image anomaly detection. Unlike existing Proto-MAML approaches that rely on deterministic class prototypes and distance-based adaptation, BayPrAnoMeta replaces prototypes with task-specific probabilistic normality models and performs inner-loop adaptation via a Bayesian posterior predictive likelihood. We model normal support embeddings with a Normal-Inverse-Wishart (NIW) prior, producing a Student-$t$ predictive distribution that enables uncertainty-aware, heavy-tailed anomaly scoring and is essential for robustness in extreme few-shot settings. We further extend BayPrAnoMeta to a federated meta-learning framework with supervised contrastive regularization for heterogeneous industrial clients and prove convergence to stationary points of the resulting nonconvex objective. Experiments on the MVTec AD benchmark demonstrate consistent and significant AUROC improvements over MAML, Proto-MAML, and PatchCore-based methods in few-shot anomaly detection settings.

In high-dimensional robotic path planning, traditional sampling-based methods often struggle to efficiently identify both feasible and optimal paths in complex, multi-obstacle environments. This challenge is intensified in robotic manipulators, where the risk of kinematic singularities and self-collisions further complicates motion efficiency and safety. To address these issues, we introduce the Just-in-Time Informed Trees (JIT*) algorithm, an enhancement over Effort Informed Trees (EIT*), designed to improve path planning through two core modules: the Just-in-Time module and the Motion Performance module. The Just-in-Time module includes "Just-in-Time Edge," which dynamically refines edge connectivity, and "Just-in-Time Sample," which adjusts sampling density in bottleneck areas to enable faster initial path discovery. The Motion Performance module balances manipulability and trajectory cost through dynamic switching, optimizing motion control while reducing the risk of singularities. Comparative analysis shows that JIT* consistently outperforms traditional sampling-based planners across $\mathbb{R}^4$ to $\mathbb{R}^{16}$ dimensions. Its effectiveness is further demonstrated in single-arm and dual-arm manipulation tasks, with experimental results available in a video at https://youtu.be/nL1BMHpMR7c.

Human-in-the-loop guidance has emerged as an effective approach for enabling faster convergence in online reinforcement learning (RL) of complex real-world manipulation tasks. However, existing human-in-the-loop RL (HiL-RL) frameworks often suffer from low sample efficiency, requiring substantial human interventions to achieve convergence and thereby leading to high labor costs. To address this, we propose a sample-efficient real-world human-in-the-loop RL framework named \method, which requires fewer human intervention by actively selecting informative samples. Specifically, stable reduction of policy entropy enables improved trade-off between exploration and exploitation with higher sample efficiency. We first build influence functions of different samples on the policy entropy, which is efficiently estimated by the covariance of action probabilities and soft advantages of policies. Then we select samples with moderate values of influence functions, where shortcut samples that induce sharp entropy drops and noisy samples with negligible effect are pruned. Extensive experiments on four real-world manipulation tasks demonstrate that \method achieves a 42.1\% higher success rate while requiring 10.1\% fewer human interventions compared to the state-of-the-art HiL-RL method, validating its effectiveness. The project page providing code, videos, and mathematical formulations can be found at https://e2hil.github.io/.

Neuroscience and Artificial Intelligence (AI) have made significant progress in the past few years but have only been loosely inter-connected. Based on a workshop held in August 2025, we identify current and future areas of synergism between these two fields. We focus on the subareas of embodiment, language and communication, robotics, learning in humans and machines and Neuromorphic engineering to take stock of the progress made so far, and possible promising new future avenues. Overall, we advocate for the development of NeuroAI, a type of Neuroscience-informed Artificial Intelligence that, we argue, has the potential for significantly improving the scope and efficiency of AI algorithms while simultaneously changing the way we understand biological neural computations. We include personal statements from several leading researchers on their diverse views of NeuroAI. Two Strength-Weakness-Opportunities-Threat (SWOT) analyses by researchers and trainees are appended that describe the benefits and risks offered by NeuroAI.

Extending the intelligence of sensors to the data-acquisition process - deciding whether to sample or not - can result in transformative energy-efficiency gains. However, making such a decision in a deterministic manner involves risk of losing information. Here we present a sensing paradigm that enables making such a decision in a probabilistic manner. The paradigm takes inspiration from the autonomous nervous system and employs a probabilistic neuron (p-neuron) driven by an analog feature extraction circuit. The response time of the system is on the order of microseconds, over-coming the sub-sampling-rate response time limit and enabling real-time intelligent autonomous activation of data-sampling. Validation experiments on active seismic survey data demonstrate lossless probabilistic data acquisition, with a normalized mean squared error of 0.41%, and 93% saving in the active operation time of the system and the number of generated samples.

Real-time voice agents face a dilemma: end-to-end models often lack deep reasoning, while cascaded pipelines incur high latency by executing ASR, LLM reasoning, and TTS strictly in sequence, unlike human conversation where listeners often start thinking before the speaker finishes. Since cascaded architectures remain the dominant choice for complex tasks, existing cascaded streaming strategies attempt to reduce this latency via mechanical segmentation (e.g., fixed chunks, VAD-based splitting) or speculative generation, but they frequently either break semantic units or waste computation on predictions that must be rolled back. To address these challenges, we propose LTS-VoiceAgent, a Listen-Think-Speak framework that explicitly separates when to think from how to reason incrementally. It features a Dynamic Semantic Trigger to detect meaningful prefixes, and a Dual-Role Stream Orchestrator that coordinates a background Thinker (for state maintenance) and a foreground Speaker (for speculative solving). This parallel design enables "thinking while speaking" without blocking responses. We also introduce a Pause-and-Repair benchmark containing natural disfluencies to stress-test streaming robustness. Experiments across VERA, Spoken-MQA, BigBenchAudio, and our benchmark show that LTS-VoiceAgent achieves a stronger accuracy-latency-efficiency trade-off than serial cascaded baselines and existing streaming strategies.

Symbolic music representation is a fundamental challenge in computational musicology. While grid-based representations effectively preserve pitch-time spatial correspondence, their inherent data sparsity leads to low encoding efficiency. Discrete-event representations achieve compact encoding but fail to adequately capture structural invariance and spatial locality. To address these complementary limitations, we propose Pianoroll-Event, a novel encoding scheme that describes pianoroll representations through events, combining structural properties with encoding efficiency while maintaining temporal dependencies and local spatial patterns. Specifically, we design four complementary event types: Frame Events for temporal boundaries, Gap Events for sparse regions, Pattern Events for note patterns, and Musical Structure Events for musical metadata. Pianoroll-Event strikes an effective balance between sequence length and vocabulary size, improving encoding efficiency by 1.36\times to 7.16\times over representative discrete sequence methods. Experiments across multiple autoregressive architectures show models using our representation consistently outperform baselines in both quantitative and human evaluations.

Automatic Speech Recognition (ASR) offers significant potential to reduce the workload of medical personnel, for example, through the automation of documentation tasks. While numerous benchmarks exist for the English language, specific evaluations for the German-speaking medical context are still lacking, particularly regarding the inclusion of dialects. In this article, we present a curated dataset of simulated doctor-patient conversations and evaluate a total of 29 different ASR models. The test field encompasses both open-weights models from the Whisper, Voxtral, and Wav2Vec2 families as well as commercial state-of-the-art APIs (AssemblyAI, Deepgram). For evaluation, we utilize three different metrics (WER, CER, BLEU) and provide an outlook on qualitative semantic analysis. The results demonstrate significant performance differences between the models: while the best systems already achieve very good Word Error Rates (WER) of partly below 3%, the error rates of other models, especially concerning medical terminology or dialect-influenced variations, are considerably higher.

We study model-agnostic post-hoc calibration methods intended to improve probabilistic predictions in supervised binary classification on real i.i.d. tabular data, with particular emphasis on conformal and Venn-based approaches that provide distribution-free validity guarantees under exchangeability. We benchmark 21 widely used classifiers, including linear models, SVMs, tree ensembles (CatBoost, XGBoost, LightGBM), and modern tabular neural and foundation models, on binary tasks from the TabArena-v0.1 suite using randomized, stratified five-fold cross-validation with a held-out test fold. Five calibrators; Isotonic regression, Platt scaling, Beta calibration, Venn-Abers predictors, and Pearsonify are trained on a separate calibration split and applied to test predictions. Calibration is evaluated using proper scoring rules (log-loss and Brier score) and diagnostic measures (Spiegelhalter's Z, ECE, and ECI), alongside discrimination (AUC-ROC) and standard classification metrics. Across tasks and architectures, Venn-Abers predictors achieve the largest average reductions in log-loss, followed closely by Beta calibration, while Platt scaling exhibits weaker and less consistent effects. Beta calibration improves log-loss most frequently across tasks, whereas Venn-Abers displays fewer instances of extreme degradation and slightly more instances of extreme improvement. Importantly, we find that commonly used calibration procedures, most notably Platt scaling and isotonic regression, can systematically degrade proper scoring performance for strong modern tabular models. Overall classification performance is often preserved, but calibration effects vary substantially across datasets and architectures, and no method dominates uniformly. In expectation, all methods except Pearsonify slightly increase accuracy, but the effect is marginal, with the largest expected gain about 0.008%.

How can populations of learners develop coordinated, diverse behaviors without explicit communication or diversity incentives? We demonstrate that competition alone is sufficient to induce emergent specialization -- learners spontaneously partition into specialists for different environmental regimes through competitive dynamics, consistent with ecological niche theory. We introduce the NichePopulation algorithm, a simple mechanism combining competitive exclusion with niche affinity tracking. Validated across six real-world domains (cryptocurrency trading, commodity prices, weather forecasting, solar irradiance, urban traffic, and air quality), our approach achieves a mean Specialization Index of 0.75 with effect sizes of Cohen's d > 20. Key findings: (1) At lambda=0 (no niche bonus), learners still achieve SI > 0.30, proving specialization is genuinely emergent; (2) Diverse populations outperform homogeneous baselines by +26.5% through method-level division of labor; (3) Our approach outperforms MARL baselines (QMIX, MAPPO, IQL) by 4.3x while being 4x faster.

We study the emergence of multi-step reasoning in deep Transformer language models through a geometric and statistical-physics lens. Treating the hidden-state trajectory as a flow on an implicit Riemannian manifold, we analyze the layerwise covariance spectrum of activations, where $C^{(\ell)}=\mathbb{E}[h^{(\ell)}h^{(\ell)\top}]$, and track deviations from a random-matrix bulk. Across model scales (1.5B--30B), we observe a sharp reduction in effective dimensionality consistent with a phase transition: an order parameter based on sparsity/localization, $Ω(h)=1-\|h\|_1/(\sqrt{d}\|h\|_2)$, exhibits a discontinuity near a critical normalized depth $γ_c\approx 0.42$ in sufficiently large models. We formalize the forward pass as a discrete coarse-graining map and relate the appearance of stable "concept basins" to fixed points of this renormalization-like dynamics. The resulting low-entropy regime is characterized by a spectral tail collapse and by the formation of transient, reusable object-like structures in representation space, which we call Transient Class Objects (TCOs). We provide theoretical conditions connecting logical separability to spectral decay and validate the predicted signatures with layerwise probes on multiple open-weight model families.

Oculomics - the concept of predicting systemic diseases, such as cardiovascular disease and dementia, through retinal imaging - has advanced rapidly due to the data efficiency of transformer-based foundation models like RETFound. Image-level mixed sample data augmentations, such as CutMix and MixUp, are frequently used for training transformers, yet these techniques perturb patient-specific attributes, such as medical comorbidity and clinical factors, since they only account for images and labels. To address this limitation, we propose a hierarchical sampling strategy, Oculomix, for mixed sample augmentations. Our method is based on two clinical priors. First (exam level), images acquired from the same patient at the same time point share the same attributes. Second (patient level), images acquired from the same patient at different time points have a soft temporal trend, as morbidity generally increases over time. Guided by these priors, our method constrains the mixing space to the patient and exam levels to better preserve patient-specific characteristics and leverages their hierarchical relationships. The proposed method is validated using ViT models on a five-year prediction of major adverse cardiovascular events (MACE) in a large ethnically diverse population (Alzeye). We show that Oculomix consistently outperforms image-level CutMix and MixUp by up to 3% in AUROC, demonstrating the necessity and value of the proposed method in oculomics.

Decentralized Federated Learning (DFL) is a serverless collaborative machine learning paradigm where devices collaborate directly with neighbouring devices to exchange model information for learning a generalized model. However, variations in individual experiences and different levels of device interactions lead to data and model initialization heterogeneities across devices. Such heterogeneities leave variations in local model parameters across devices that leads to slower convergence. This paper tackles the data and model heterogeneity by explicitly addressing the parameter level varying evidential credence across local models. A novel aggregation approach is introduced that captures these parameter variations in local models and performs robust aggregation of neighbourhood local updates. Specifically, consensus weights are generated via approximation of second-order information of local models on their local datasets. These weights are utilized to scale neighbourhood updates before aggregating them into global neighbourhood representation. In extensive experiments with computer vision tasks, the proposed approach shows strong generalizability of local models at reduced communication costs.

Large Language Model (LLM)-based agents are increasingly deployed for complex, tool-based tasks where long-term memory is critical to driving actions. Existing benchmarks, however, primarily test a angent's ability to passively retrieve isolated facts in response to explicit questions. They fail to evaluate the more crucial capability of actively applying memory to execute tasks. To address this gap, we introduce \textsc{Mem2ActBench}, a benchmark for evaluating whether agents can proactively leverage long-term memory to execute tool-based actions by selecting appropriate tools and grounding their parameters. The benchmark simulates persistent assistant usage, where users mention the same topic across long, interrupted interactions and expect previously established preferences and task states to be implicitly applied. We build the dataset with an automated pipeline that merges heterogeneous sources (ToolACE, BFCL, Oasst1), resolves conflicts via consistency modeling, and synthesizes 2,029 sessions with 12 user--assistant--tool turns on average. From these memory chains, a reverse-generation method produces 400 tool-use tasks, with human evaluation confirming 91.3\% are strongly memory-dependent. Experiments on seven memory frameworks show that current systems remain inadequate at actively utilizing memory for parameter grounding, highlighting the need for more effective approaches to evaluate and improve memory application in task execution.

Large language models (LLMs) are widely used for tasks ranging from summarisation to decision support. In practice, identical prompts do not always produce identical outputs, even when temperature and other decoding parameters are fixed. In this work, we conduct repeated-run experiments to empirically quantify baseline behavioural drift, defined as output variability observed when the same prompt is issued multiple times under operator-free conditions. We evaluate two publicly accessible models, gpt-4o-mini and llama3.1-8b, across five prompt categories using exact repeats, perturbed inputs, and reuse modes at temperatures of 0.0 and 0.7. Drift is measured using unique output fractions, lexical similarity, and word count statistics, enabling direct comparison across models, prompting modes, and deployment types. The results show that nondeterminism persists even at temperature 0.0, with distinct variability patterns by model size, deployment, and prompt type. We situate these findings within existing work on concept drift, behavioural drift, and infrastructure-induced nondeterminism, discuss the limitations of lexical metrics, and highlight emerging semantic approaches. By establishing a systematic empirical baseline in the absence of stabilisation techniques, this study provides a reference point for evaluating future drift mitigation and control methods.

We introduce a new benchmark for Danish culture via cultural heritage, Daisy, based on the curated topics from the Danish Culture Canon 2006. For each artifact in the culture canon, we query the corresponding Wikipedia page and have a language model generate random questions. This yields a sampling strategy within each work, with a mix of central of peripheral questions for each work, not only knowledge of mainstream information, but also in-depth cornerstones defining the heritage of Danish Culture, defined by the Canon committee. Each question-answer pair is humanly approved or corrected in the final dataset consisting of 741 close-ended question answer pairs covering topics, from 1300 BC. archaeological findings, 1700 century poems and musicals pieces to contemporary pop music and Danish design and architecture.