Abstract Recent advances in machine learning, particularly in multimodal models, have created new opportunities for analyzing complex data in high-energy physics, where accurate identification of particle interactions is critical for scientific discovery. However, existing approaches rely heavily on convolutional neural networks, which lack interpretability and do not fully leverage multimodal reasoning capabilities. Here we show that a fine-tuned Vision Language Model (VLM) based on LLaMA 3.2 can effectively identify neutrino interactions in pixelated detector data, outperforming both a state-of-the-art convolutional neural network and a Vision Transformer baseline in classification accuracy and robustness. In addition, the VLM provides improved explainability through reasoning-based, interpretable predictions and supports integration of auxiliary semantic information. These results demonstrate the potential of multimodal transformer architectures as general-purpose tools for physics event classification, paving the way for more transparent, flexible, and scalable analysis methods in future high-energy physics experiments.

Finding approximate equilibria for large-scale imperfect-information competitive games such as StarCraft, Dota, and CounterStrike remains computationally infeasible due to sparse rewards and challenging exploration over long horizons. In this paper, we propose a multi-agent starting-state sampling strategy designed to substantially accelerate online exploration in regularized policy-gradient game methods for two-player zero-sum (2p0s) games. Motivated by an assumption that offline demonstrations from skilled humans can provide good coverage of high-level strategies relevant to equilibrium play, we propose the initialization of reinforcement learning data collection at intermediate states sampled from offline data to facilitate exploration of strategically relevant subgames. Referring to this method as Data-Augmented Game Starts (DAGS), we perform experiments using synthetic datasets and analytically tractable, long-horizon control variants of two-player Kuhn Poker, Goofspiel, and a counterexample game designed to penalize biased beliefs over hidden information. Under fixed computational budgets, DAGS enables regularized policy gradient methods to achieve lower exploitability in games with significantly more challenging exploration. We show that augmenting starting state distributions when solving imperfect information games can lead to biased equilibria, and we provide a straightforward mitigation to this in the form of multi-task observation flags. Finally, we release a new set of benchmark environments that drastically increase exploration challenges and state counts in existing OpenSpiel games while keeping exploitability measurements analytically tractable.

The gut microbiota is increasingly recognized as a regulator of brain function, yet its role in experience-dependent plasticity during postnatal development remains largely unknown. Here, we show that disrupting the gut microbiota with antibiotics during critical periods of visual cortex development impairs ocular dominance plasticity (ODP) in juvenile mice. Antibiotic treatment induces marked changes in microbial community composition and is accompanied by extensive transcriptional remodeling of the visual cortex, including pathways involved in extracellular matrix organization, blood-brain barrier function, and myelination. Remarkably, fecal transplantation of the juvenile microbiota into adult recipients restores ODP. These findings identify the gut microbiota as a previously unrecognized regulator of neurodevelopmental plasticity and support the existence of microbiota-dependent critical periods of brain development. More broadly, our results suggest that early-life microbial perturbations may have lasting consequences for lifelong brain function and reveal that juvenile microbiota-derived signals could be exploited to promote plasticity in the adult brain.

Abstract Biologically informed neural networks (BINNs), also known as visible neural networks (VNNs), are widely adopted in omics because their architectures mirror known biological structures, such as gene-to-pathway relationships, and are therefore often assumed to be inherently interpretable. This assumption implies that learned gene-to-pathway weights and pathway node activations reflect meaningful biological mechanisms. Here, we show that this premise fails for a classical reason: nonidentifiability. Using a controlled teacher and student framework, we demonstrate that even under ideal conditions, including noiseless data, the correct model class, and identical sparse wiring, a BINN can perfectly recover the input-to-output mapping while failing to recover both gene-to-pathway weights and pathway activations. This failure persists across classification, regression, and survival tasks, and remains robust to variations in biological structure and network depth. Thus, the problem is not merely overparameterization or poor optimization: learning from outputs alone does not identify internal structure. Since biological mechanisms are not directly observed, recovering them from predictions alone is harder, not easier , than recovering neural network parameters, which are already known to be nonidentifiable. Critically, this failure reflects standard practice: widely used BINNs do not impose objective level constraints on gene-to-pathway weights or pathway activations, and therefore operate precisely in the regime modeled by our teacher-student framework. These results indicate that architectural transparency does not imply mechanistic interpretability. Without constraints that explicitly enforce identifiability, the apparent interpretability of BINNs reflects their design rather than what they actually learn.

Circadian clocks present throughout the brain and body coordinate diverse physiological processes to support daily homeostasis, yet the specific interorgan signaling axes involved are not well defined. We previously demonstrated that the skeletal muscle clock controls transcript oscillations of genes involved in fatty acid metabolism in the liver, yet the impact of the liver clock on the muscle remained unknown. Here, we use male hepatocyte-specific Bmal1 KO mice (Bmal1hep-/-) to reveal that approximately one-third of transcript rhythms in skeletal muscle are influenced by the liver clock in vivo. Treatment of myotubes with serum harvested from Bmal1hep-/- mice inhibits expression of genes involved in metabolic pathways, including oxidative phosphorylation. Only small transcriptional changes were induced by liver clock-driven endocrine communication in vitro, leading us to surmise that the liver clock acts to fine-tune metabolic gene expression in muscle. Consistent with functional tuning, treatment of myotubes with serum collected from Bmal1hep-/- mice during the dark phase lowers mitochondrial ATP production compared with serum from wild-type mice. Overall, our results reveal communication between the liver clock and skeletal muscle, uncovering a bidirectional endocrine communication pathway that may contribute to the metabolic phenotypes of circadian disruption.

Chromatin-modifying and -remodeling machineries are important for learning-induced transcriptional activity, yet it remains unclear how they coordinate to drive de novo gene expression for memory formation. Here, we examine the transcription factor known as calcium-responsive transactivator (CREST) in memory formation, synaptic plasticity, and learning-induced gene expression. CREST is known to bind major chromatin-modifying and -remodeling machineries via interaction with CREB-binding protein (CBP) and brahma-related gene 1 (BRG1), respectively. In silico modeling of CREST identified tyrosine 397 (Y397) within the CBP-binding domain. Expression of a CREST Y397F point mutant impairs long-term potentiation and memory. Conversely, expression of a CREST Y397D point mutant enhances memory in a CBP-dependent manner. Differential gene expression analysis reveals distinct CREST Y397-regulated signatures during memory consolidation. CBP acts through CREB and post-translation modifications to affect memory, but the findings of this study argue for consideration of the CREST-CBP interaction and Y397 accessibility as factors in memory processes.

Circadian misalignment of the feeding behavior and the terrestrial cycle is associated with obesity and metabolic perturbations. However, it remains unclear whether the quantity and timing of dietary salt intake influence temporal sodium handling and blood pressure regulation. Here, we demonstrate that the colonic mineralocorticoid receptor (MR) and peripheral clock affect the daily sodium absorption and blood pressure variations. Genes related to sodium handling display diurnal rhythms in synchrony with the daily rhythms of aldosterone and the colonic circadian clock. Cistromic analysis substantiated the overlap of occupancy between the MR and brain and muscle ARNT-like 1 (BMAL1). Diurnal oscillation of aldosterone and peripheral clocks, as well as blood pressure, was robustly driven by nighttime feeding of a low-salt diet but markedly disrupted by daytime feeding of a high-salt diet in nocturnal mice. These findings delineate the colonic temporal sensing of dietary sodium abundance and highlight the transcriptional mechanisms of daily salt handling and blood pressure variations.

Obstructive sleep apnea (OSA), characterized by chronic intermittent hypoxia (IH) during sleep, is increasingly recognized as a driver of metabolic dysfunction. However, its impact on circadian metabolic regulation remains poorly understood. Here, we investigated how chronic IH reshapes 24-hour hepatic and systemic metabolic programs in a mouse model mimicking OSA-associated chronic hypoxia. Through integrated circadian transcriptomic, metabolomic, and physiological 24-hour profiling, we show that 4 weeks of rest phase-restricted IH reprograms hepatic and systemic metabolism in a time-specific manner. This reorganization involves the coordinated circadian regulation of glucose, lipid, and redox pathways, with a shift away from oxidative metabolism toward oxygen-sparing processes such as gluconeogenesis, glycogen turnover, and lipid mobilization. These changes align with the hypoxic phase exposure and coincide with reshaped hepatic metabolite oscillations and systemic autonomic rhythms, supporting a functional adaptation to intermittent oxygen availability. Mechanistically, we identify the cAMP-CREB1 pathway as a driver of circadian transcriptional remodeling in the liver and a central integrator of IH-dependent adrenergic stress. Our findings establish chronic IH as a potent metabolic zeitgeber that rewires hepatic transcriptional and metabolic programs, revealing a circadian dimension to the metabolic consequences of sleep-disordered breathing.

Electromagnetic field reconstruction is crucial in many applications, including antenna diagnostics, electromagnetic interference analysis, and system modeling. This paper presents a deep learning-based approach for Far-Field to Near-Field (FF-NF) transformation using Convolutional Neural Networks (CNNs). The goal is to reconstruct near-field distributions from the far-field data of an antenna without relying on explicit analytical transformations. The CNNs are trained on paired far-field and near-field data and evaluated using mean squared error (MSE). The best model achieves a training error of 0.0199 and a test error of 0.3898. Moreover, visual comparisons between the predicted and true near-field distributions demonstrate the model's effectiveness in capturing complex electromagnetic field behavior, highlighting the potential of deep learning in electromagnetic field reconstruction.

Modern climate projections lack adequate spatial and temporal resolution due to computational constraints, leading to inaccuracies in representing critical processes like thunderstorms that occur on the sub-resolution scale. Hybrid methods combining physics with machine learning (ML) offer faster, higher fidelity climate simulations by outsourcing compute-hungry, high-resolution simulations to ML emulators. However, these hybrid physics-ML simulations require domain-specific data and workflows that have been inaccessible to many ML experts. This paper is an extended version of our NeurIPS award-winning ClimSim dataset paper (Yu et al., 2024). The ClimSim dataset includes 5.7 billion pairs of multivariate input/output vectors spanning ten years at high temporal resolution, capturing the influence of high-resolution, high-fidelity physics on a host climate simulator's macro-scale state. In this extended version, we introduce a significant new contribution in Section 5, which provides a cross-platform, containerized pipeline to integrate ML models into operational climate simulators for hybrid testing. We also implement various baselines of ML models and hybrid simulators to highlight the ML challenges of building stable, skillful emulators.

We develop a general theory of synaptic neural balance and how it can emerge or be enforced in neural networks. For a given additive cost function R (regularizer), a neuron is said to be in balance if the total cost of its input weights is equal to the total cost of its output weights. The basic example is provided by feedforward networks of ReLU units trained with L2 regularizers, which exhibit balance after proper training. The theory explains this phenomenon and extends it in several directions. The first direction is the extension to bilinear and other activation functions. The second direction is the extension to more general regularizers, including all Lp (p>0) regularizers. The third direction is the extension to non-layered architectures, recurrent architectures, convolutional architectures, as well as architectures with mixed activation functions and to different balancing algorithms. Gradient descent on the error function alone does not converge in general to a balanced state, where every neuron is in balance, even when starting from a balanced state. However, gradient descent on the regularized error function ought to converge to a balanced state, and thus network balance can be used to assess learning progress. The theory is based on two local neuronal operations: scaling which is commutative, and balancing which is not commutative. Finally, and most importantly, given any set of weights, when local balancing operations are applied to each neuron in a stochastic manner, global order always emerges through the convergence of the stochastic balancing algorithm to the same unique set of balanced weights. The reason for this convergence is the existence of an underlying strictly convex optimization problem where the relevant variables are constrained to a linear, only architecture-dependent, manifold. Simulations show that balancing neurons prior to learning, or during learning in alternation with gradient descent steps, can improve learning speed and performance thereby expanding the arsenal of available training tools. Scaling and balancing operations are entirely local and thus physically plausible in biological and neuromorphic neural networks.

BACKGROUND:Postinduction hypotension is a well-known risk factor for adverse postoperative outcomes. Anesthesiologists estimate anesthetic dosages based on a patient's chart and domain knowledge. Machine learning is increasingly applied in predicting postinduction hypotension, with neural networks providing a robust and accurate approach to model complex relationships. This study aims to use machine learning to suggest anesthetic doses, both generalized to an average patient population and personalized for specific patients, incorporating domain knowledge into the modeling process. METHODS:In this study, postinduction hypotension is defined as a mean arterial pressure (<65 mm Hg) occurring during the first 10 minutes after anesthesia induction. The dataset includes 201,000 patient records, after exclusion criteria, containing clinical data, medication history, procedure descriptions, and anesthetic dosages for fentanyl and propofol. Several classification algorithms were implemented to model postinduction hypotension, and likelihood calculations were made by fixing values of fentanyl and propofol dosages to assess patient risk. RESULTS:Gradient boosting and neural network models demonstrated the highest performance. However, these models did not account for domain experts' knowledge that anesthetic dosage and postinduction hypotension have a monotonically increasing relationship. To address this limitation, we developed a monotonic neural network (MNN), which integrates this domain knowledge. The models' results are presented through heatmaps, illustrating the likelihood of postinduction hypotension for both average and specific patients, with the MNN generating smoother, more plausible predictions compared to traditional models. CONCLUSIONS:We successfully predicted postinduction hypotension using the MNN, achieving performance comparable to existing methods. This model, by encoding clinically relevant monotonic relationships, provides anesthesiologists with a tool to assist in patient-specific fentanyl and propofol dosages, improving both the interpretability and clinical relevance of anesthetic dosing strategies.

The reaction predictor expands and searches the synthesis pathway tree through a series of exponentially growing predictions that can ultimately explain the reasons behind the ultra-stretchability of our hydrogel.

Antenna arrays are widely used in wireless communication, radar systems, radio astronomy, and military defense to enhance signal strength, directivity, and interference suppression. We introduce a deep learning-based optimization approach that enhances the design of sparse phased arrays by reducing grating lobes. This approach begins by generating sparse array configurations to address the non-convex challenges and extensive degrees of freedom inherent in array design. We use neural networks to approximate the non-convex cost function that estimates the energy ratio between the main and side lobes. This differentiable approximation facilitates cost function minimization through gradient descent, optimizing the antenna elements' coordinates and leading to an improved layout. Additionally, we incorporate a tailored penalty mechanism that includes various physical and design constraints into the optimization process, enhancing its robustness and practical applicability. We demonstrate the effectiveness of our method by applying it to the ten array configurations with the lowest initial costs, achieving further cost reductions ranging from 411 reducing side lobe levels in antenna arrays, this breakthrough paves the way for ultra-precise beamforming, enhanced interference mitigation, and next-generation wireless and radar systems with unprecedented efficiency and clarity.

The measurements performed by particle physics experiments must account for the imperfect response of the detectors used to observe the interactions. One approach, unfolding, statistically adjusts the experimental data for detector effects. Recently, generative machine learning models have shown promise for performing unbinned unfolding in a high number of dimensions. However, all current generative approaches are limited to unfolding a fixed set of observables, making them unable to perform full-event unfolding in the variable dimensional environment of collider data. A novel modification to the variational latent diffusion model (VLD) approach to generative unfolding is presented, which allows for unfolding of high- and variable-dimensional feature spaces. The performance of this method is evaluated in the context of semi-leptonic top quark pair production at the Large Hadron Collider.

A large data set of kinetically plausible proton transfer steps was created. A set of over 48 million proton transfer steps, between heteroatoms, was generated combinatorially from a set of about 8,000 acids and conjugate bases for which experimental aqueous pKas around room temperature were available. The set was augmented with about 100 estimated pKas of highly reactive species important for reaction mechanisms. The resulting set of pKas span a range from -15 to +37. Rate constants were estimated at 25 °C using the pKas and utilizing a simplified Eigen equation without statistical factors. Steps with estimated rate constants ≥ 10^3 M^-1 s^-1 – a conservative boundary – were included in the data set. An additional set of 15,138 proton transfer steps were estimated using the Eigen-Bernasconi equation for proton transfers from carbon acids to heteroatom bases for which intrinsic rate constants and Brønsted 𝛽 values were known. Steps for proton transfers from carbon with estimated rate constants ≥ 10^-1 M^-1 s^-1 were added to the data set. Each entry was encoded in SMIRKS format, which is commonly used for machine learning, with electron-flow specification. The objective of this work was the creation of a structurally rich data set rather than accurate calculation of rate constants.

We present theory of synaptic neural balance and we show experimentally that synaptic neural balance can improve deep learning speed, and accuracy, even in data-scarce environments. Given an additive cost function (regularizer) of the synaptic weights, a neuron is said to be in balance if the total cost of its incoming weights is equal to the total cost of its outgoing weights. For large classes of networks, activation functions, and regularizers, neurons can be balanced fully or partially using scaling operations that do not change their functionality. Furthermore, these balancing operations are associated with a strictly convex optimization problem with a single optimum and can be carried out in any order. In our simulations, we systematically observe that: (1) Fully balancing before training results in better performance as compared to several other training approaches; (2) Interleaving partial (layer-wise) balancing and stochastic gradient descent steps during training results in faster learning convergence and better overall accuracy (with L1 balancing converging faster than L2 balancing); and (3) When given limited training data, neural balanced models outperform plain or regularized models; and this is observed in both feedforward and recurrent networks. In short, the evidence supports that neural balancing operations could be added to the arsenal of methods used to regularize and train neural networks. Furthermore, balancing operations are entirely local and can be carried out asynchronously, making them plausible for biological or neuromorphic systems.

Measuring observables to constrain models using maximum-likelihood estimation is fundamental to many physics experiments. Wilks' theorem provides a simple way to construct confidence intervals on model parameters, but it only applies under certain conditions. These conditions, such as nested hypotheses and unbounded parameters, are often violated in neutrino oscillation measurements and other experimental scenarios. Monte Carlo methods can address these issues, albeit at increased computational cost. In the presence of nuisance parameters, however, the best way to implement a Monte Carlo method is ambiguous. This paper documents the method selected by the NOvA experiment, the profile construction. It presents the toy studies that informed the choice of method, details of its implementation, and tests performed to validate it. It also includes some practical considerations which may be of use to others choosing to use the profile construction.

Recent advances in Large Language Models (LLMs) have demonstrated their remarkable capacity to process and reason over structured and unstructured data modalities beyond natural language. In this work, we explore the applications of Vision Language Models (VLMs), specifically a fine-tuned variant of LLaMa 3.2, to the task of identifying neutrino interactions in pixelated detector data from high-energy physics (HEP) experiments. We benchmark this model against a state-of-the-art convolutional neural network (CNN) architecture, similar to those used in the NOvA and DUNE experiments, which have achieved high efficiency and purity in classifying electron and muon neutrino events. Our evaluation considers both the classification performance and interpretability of the model predictions. We find that VLMs can outperform CNNs, while also providing greater flexibility in integrating auxiliary textual or semantic information and offering more interpretable, reasoning-based predictions. This work highlights the potential of VLMs as a general-purpose backbone for physics event classification, due to their high performance, interpretability, and generalizability, which opens new avenues for integrating multimodal reasoning in experimental neutrino physics.