Papers

Epileptic seizure detection from EEG signals remains challenging due to the high dimensionality and nonlinear, potentially stochastic, dynamics of neural activity. In this work, we investigate whether features derived from topological data analysis (TDA) can improve the classification of brain states in preictal, ictal and interictal iEEG recordings from epilepsy patients using multichannel data. We analyze data from 55 patients, significantly larger than many previous studies that rely on patient-specific models. Persistence diagrams derived from iEEG signals are vectorized using several TDA representations, including Carlsson coordinates, persistence images, and template functions. To understand how topological representations interact with modern machine learning pipelines, we conduct a large-scale ablation study across multiple iEEG frequency bands, dimensionality reduction techniques, feature representations, and classifier architectures. Our experiments show that dimension-reduced topological representations achieve up to 80% balanced accuracy for three-class classification. Interestingly, classical machine learning models perform comparably to deep learning models, achieving up to 79.17% balanced accuracy, suggesting that carefully designed topological features can substantially reduce model complexity requirements. In contrast, pipelines preserving the full multichannel feature structure exhibit severe overfitting due to the high-dimensional feature space. These findings highlight the importance of structure-preserving dimensionality reduction when applying topology-based representations to multichannel neural data.

We introduce INDOTABVQA, a benchmark for evaluating cross-lingual Table Visual Question Answering (VQA) on real-world document images in Bahasa Indonesia. The dataset comprises 1,593 document images across three visual styles (bordered, borderless, and colorful) with one or more than one tables, and 1,593 question-answer sets in four languages: Bahasa Indonesia, English, Hindi, and Arabic. This enables evaluation of Vision-Language Models (VLMs) in both monolingual (Bahasa documents with Bahasa questions) and cross-lingual settings (Bahasa documents with questions in other languages). We benchmark leading open-source VLMs (Qwen2.5-VL, Gemma-3, LLaMA-3.2) and GPT-4o and reveal substantial performance gaps, particularly on structurally complex tables and in low-resource languages. Fine-tuning a compact 3B and LoRA-finetuned 7B model on our dataset yields 11.6% and 17.8% improvements in accuracy. Providing explicit table region coordinates as additional input further improves performance by 4-7%, demonstrating the value of Spatial priors for table-based reasoning. Our findings underscore the importance of language-diverse, domain-specific datasets and demonstrate that targeted fine-tuning can significantly enhance VLM performance on specialized document understanding tasks. INDOTABVQA provides a valuable resource for advancing research in cross-lingual, structure-aware document understanding, especially in underrepresented regions of the world. Full dataset can be accessed in huggingface at: https://huggingface.co/datasets/NusaBharat/INDOTABVQA}

The Linear Representation Hypothesis (LRH) identifies features of a trained deep network (DN) as linear directions in the activation spaces, i.e., output spaces of intermediate layers. This characterization decouples the input-output maps learned by a DN from the organization of feature directions in its activation spaces. We introduce the Linear Centroids Hypothesis (LCH), which instead identifies features with linear directions among a DN's centroid spaces – where any vector denotes a centroid or summary of a local affine expert characterizing the learned input-output maps of the DN exactly (e.g., for piecewise-affine DNs) or approximately (e.g., for smooth DNs like transformers). We show that replacing intermediate activations with centroids yields a functional drop-in alternative for standard interpretability tools. Empirically, this change yields sparser, more downstream-useful feature dictionaries on DINO ViTs, suppresses spurious directions on a controlled task, recovers interpretable circuits in GPT2-Large, and produces faithful gradient-based saliency maps. LCH unifies dictionaries, probing, circuits, and saliency maps into a single geometric object grounded in the network's input-output map – making interpretability mechanistic by construction rather than post hoc. Code to study the LCH https://github.com/ThomasWalker1/LinearCentroidsHypothesis .

Self-driving laboratories promise to accelerate materials discovery. Yet current automated solid-state synthesis platforms are limited to ambient conditions, thereby precluding their use for air-sensitive materials. Here, we present A-Lab for Glovebox Powder Solid-state Synthesis (A-Lab GPSS), a robotic platform capable of synthesizing and characterizing air-sensitive inorganic materials under strict air-free conditions. By integrating an agentic AI framework into the A-Lab GPSS platform, we structure autonomous experimental design through abductive and inductive reasoning. We deploy this platform to explore the vast compositional space of lithium halide spinel solid-state ionic conductors. Across a synthesis campaign comprising 352 samples with diverse compositions, the system explores a broad chemical space, experimentally realizing 72% of the 171 possible pairwise combinations among the 19 metals considered in this study. Over the course of the campaign, the fraction of compositions exhibiting both good ionic conductivity (> 0.05 mS/cm) and high halide spinel phase purity increases from 1.33% in the first 75 agent-proposed samples to 5.33% in the final 75. Furthermore, by inspecting the AI's reasoning processes, we reveal distinct yet complementary discovery strategies: abductive reasoning interrogates abnormal observations within already explored regions, whereas inductive reasoning expands the search into broader, previously unvisited chemical space. This work establishes a scalable platform for the autonomous discovery of complex, air-sensitive solid-state materials.

Large Foundation Model (LFM) inference is both memory- and compute-intensive, traditionally relying on GPUs. However, the limited availability and high cost have motivated the adoption of high-performance general-purpose CPUs, especially emerging 3D-stacked Static Non-Uniform Cache Architecture (3D S-NUCA) systems. These architectures offer enhanced bandwidth and locality but suffer from severe thermal challenges and uneven cache latencies due to 3D Networks-on-Chip (NoC). Optimal management of thread migration and V/f scaling is non-trivial due to LFM kernel diversity and system heterogeneity. Existing thermal management approaches often rely on oversimplified analytical models and lack adaptability. We propose AILFM, an Active Imitation Learning (AIL)-based scheduling framework that learns near-optimal thermal-aware scheduling policies from Oracle demonstrations with minimal run-time overhead. AILFM accounts for both core-level performance heterogeneity and kernel-specific behavior in LFMs to maintain thermal safety while maximizing performance. Extensive experiments show that AILFM outperforms state-of-the-art baselines and generalizes well across diverse LFM workloads.

Unlocking large-scale low-bandwidth decentralized training has the potential to utilize otherwise untapped compute resources. In centralized settings, large-scale multi-node training is primarily enabled by data and pipeline parallelism, two techniques that require ultra-high-bandwidth communication. While efficient methods now exist for decentralized data parallelism, pipeline parallelism remains the primary challenge. Recent efforts, such as Subspace Models (SM), have claimed up to 100x activation compression but rely on complex constrained optimization and diverge from true end-to-end training. In this paper, we propose a different approach, based on an architecture designed from the ground up to be native to low-bandwidth communication environments while still applicable to any standard transformer-based architecture. We call this architecture the Residual Bottleneck Model or ResBM, it introduces a residual encoder-decoder bottleneck module across pipeline boundaries that can be trained end-to-end as part of the model's parameters while preserving an explicit low-rank identity path. We show that ResBMs achieve state-of-the-art 128x activation compression without significant loss in convergence rates and without significant memory or compute overhead.

High-fidelity numerical simulation of subsurface flow is computationally intensive, especially for many-query tasks such as uncertainty quantification and data assimilation. Deep learning (DL) surrogates can significantly accelerate forward simulations, yet constructing them requires substantial machine learning (ML) expertise - from architecture design to hyperparameter tuning - that most domain scientists do not possess. Furthermore, the process is predominantly manual and relies heavily on heuristic choices. This expertise gap remains a key barrier to the broader adoption of DL surrogate techniques. For this reason, we present AutoSurrogate, a large-language-model-driven multi-agent framework that enables practitioners without ML expertise to build high-quality surrogates for subsurface flow problems through natural-language instructions. Given simulation data and optional preferences, four specialized agents collaboratively execute data profiling, architecture selection from a model zoo, Bayesian hyperparameter optimization, model training, and quality assessment against user-specified thresholds. The system also handles common failure modes autonomously, including restarting training with adjusted configurations when numerical instabilities occur and switching to alternative architectures when predictive accuracy falls short of targets. In our setting, a single natural-language sentence can be sufficient to produce a deployment-ready surrogate model, with minimum human intervention required at any intermediate stage. We demonstrate the utility of AutoSurrogate on a 3D geological carbon storage modeling task, mapping permeability fields to pressure and CO$_2$ saturation fields over 31 timesteps. Without any manual tuning, AutoSurrogate is able to outperform expert-designed baselines and domain-agnostic AutoML methods, demonstrating strong potential for practical deployment.

Diabetes devices, including Continuous Glucose Monitoring (CGM), Smart Insulin Pens, and Automated Insulin Delivery systems, generate rich time-series data widely used in research and machine learning. However, inconsistent data formats across sources hinder sharing, integration, and analysis. We present DIAX (DIAbetes eXchange), a standardized JSON-based format for unifying diabetes time-series data, including CGM, insulin, and meal signals. DIAX promotes interoperability, reproducibility, and extensibility, particularly for machine learning applications. An open-source repository provides tools for dataset conversion, cross-format compatibility, visualization, and community contributions. DIAX is a translational resource, not a data host, ensuring flexibility without imposing data-sharing constraints. Currently, DIAX is compatible with other standardization efforts and supports major datasets (DCLP3, DCLP5, IOBP2, PEDAP, T1Dexi, Loop), totaling over 10 million patient-hours of data. https://github.com/Center-for-Diabetes-Technology/DIAX

An OS kernel that runs LLM inference internally can read the model's own next-token logit distribution before any text is generated, and act on it as a governance primitive. I present ProbeLogits, a kernel-level operation that performs a single forward pass and reads specific token logits to classify an agent's action as safe or dangerous, with zero learned parameters. Because the probe reads a logit from the same base model the agent already runs, it removes the second model a fine-tuned guard requires: the marginal cost of a safety check becomes a single logit read. I evaluate ProbeLogits on three base models (Qwen2.5-7B, Llama-3-8B, Mistral-7B) across three external benchmarks (HarmBench, XSTest, ToxicChat). On HarmBench non-copyright, all three reach a 97-99% block rate. On ToxicChat (n=1,000), ProbeLogits attains F1 parity-or-better against Llama Guard 3: Qwen2.5-7B Safe/Dangerous reaches F1=0.812 (+13.7 pp, bootstrap 95% CIs disjoint), Llama-3 matches within CI (+0.4 pp), and Mistral exceeds by +4.4 pp. Classification is a measured 2.4-3.4x faster than Llama Guard 3 (332-556 ms vs. 851-1,142 ms), because it reads a single logit position instead of generating tokens. A calibration strength alpha acts as a deployment-time policy knob rather than a learned hyperparameter, trading recall for precision per operation class. I implement ProbeLogits within Anima OS, a bare-metal x86-64 kernel written in ~285,000 lines of Rust. Because agent actions must pass through kernel-mediated host functions, enforcement operates below the WASM sandbox boundary, making it substantially harder to circumvent than application-layer classifiers.

Our ability to predict, control, and ultimately understand complex systems rests on discovering the equations that govern their dynamics. Identifying these equations directly from noisy, limited observations has therefore become a central challenge in data-driven science, yet existing library-based sparse regression methods force a compromise between automation, statistical rigor, and computational efficiency. Here we develop Bayesian-ARGOS, a hybrid framework that reconciles these demands by combining rapid frequentist screening with focused Bayesian inference, enabling automated equation discovery with principled uncertainty quantification at a fraction of the computational cost of existing methods. Tested on seven chaotic systems under varying data scarcity and noise levels, Bayesian-ARGOS outperforms two state-of-the-art methods in most scenarios. It surpasses SINDy in data efficiency for all systems and noise tolerance for six out of the seven, with a two-order-of-magnitude reduction in computational cost compared to bootstrap-based ARGOS. The probabilistic formulation additionally enables a suite of standard statistical diagnostics, including influence analysis and multicollinearity detection that expose failure modes otherwise opaque. When integrated with representation learning (SINDy-SHRED) for high dimensional sea surface temperature reconstruction, Bayesian-ARGOS increases the yield of valid latent equations with significantly improved long horizon stability. Bayesian-ARGOS thus provides a principled, automated, and computationally efficient route from scarce and noisy observations to interpretable governing equations, offering a practical framework for equation discovery across scales, from benchmark chaotic systems to the latent dynamics underlying global climate patterns.

Time-series forecasting aims to predict future values by modeling temporal dependencies in historical observations. It is a critical component of many real-world systems, where accurate forecasts improve operational efficiency and help mitigate uncertainty and risk. More recently, machine learning (ML), and especially deep learning (DL)-based models, have gained widespread adoption for time-series forecasting, but they remain vulnerable to adversarial attacks. However, many state-of-the-art attack methods are not directly applicable in time-series settings, where storing complete historical data or performing attacks at every time step is often impractical. This paper proposes an adversarial attack framework for time-series forecasting under an online bounded-buffer setting, leveraging an informed and selective attack strategy. By selectively targeting time steps where the model exhibits high confidence and the expected prediction error is maximal, our framework produces fewer but substantially more effective attacks. Experiments show that our framework can increase the prediction error up to 2.42x, while performing attacks in fewer than 10% of time steps.

Using FlowBoost, a closed-loop deep generative optimization framework for extremal structure discovery, we investigate $\ell^p$-generalizations of the finite free Stam inequality for real-rooted polynomials under finite free additive convolution $\boxplus_n$. At $p=2$, FlowBoost finds the Hermite pair as the unique equality case and reveals the spectral structure of the linearized convolution map at this extremal point. As a result, we conjecture that the singular values of the doubly stochastic coupling matrix $E_n$ on the mean-zero subspace are ${2^{-k/2}:k=1,\ldots,n-1}$, independent of $n$. Conditional on this conjecture, we obtain a sharp local stability constant and the finite free CLT convergence rate, both uniform in $n$. We introduce a one-parameter family of $p$-Stam inequalities using $\ell^p$-Fisher information and prove that the Hermite pair itself violates the inequality for every $p>2$, with the sign of the deficit governed by the $\ell^p$-contraction ratio of $E_n$. Systematic computation via FlowBoost supports the conjecture that $p^*\!=2$ is the sharp critical exponent. For $p<2$, the extremal configurations undergo a bifurcation, meaning that they become non-matching pairs with bimodal root structure, converging back to the Hermite diagonal only as $p\to 2^-$. Our findings demonstrate that FlowBoost, can be an effective tool of mathematical discovery in infinite-dimensional extremal problems.

Modern machine learning methods have been proposed to detect life in extraterrestrial samples, drawing on their ability to distinguish biotic from abiotic samples based on training models using natural and synthetic organic molecular mixtures. Here we show using Artificial Life that such methods are easily fooled into detecting life with near 100% confidence even if the analyzed sample is not capable of life. This is due to modern machine learning methods' propensity to be easily fooled by out-of-distribution samples. Because extra-terrestrial samples are very likely out of the distribution provided by terrestrial biotic and abiotic samples, using AI methods for life detection is likely to yield significant false positives.

While next-token prediction (NTP) has been the standard objective for training language models, it often struggles to capture global structure in reasoning tasks. Multi-token prediction (MTP) has recently emerged as a promising alternative, yet its underlying mechanisms remain poorly understood. In this paper, we study how MTP facilitates reasoning, with a focus on planning. Empirically, we show that MTP consistently outperforms NTP on both synthetic graph path-finding tasks and more realistic reasoning benchmarks, such as Countdown and boolean satisfiability problems. Theoretically, we analyze a simplified two-layer Transformer on a star graph task. We prove that MTP induces a two-stage reverse reasoning process: the model first attends to the end node and then reconstructs the path by tracing intermediate nodes backward. This behavior arises from a gradient decoupling property of MTP, which provides a cleaner training signal compared to NTP. Ultimately, our results highlight how multi-token objectives inherently bias optimization toward robust and interpretable reasoning circuits.

The stable operation of autonomous off-grid photovoltaic systems requires solar forecasting algorithms that respect atmospheric thermodynamics. Contemporary deep learning models consistently exhibit critical anomalies, primarily severe temporal phase lags during cloud transients and physically impossible nocturnal power generation. To resolve this divergence between data-driven modeling and deterministic celestial mechanics, this research introduces the Thermodynamic Liquid Manifold Network. The methodology projects 22 meteorological and geometric variables into a Koopman-linearized Riemannian manifold to systematically map complex climatic dynamics. The architecture integrates a Spectral Calibration unit and a multiplicative Thermodynamic Alpha-Gate. This system synthesizes real-time atmospheric opacity with theoretical clear-sky boundary models, structurally enforcing strict celestial geometry compliance. This completely neutralizes phantom nocturnal generation while maintaining zero-lag synchronization during rapid weather shifts. Validated against a rigorous five-year testing horizon in a severe semi-arid climate, the framework achieves an RMSE of 18.31 Wh/m2 and a Pearson correlation of 0.988. The model strictly maintains a zero-magnitude nocturnal error across all 1826 testing days and exhibits a sub-30-minute phase response during high-frequency optical transients. Comprising exactly 63,458 trainable parameters, this ultra-lightweight design establishes a robust, thermodynamically consistent standard for edge-deployable microgrid controllers.

We study signal propagation at initialization in transformers through the averaged partial Jacobian norm (APJN), a measure of gradient amplification across layers. We extend APJN analysis to transformers with bidirectional attention and permutation-symmetric input token configurations by deriving recurrence relations for activation statistics and APJNs across layers. Our theory predicts how attention modifies the asymptotic behavior of the APJN at large depth and matches APJNs measured in deep vision transformers. The criticality picture known from residual networks carries over to transformers: the pre-LayerNorm architecture exhibits power-law APJN growth, whereas transformers with LayerNorm replaced by elementwise $\tanh$-like nonlinearities have stretched-exponential APJN growth, indicating that the latter are subcritical. Applied to Dynamic Tanh (DyT) and Dynamic erf (Derf) transformers, the theory explains why these architectures can be more sensitive to initialization and optimization choices and require careful tuning for stable training.

We set out to train behavioral dispositions (self-verification, uncertainty acknowledgment, feedback integration) into small language models (0.6B to 2.3B effective parameters) through a four-stage all-MIT distillation pipeline, with follow-on experiments on inference-time attention-head interventions and a frozen-base confidence-gated sidecar. An internal draft reported +33.9-point MCAS and +15.3-point HumanEval gains on a Qwen3-0.6B student; a second-pass sanity check falsified both numbers before publication. The HumanEval delta was a truncation artifact (n_predict=512) that inverted to -8.0 points at n_predict=1024; the MCAS gain disappeared under apples-to-apples scoring. That falsification triggered three subsequent arcs. Across (1) SFT/DPO LoRA on three model families and two domains, (2) inference-time attention-head tempering on o_proj, and (3) a training-free frozen-base sidecar reading the final-token hidden state h_last, we find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry. The failure is consistent across five models (Qwen3-0.6B, Qwen3-1.7B, Qwen3.5-0.8B, Gemma 4 E2B, and SmolLM2-1.7B-Instruct). A within-distribution cross-validation pass (AUC=0.683) collapsed to chance on fresh prompts (AUC=0.516). We contribute a three-arc negative result with mechanism, a two-failure-mode taxonomy for linear h_last probes, and an honest falsification pipeline that converts the class of false positives we ourselves produced into publishable negatives. As an independent finding, Gemma 4 E2B exhibits near-complete confidence-correctness decoupling on the Chef domain (assertion asymmetry -0.009; the model asserts at 91% regardless of correctness).

In optimization problems, some variable subsets may have a joint non-linear or non-monotonical influence on the function value. Therefore, knowledge of variable dependencies may be crucial for effective optimization, and many state-of-the-art optimizers leverage it to improve performance. However, some real-world problem instances may be the subject of noise of various origins. In such a case, variable dependencies relevant to optimization may be hard or impossible to tell using dependency checks sufficient for problems without noise, making highly effective operators, e.g., Partition Crossover (PX), useless. Therefore, we use Statistical Linkage Learning (SLL) to decompose problems with noise and propose a new SLL-dedicated mask construction algorithm. We prove that if the quality of the SLL-based decomposition is sufficiently high, the proposed clustering algorithm yields masks equivalent to PX masks for the noise-free instances. The experiments show that the optimizer using the proposed mechanisms remains equally effective despite the noise level and outperforms state-of-the-art optimizers for the problems with high noise.

The identification of reliable molecular biomarkers for Parkinson's disease remains challenging due to its multifactorial nature. Although protein sequences constitute a fundamental and widely available source of biological information, their standalone discriminative capacity for complex disease classification remains unclear. In this work, we present a controlled and leakage-free evaluation of multiple representations derived exclusively from protein primary sequences, including amino acid composition, k-mers, physicochemical descriptors, hybrid representations, and embeddings from protein language models, all assessed under a nested stratified cross-validation framework to ensure unbiased performance estimation. The best-performing configuration (ProtBERT + MLP) achieves an F1-score of 0.704 +/- 0.028 and ROC-AUC of 0.748 +/- 0.047, indicating only moderate discriminative performance. Classical representations such as k-mers reach comparable F1 values (up to approximately 0.667), but exhibit highly imbalanced behavior, with recall close to 0.98 and precision around 0.50, reflecting a strong bias toward positive predictions. Across representations, performance differences remain within a narrow range (F1 between 0.60 and 0.70), while unsupervised analyses reveal no intrinsic structure aligned with class labels, and statistical testing (Friedman test, p = 0.1749) does not indicate significant differences across models. These results demonstrate substantial overlap between classes and indicate that primary sequence information alone provides limited discriminative power for Parkinson's disease classification. This work establishes a reproducible baseline and provides empirical evidence that more informative biological features, such as structural, functional, or interaction-based descriptors, are required for robust disease modeling.

The stable operation of off-grid photovoltaic systems requires accurate, computationally efficient solar forecasting. Contemporary deep learning models often suffer from massive computational overhead and physical blindness, generating impossible predictions. This paper introduces the Physics-Informed State Space Model (PISSM) to bridge the gap between efficiency and physical accuracy for edge-deployed microcontrollers. PISSM utilizes a dynamic Hankel matrix embedding to filter stochastic sensor noise by transforming raw meteorological sequences into a robust state space. A Linear State Space Model replaces heavy attention mechanisms, efficiently modeling temporal dependencies for parallel processing. Crucially, a novel Physics-Informed Gating mechanism leverages the Solar Zenith Angle and Clearness Index to structurally bound outputs, ensuring predictions strictly obey diurnal cycles and preventing nocturnal errors. Validated on a multi-year dataset for Omdurman, Sudan, PISSM achieves superior accuracy with fewer than 40,000 parameters, establishing an ultra-lightweight benchmark for real-time off-grid control.

We have witnessed remarkable advances in LLM reasoning capabilities with the advent of DeepSeek-R1. However, much of this progress has been fueled by the abundance of internet question-answer (QA) pairs, a major bottleneck going forward, since such data is limited in scale and concentrated mainly in domains like mathematics. In contrast, other sciences such as physics lack large-scale QA datasets to effectively train reasoning-capable models. In this work, we show that physics simulators can serve as a powerful alternative source of supervision for training LLMs for physical reasoning. We generate random scenes in physics engines, create synthetic question-answer pairs from simulated interactions, and train LLMs using reinforcement learning on this synthetic data. Our models exhibit zero-shot sim-to-real transfer to real-world physics benchmarks: for example, training solely on synthetic simulated data improves performance on IPhO (International Physics Olympiad) problems by 5-10 percentage points across model sizes. These results demonstrate that physics simulators can act as scalable data generators, enabling LLMs to acquire deep physical reasoning skills beyond the limitations of internet-scale QA data. Code available at: https://sim2reason.github.io/.

Reasoning has become a central capability in large language models. Recent research has shown that reasoning performance can be improved by looping an LLM's layers in the latent dimension, resulting in looped reasoning language models. Despite promising results, few works have investigated how their internal dynamics differ from those of standard feedforward models. In this paper, we conduct a mechanistic analysis of the latent states in looped language models, focusing in particular on how the stages of inference observed in feedforward models compare to those observed in looped ones. To this end, we analyze cyclic recurrence and show that for many of the studied models each layer in the cycle converges to a distinct fixed point; consequently, the recurrent block follows a consistent cyclic trajectory in the latent space. We provide evidence that as these fixed points are reached, attention-head behavior stabilizes, leading to constant behavior across recurrences. Empirically, we discover that recurrent blocks learn stages of inference that closely mirror those of feedforward models, repeating these stages in depth with each iteration. We study how recurrent block size, input injection, and normalization influence the emergence and stability of these cyclic fixed points. We believe these findings help translate mechanistic insights into practical guidance for architectural design.

GUI agents drive applications through their visual interfaces instead of programmatic APIs, interacting with arbitrary software via taps, swipes, and keystrokes, reaching a long tail of applications that CLI-based agents cannot. Yet progress in this area is bottlenecked less by modeling capacity than by the absence of a coherent full-stack infrastructure: online RL training suffers from environment instability and closed pipelines, evaluation protocols drift silently across works, and trained agents rarely reach real users on real devices. We present ClawGUI, an open-source framework addressing these three gaps within a single harness. ClawGUI-RL provides the first open-source GUI agent RL infrastructure with validated support for both parallel virtual environments and real physical devices, integrating GiGPO with a Process Reward Model for dense step-level supervision. ClawGUI-Eval enforces a fully standardized evaluation pipeline across 6 benchmarks and 11+ models, achieving 95.8% reproduction against official baselines. ClawGUI-Agent brings trained agents to Android, HarmonyOS, and iOS through 12+ chat platforms with hybrid CLI-GUI control and persistent personalized memory. Trained end to end within this pipeline, ClawGUI-2B achieves 17.1% Success Rate on MobileWorld GUI-Only, outperforming the same-scale MAI-UI-2B baseline by 6.0%.

Automation underpins progress across scientific and industrial disciplines. Yet, automating tasks requiring interpretation of abstract visual information remain challenging. For example, crystal alignment strongly relies on humans with the ability to comprehend diffraction patterns. Here we introduce an autonomous system that aligns single crystals without access to crystallography and diffraction theory. Using a model-free reinforcement learning framework, an agent learns to identify and navigate towards high-symmetry orientations directly from Laue diffraction patterns. Despite the absence of human supervision, the agent develops human-like strategies to achieve time-efficient alignment across different crystal symmetry classes. With this, we provide a computational framework for intelligent diffractometers. As such, our approach advances the development of automated experimental workflows in materials science.

Deep learning underpins a wide range of applications in MRI, including reconstruction, artifact removal, and segmentation. However, progress has been driven largely by public datasets focused on brain and knee imaging, shaping how models are trained and evaluated. As a result, careful studies of the reliability of these models across diverse anatomical settings remain limited. In this work, we introduce MosaicMRI, a large and diverse collection of fully sampled raw musculoskeletal (MSK) MR measurements designed for training and evaluating machine-learning–based methods. MosaicMRI is the largest open-source raw MSK MRI dataset to date, comprising 3,150 volumes and 89,830 slices, including a 3,096-volume 1.5~T core dataset and a 54-volume 0.55~T low-field set. The dataset offers substantial diversity in volume orientation (e.g., axial, sagittal), imaging contrasts (e.g., PD, T1, T2), anatomies (e.g., spine, knee, hip, ankle, and others), numbers of acquisition coils, and field strengths. Using accelerated reconstruction as a testbed, we study scaling, robustness, and data selection under realistic MSK distribution shifts. Across E2E-VarNet, U-Net, and ViT baselines, mixed-anatomy training improves over anatomy-specific training, showing that the benefit of anatomical diversity is not architecture-specific. Zero-shot experiments on the 0.55~T low-field set highlight that using more training data improves average reconstruction under field-strength shift. Controlled subset experiments further suggest that these gains come from transferable cross-anatomy structure, rather than anatomy or contrast matching alone.

Continuous diffusion has served as a foundation for high-fidelity, controllable, and few-step generation across continuous data modalities such as images, videos, and molecular structures. However, in language modeling, prior continuous diffusion language models (DLMs) lag behind discrete counterparts. Existing categorical and simplex-based approaches operate over extremely large and sparse language spaces, while prior embedding-space approaches avoid this sparsity but lack a well-selected design space. In this work, we close this gap with LangFlow by connecting embedding-space DLMs to Flow Matching, alongside three key innovations: (1) we derive a novel ODE-based NLL bound for principled evaluation of continuous flow-based language models; (2) we propose an information-uniform principle for setting the noise schedule, which motivates a learnable noise scheduler based on a Gumbel distribution; and (3) we revise prior training protocols by incorporating self-conditioning, which improves both likelihood and sample quality for embedding-space DLMs and behaves differently from its use in discrete diffusion. Putting everything together, LangFlow is competitive with top discrete DLMs on generative quality, reaching a generative perplexity of 92.3 on LM1B and 36.5 on OpenWebText. It also outperforms discrete counterparts on reasoning tasks like Sudoku. As the first continuous DLM shown to rival discrete diffusion in generative quality, LangFlow provides clear evidence that continuous diffusion is a promising paradigm for language modeling. Code: https://github.com/nealchen2003/LangFlow.

Kullback-Leibler (KL) divergence is a fundamental concept in information theory that quantifies the discrepancy between two probability distributions. In the context of Variational Autoencoders (VAEs), it serves as a central regularization term, imposing structure on the latent space and thereby enabling the model to exhibit generative capabilities. In this work, we present a detailed derivation of the closed-form expression for the KL divergence between Gaussian distributions, a case of particular importance in practical VAE implementations. Starting from the general definition for continuous random variables, we derive the expression for the univariate case and extend it to the multivariate setting under the assumption of diagonal covariance. Finally, we discuss the interpretation of each term in the resulting expression and its impact on the training dynamics of the model.

Using behavioural science, health interventions focus on behaviour change by providing a framework to help patients acquire and maintain healthy habits that improve medical outcomes. In-person interventions are costly and difficult to scale, especially in resource-limited regions. Digital health interventions offer a cost-effective approach, potentially supporting independent living and self-management. Automating such interventions, especially through machine learning, has recently gained considerable attention. Ambivalence and hesitancy (A/H) play a primary role for individuals to delay, avoid, or abandon health interventions. A/H are subtle and conflicting emotions that place a person in a state between positive and negative evaluations of a behaviour, or between acceptance and refusal to engage in it. They manifest as affective inconsistency across modalities or within a modality, such as language, facial, vocal expressions, and body language. While experts can be trained to recognize A/H, integrating them into digital health interventions is costly and less effective. Automatic A/H recognition is therefore critical for the personalization and cost-effectiveness of digital health interventions. Here, we explore the application of deep learning models for A/H recognition in videos, a multi-modal task by nature. In particular, this paper covers three learning setups: supervised learning, unsupervised domain adaptation for personalization, and zero-shot inference via large language models (LLMs). Our experiments are conducted on the unique and recently published BAH video dataset for A/H recognition. Our results show limited performance, suggesting that more adapted multi-modal models are required for accurate A/H recognition. Better methods for modeling spatio-temporal and multimodal fusion are necessary to leverage conflicts within/across modalities.

General first-order methods (GFOM) are a flexible class of iterative algorithms which update a state vector by matrix-vector multiplications and entrywise nonlinearities. A long line of work has sought to understand the large-n dynamics of GFOM, mostly focusing on "very random" input matrices and the approximate message passing (AMP) special case of GFOM whose state is asymptotically Gaussian. Yet, it has long remained unknown how to construct iterative algorithms that retain this Gaussianity for more structured inputs, or why existing AMP algorithms can be as effective for some deterministic matrices as they are for random matrices. We analyze diagrammatic expansions of GFOM via the limiting traffic distribution of the input matrix, the collection of all limiting values of permutation-invariant polynomials in the matrix entries, to obtain the following results: 1. We calculate the traffic distribution for the first non-trivial deterministic matrices, including (minor variants of) the Walsh-Hadamard and discrete sine and cosine transform matrices. This determines the limiting dynamics of GFOM on these inputs, resolving parts of longstanding conjectures of Marinari, Parisi, and Ritort (1994). 2. We design a new AMP iteration which unifies several previous AMP variants and generalizes to new input types, whose limiting dynamics are Gaussian conditional on some latent random variables. The asymptotic dynamics hold for a large and natural class of traffic distributions (encompassing both random and deterministic input matrices) and the algorithm's analysis gives a simple combinatorial interpretation of the Onsager correction, answering questions posed recently by Wang, Zhong, and Fan (2022).

Governing common-pool resources requires agents to develop enduring strategies through cooperation and self-governance to avoid collective failure. While foundation models have shown potential for cooperation in these settings, existing multi-agent research provides little insight into whether structured leadership and election mechanisms can improve collective decision making. The lack of such a critical organizational feature ubiquitous in human society presents a significant shortcoming of the current methods. In this work we aim to directly address whether leadership and elections can support improved social welfare and cooperation through multi-agent simulation with LLMs. We present our open-source framework that simulates leadership through elected personas and candidate-driven agendas and carry out an empirical study of LLMs under controlled governance conditions. Our experiments demonstrate that having elected leadership improves social welfare scores by 55.4% and survival time by 128.6% across a range of high performing LLMs. Through the construction of an agent social graph we compute centrality metrics to assess the social influence of leader personas and also analyze rhetorical and cooperative tendencies revealed through a sentiment analysis on leader utterances. This work lays the foundation for further study of election mechanisms in multi-agent systems toward navigating complex social dilemmas.