Papers

Does representation learning stop when the training loss stops improving? We study this question for matrix Muon, whose polar-normalised updates have a step length set by the gradient's rank rather than its norm. Near the edge of stability, full-batch Muon on teacher-student problems enters approximately period-2 loss oscillations that persist for thousands of steps: the cycle-mean loss stays flat or rises, yet the weights keep moving and the learned features continue to align with the teacher subspace. For linear teacher-student learning toys, we derive explicit cycle and alignment formulas and conditional plateau and decay bounds. For a population mean-field ReLU model, we prove that, under stated dimension, initialisation and small-head conditions, the leading eigenspace of the average gradient outer product (AGOP) recovers the teacher subspace exactly during a loss plateau, before the loss later drops. In all 33 ReLU, GELU and SiLU teacher configurations we study, direction-only alignment metrics show the student AGOP aligned with, or still aligning to, the teacher subspace during the period-2 oscillations; projected head refitting on selected configurations shows that the learned directions are useful for prediction, and further measurements distinguish AGOP alignment from weight-mass concentration. In deep residual ReLU students, freezing the downstream layers while the first layer trains with full-batch exact polar updates recreates a nearly flat cycle-mean loss with improving input-AGOP alignment; freezing and unfreezing switch between this plateau and loss decrease, and the effect is sensitive to momentum and to the choice of orthogonaliser.

Many learning tasks require stability to small transformations while retaining sensitivity to larger ones. We introduce transversal pooling neural networks, which generalize spatial max pooling to affine group actions. We establish equivariance to a chosen subgroup and derive explicit stability bounds for individual pooled wavelet coefficients under affine perturbations of the input. Experimentally, we demonstrate the utility of our networks in low-data environments and for predicting tropical cyclone intensification.

Major depressive disorder (MDD) severely impacts daily activities and quality of life. Detecting MDD involves multimodal data, such as interview recordings and sensor measurements. This is particularly challenging, as these heterogeneous modalities often demand distinct, customized prediction pipelines. Existing efforts to address this challenge have explored both manually engineered multimodal architectures and agent-assisted pipeline development. Despite their progress, it remains challenging to autonomously revise pipelines based on experimental feedback and carry verified improvements forward into subsequent designs. To this end, we propose Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis (MERID). The framework develops depression pipelines through experience-based recursive self-improvement (RSI). Grounded State Construction (GSC) grounds experience by aligning multimodal records with subject-level depression targets. Coupled Pipeline Exploration (CPE) jointly modifies representations, fusion, and predictors to build successor pipelines for classification and severity estimation. Evidence-Guided Evolution (EGE) guides revisions through feedback and verifies gains under uncertainty in small depression cohorts before inheritance. Extensive experiments on depression benchmarks show that MERID achieves the best results on multiple tasks compared with multimodal and agent-based baselines. Further analysis highlights the value of acoustic and linguistic cues for depression detection. Our code is available at https://github.com/DiscoAILab/MERID

We study the dynamical behavior of tokens in transformers from a control-theoretic perspective. Our model includes the feed-forward layer present after the self-attention mechanism, with the self-attention mechanism interpreted as an interacting particle system and the feed-forward layer as an independent control. Our main theoretical result establishes that the feed-forward network can steer the tokens arbitrarily close to consensus regardless of the key, query, and value matrices. Our result are easily extended to convergence to many clusters and to multi-head attention. We conduct numerical experiments to verify our results, and compare thresholding behavior from our theory to real-world LLMs.

Next-latent prediction fits a map from the current embedding to the next one. LeNEPA carries this objective to time series, replacing the stop-gradient of next-embedding prediction with the isotropy penalty of LeJEPA. A world model is a transition kernel that can be rolled out. The one-step regression identifies a conditional mean, and a mean is a kernel only in special cases. For a linear-Gaussian Markov latent, the mean transition and the innovation covariance are fixed by the one-step problem, and the open-loop squared error at horizon $K$ equals the trace of the sum of the pushed-forward innovation covariances. That error grows with $K$ after the one-step fit is exact. If the conditional mean is nonlinear, composing it is not the multi-step conditional mean. If the observation is a non-injective function of a Markov state, a memoryless one-step map does not determine future observations, while a short window can. An isotropy penalty is a function of the embedding marginal, so its partial derivative in the transition weights is zero. On a scalar autoregression with coefficient $0.9$, the one-step mean squared error is $0.998$ and the $16$-step open-loop error is $5.10$. On a hidden rotation, an eight-step window reaches $16$-step error $0.056$, while the current scalar alone reaches $0.778$. Raising the isotropy weight from $0.1$ to $10$ leaves eight-step latent error inside $[0.78,0.85]$ on three seeds.

Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferring model weights from accelerator high-bandwidth memory into on-chip SRAM. Mixture-of-experts (MoE) models reduce computation by activating only a small subset of experts per token, but this sparsity does not translate directly to batched decoding. Different requests select different experts; therefore, the combined active set across many concurrent requests can span a substantial fraction of the expert pool and require significantly more expert weights to be transferred. Most expert-reduction techniques make retention decisions independently for each token and therefore do not address this batch-level expansion. More recently, batch-aware methods have attempted to coordinate expert use across concurrent requests and reuse experts already fetched for the batch. Yet their selection criteria are based primarily on router rankings or expert statistics collected during calibration. Consequently, these criteria are not directly tied to the output error caused by dropping an expert, nor do they capture how an expert's contribution changes across tokens at inference time. We instead rank experts according to how much their removal would change the MoE-layer output. To apply this criterion during serving, we train a lightweight linear predictor during calibration that estimates the expert removal cost for each incoming token, and develop custom GPU kernels for cost prediction and expert selection. Across three MoE architectures, BASE improves the quality-efficiency tradeoff without retraining. On Qwen3-30B-A3B, it improves average accuracy by 29.5 points over the strongest baseline at comparable throughput under a tight expert budget. At a higher expert budget, it is 60% faster than dense inference while remaining within 0.4 accuracy points.

The way a model distributes activity over each layer's attention heads offers a coarse view of how it routes information through depth; how this changes with the task is part of what a mechanistic account must explain. Holding prompt length fixed, we vary how many serial steps a task demands and measure, in every layer of 17 open-weight models, whether activity concentrates on a few heads or spreads across many as demand rises. Both occur: in most models, layers just before mid-depth concentrate activity and later layers spread it. Models differ in where and how strongly this happens. The Qwen2.5 base models from 0.5B to 7B, for example, spread less than the average model in every task and concentrate activity in parts of their second half, where Llama models from 1B to 8B and OLMo-2 spread; the contrast largely holds between Llama-3.1-70B and Qwen2.5-72B, which have the same number of layers and heads. These differences are reproducible, and post-trained models keep much of their base model's pattern. An ablation study suggests that, within a task, models whose activity is more concentrated on their top heads also depend more on those heads for the answer. Concentration and spreading across layers thus offer a new way to compare models, by how they route information through depth. Code is available at https://github.com/johnnyjli/serial-demand-heads.

Large language models are increasingly evaluated in specialized domains such as law, medicine, software engineering, and cybersecurity, yet film remains comparatively underexplored despite requiring long-form narrative integration, multilingual interpretation, and culturally situated audience judgments. We introduce CineSubBench, a benchmark for evaluating long-context film understanding from multilingual movie subtitles. A subtitle track represents a film as thousands of short, temporally ordered utterances from which models must reconstruct characters, relationships, events, causal progression, and themes without explicit scene or event structure. CineSubBench contains 1,012 films with complete subtitle coverage in six languages, yielding 6,072 tracks and 8.13M timestamped subtitle entries. It provides a matched multi-task, multilingual, and multicultural (MultiX) evaluation setting: seven tasks span narrative reconstruction and abstraction, genre prediction, age suitability, country-specific motion-picture ratings across ten national classification systems, and subtitle-grounded language safety. Across nine LLMs, plot premises are recovered more reliably than event-complete synopses; cross-lingual consistency varies substantially across models and languages; national rating systems expose distinct calibration patterns; and strong profanity is far easier to ground than mild obscenity. CineSubBench establishes film as a long-context LLM evaluation domain and provides a unified benchmark for measuring narrative, multilingual, cultural, and evidence-grounding capabilities.

Neural operators are typically trained in a supervised fashion, which requires a dataset to be generated with a classical solver. Training them physics-informed, i.e., purely from the governing equations, removes this large offline cost and allows fresh samples to be drawn at every optimization step, but has so far been limited to simplified problems and trails supervised training in accuracy. The obstacle is the ill-conditioning of physics-informed losses, which differential operators induce and which worsens as the discretization is refined. We therefore propose a preconditioned residual loss function and show mesh-independent conditioning for elliptic problems and greatly improved conditioning for saddle point problems. Realized through geometric and algebraic multigrid, the construction applies to linear and nonlinear equations, steady or time-dependent, on structured and unstructured meshes, is agnostic to the neural operator architecture, and adds no cost at inference. On the Poisson, Allen-Cahn and stationary Stokes equations, the resulting label-free training matches supervised training and is four to twenty-five times more accurate than previous physics-informed operator learning methods.

Power-grid digital twins must combine data-driven prediction with physically meaningful state evolution while preserving the provenance of environmental observations. This paper presents an early-stage EnergyEminence testbed that couples an IEEE 118-bus-style graph-temporal predictor, nonlinear AC cascade simulation, and operator-dashboard-like temporal replay. In addition, we introduce a shared bounded calibration that converts wildfire-detection confidence and spatial extent into source-comparable wildfire interpretable and explainable evidence. We then evaluate it with visually diverse fire and hard-negative videos. Sixteen synthetic environmental videos are curated to generate 160 source-tracked grid scenarios, and a source-video-disjoint test yields 10 true positives, 8 false positives, 22 true negatives, and no false negatives. The errors occur in stressed, non-cascading scenarios conditioned on an unseen growing-fire source. Our diagnostic then reveals environmental shortcut learning that is obscured by scenario-level random splitting. The paper therefore contributes a data-centric and inspectable evaluation workflow for multimodal grid-resilience models, together with evidence supporting separation of environmental alerting from electrical cascade inference

Generative denoising models, such as diffusion and flow-matching, learn to sample from complex distributions by training a deep neural network denoiser to recover clean data from noise-corrupted samples. While such models are typically compared on the quality of their synthesized samples, these metrics provide limited insight into how the underlying denoiser, which drives generation, differs. In this work, we propose to analyze the spectrum of the denoiser Jacobian as a tool to characterize these differences. Across pre-trained denoising models, we observe that better generative performance is associated with larger Jacobian eigenvalues. Motivated by this, we introduce a regularization scheme that controls the Jacobian spectrum by training the denoiser on perturbed inputs, with perturbations suppressing or amplifying Jacobian responses. On ImageNet, we test whether directly modifying the Jacobian spectral properties leads to improved generations. Our findings suggest that denoisers benefit from both strengthening responses along data-relevant principal eigen-directions and suppressing the noisy, data-irrelevant ones. This establishes the denoiser Jacobian as a useful tool for identifying differences between generative denoising models.

Input encodings can restrict which measured contrasts a predictor can jointly reproduce, even when no single contrast is forced to vanish. We compute the attainable contrast space from an encoder's equivalence classes and a fixed contrast design, without labels, loss, or a fitted model; projecting the recorded contrasts onto that space gives an empirical error floor for any unrestricted decoder on those classes. On a 140-rectangle siRNA interaction panel, a graph neural network's training-only feature mask merges 165 endpoint states into 90 classes and cuts the rank of the 140 interaction contrasts to 72. The resulting floor is 0.009980, which is 14.6% of the fitted model's interaction squared error; the fitted model reaches 0.068335, slightly worse than a control predicting no interaction at all. A minimum of three restored chemistry columns recovers full rank. Refitting without the mask removes the floor entirely, yet interaction MSE improves by only 0.000017 under the reported protocol, and the restored columns remain absent from every training input. On a released RNA-splicing predictor, whose encoding is injective on the measured states, the same computation returns the full design rank of 1,986 and a floor of exactly zero. These results separate what an encoding permits from what a fitted model achieves; they do not identify what limits the remaining error. The rank check needs no fits and bounds what any amount of training under a fixed encoding can recover. The project repository is available at https://github.com/shadi97kh/REPRESENTABLE-BUT-UNLEARNED.

Gradient-based reward guidance provides a flexible way to use downstream reward models to control masked diffusion language models at inference time. However, its computational cost remains high as each decoding iteration incurs expensive diffusion model forward passes and reward model backpropagation steps. To address this, we introduce FastGuide, an adaptive hybrid of parallel and autoregressive decoding to accelerate reward guidance for diffusion language models. In analogy to parallel decoding, FastGuide amortizes the cost of reward model backpropagation by computing guidance once per decoding step and reusing it to generate multiple tokens. Within each decoding step, FastGuide makes diffusion forward passes autoregressive by unmasking tokens one at a time while efficiently recomputing token distributions after each unmasking by utilizing KV caching techniques and sparse recomputation of attention. Lastly, to adapt hybrid decoding to the model's confidence, FastGuide defers any token that the model is unconfident about under its recomputed distribution. Experiments on three reward benchmarks demonstrate that FastGuide is up to $4.4\times$ faster than sequential reward-guided decoding while retaining similar generation quality.

Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are then scored and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.

Learning from scientific measurements often requires aligning modalities with different spatial support and resolution. Subsurface characterization is an extreme dense-sparse problem in which 3D seismic provides volumetric but indirect measurements and well logs provide high-resolution 1D measurements at sparse borehole locations. We introduce a physically grounded cross-modal masked autoencoder (CM-MAE) for seismic-to-well representation learning. The model jointly tokenizes seismic volumes and well-log depth patches, embeds both modalities in continuous physical coordinates using four-axis rotary position embeddings, and reconstructs masked targets with a cross-modal decoder. Sparse Mixture-of-Experts layers provide modality-specific capacity while dense attention allows information exchange between modalities. Pretraining uses 23 seismic volumes covering approximately 178,000 square km across U.S. offshore and onshore basins, together with approximately 92,000 wells. Matched-mask ablations show strongly asymmetric information flow. Seismic context improves held-out well-log reconstruction by 9.73%, while well-log context improves seismic reconstruction by 0.65%. We evaluate seismic-only pseudo-log predictions against independent interpreter-drawn salt-geobody masks to determine whether they contain geologic signal. Across offshore U.S. surveys, compressional-slowness-derived salt scores reach an AUROC of 0.910. In onshore basins, predicted compressional slowness preserves formation-scale structure and partial relative organization in an unseen survey despite calibration drift. CM-MAE learns useful seismic-to-well representations under extreme modality asymmetry, although absolute pseudo-log calibration remains survey-dependent.

Symbolic Regression (SR) is a data-driven method for scientific discovery which searches for interpretable analytical relationships within data. Recently, Large Language Models (LLMs) have also had a significant impact on scientific discovery, enabling the automation of various stages of the process. For these reasons, the possibility of harnessing the embedded scientific knowledge and programming capabilities of LLMs to solve SR tasks has emerged, showing promising performance compared with traditional methods. We propose EvoMO-SR, a novel LLM-driven SR framework in which the LLM generates equation skeletons, with their coefficients fitted separately by an external optimizer. The framework includes a multi-objective survival selection which controls bloating by balancing accuracy and complexity, and a substructure guidance mechanism which mutates expressions with candidate reusable building blocks. EvoMO-SR achieves the best accuracy in seven of the eight in-domain and out-of-domain settings for LSR-Synth, using a small LLM model, i.e., Llama-3.1-8B-Instruct. We also evaluated structural recovery through two symbolic accuracy metrics based on canonicalized subtree overlap and term matching, showing that our method has a greater probability of recovering highly accurate symbolic structures.

We consider fair resource allocation in sequential decision-making environments modeled as major-minor weakly coupled Markov decision processes (M2WCMDP). In this framework, resource constraints couple the action spaces of a major sub-Markov decision process (sub-MDP) and a population of minor sub-MDPs that would otherwise operate independently. Instead of using the traditional utilitarian (total-sum) objective, we optimize a general class of monotone, concave, permutation-invariant, normalized fairness functions. With homogeneous minor sub-MDPs, we prove that the problem under symmetry reduces to optimizing the platform-plus-mean-participant utilitarian objective over the class of permutation-invariant policies, which allows us to exploit efficient algorithms that optimize the utilitarian-based objective to solve this fairness-aware problem. For more general settings, we introduce a count-proportion-based deep reinforcement learning approach with a priority-based sampler that generates feasible count actions. The generality of our framework means that the proposed algorithms and theoretical guarantees transfer to any domain with a symmetric M2WCMDP structure. We consider two applications: the machine replacement problem and the joint control of pricing and taxi relocation problem on a New York City-calibrated dataset. We validate our theoretical findings with comprehensive experiments, confirming the effectiveness of our proposed method in achieving strong fairness-aware performance while remaining scalable.

Parallel speculative drafting generates multiple candidates in one backbone pass, but independent token selection can produce inconsistent continuations that shorten the accepted prefix. Existing methods mostly leave conditional decoding to a lightweight module after the backbone, which limits the flow of predecessor information to successors. Our analysis of DFlash shows that early positions already form recoverable predictions in shallow layers, and that accurate adjacent predecessors help successors more when they enter earlier. We therefore propose DSpine, a drafter with causal conditioning injection throughout the backbone: at every layer, gated adjacent injection writes each predecessor's predicted feature into its successor, so the causal conditioning chain unfolds over network depth while all positions update in parallel. A unified transfer space built from the target model's output embeddings unifies layer-wise injection with predecessor-conditioned decoding, and layer-wise output-embedding supervision promotes the formation of predicted features in shallow layers. Fused kernels and a transition cache execute both efficiently in parallel within SGLang. Across seven math, code, and chat benchmarks, DSpine achieves the longest acceptance length at both temperatures on Qwen3-4B and Qwen3-8B. At temperature zero on Qwen3-8B, it raises the seven-benchmark mean from DFlash's 3.77 to 4.82 (+27.8%); in SGLang serving tests, it delivers 23.3% higher throughput than DFlash on average. Project page: https://dspine-project.github.io

Figuring out how objects relate to each other, like whether they touch, overlap, stay completely separate or one sits inside another, matters a lot in fields like GIS, biomedical imaging, and robotics. Even though machine learning has come a long way, people haven't really focused on spotting these topological relationships in images. The main roadblocks? Not enough good datasets and no clear way to measure results. So, we rolled up our sleeves and built a new dataset. It's pretty sizable: over 11,000 labelled images showing all those essential relationships. We ran tests with some classic machine learning models, Naive Bayes, KNN, Random Forest, SVM, and Artificial Neural Networks, and threw in some deep learning stars like VGG16 and InceptionResNetV2. For the dataset itself, we used segmentation, contour detection, and grayscale normalization to tease out solid feature vectors. The results? Deep learning methods, especially VGG16, pulled ahead, with validation accuracy hitting 89.55%. That's a big jump compared to the traditional models. This shows how powerful transfer learning is for analyzing topological relationships in images, and it gives researchers a new standard to aim for in future work on spatial reasoning and topological classification.

In this work, we develop a second-order optimization framework for physics-informed neural networks (PINNs) applied to high-dimensional parametric partial differential equations (PDEs). The framework is built on the Gauss–Newton pullback metric, which provides an operator-informed notion of curvature in parameter space and connects the method to the broader family of natural gradient schemes. We show that, for coordinate-separable neural architectures and linear differential operators (or linearized operators in the nonlinear case) admitting a finite separable representation, the residual Jacobian inherits a structured separable factorization. This yields an exact compressed formulation of the Gauss–Newton step in a reduced space, without assembling the full residual Jacobian on the exponentially large tensor-product collocation grid. The dimension of the reduced space (the effective compressed dimension) is determined by the local collocation grid sizes, the separable operator structure, and the contraction pattern of the architecture, thereby avoiding dependence on the full tensor-product grid size and replacing dense linear algebra in parameter space by a substantially smaller structured problem. Within our framework, we investigate canonical polyadic and tensor-train parametrizations and derive their full algebraic characterization relevant to the Gauss–Newton method, including the structure of the residual Jacobian, the resulting compressed system, and its effective compressed dimension. Numerical experiments on high-dimensional PDEs, including parametric problems, demonstrate the high efficiency of the proposed compressed Gauss–Newton method, which achieves substantially lower errors than tensor-compressed first-order baselines with orders of magnitude fewer iterations and only a fraction of the computing time.

Large language models can reason in continuous space instead of decoded text, by recurring on their own hidden states or by passing those states between agents, while training supervises only the Cross-Entropy (CE) of the final decoded answer and does not constrain the thought. Theoretical and empirical analyses establish and confirm four failures of CE-only training that lead to a lower probability of the correct answer such as collapsing thoughts across distinct questions and retaining irrelevant information. We introduce REST (REpresentation-Supervised Thoughts), a training objective that turns four properties of a valid thought representation (causality, minimality, separability, and stability) into differentiable losses added to CE. We instantiate it in latent single-agent and multi-agent systems, without architectural changes or added parameters at inference. Across 7 benchmarks spanning mathematics, science, medicine, and code generation, with the same training data, compute, and latent budget, REST increases accuracy over CE-only training across agent settings and model sizes by up to 7.5 percentage points and convergence on a final answer by 30%. Furthermore, REST thoughts encode more of what is required to achieve the correct answer, and decoding them better recovers the intended output of the agent, which makes latent communication easier to interpret. Project Website: https://fard-lab.github.io/REST

Most federated learning frameworks for vehicular ad hoc networks assume that all vehicles collaboratively train a single model for a common task. This assumption limits their applicability to practical vehicular environments, where vehicles may perform heterogeneous but related perception tasks with different output spaces. This paper proposes encoder-sharing hierarchical multi-task federated learning (EN-HMTFL), which integrates cluster-based hierarchical federated learning with a globally shared encoder and vehicle-local decoders. EN-HMTFL enables vehicles performing different tasks to collaboratively learn a transferable feature representation while preserving their task-specific models locally. Only the encoder is exchanged and aggregated through the hierarchy, whereas raw data and local decoder parameters remain at the vehicles. The proposed framework is evaluated on the MNIST and GTSRB datasets in different vehicular scenarios. Across the evaluated scenarios, EN-HMTFL improves accuracy by up to 24.0% relative to the compared representation-sharing benchmark. In scenarios where EN-HMTFL converges earlier, the reduction reaches up to 69 communication rounds (28.8%).

General-purpose models promise sensor-based decisions without training a task-specific classifier, which could reduce the dependence of Human Activity Recognition (HAR) on labeled data. Yet it remains unclear whether such models can directly interpret deterministic descriptions of physical sensor signals well enough to replace or complement trained HAR models. We study this question using Jev, a fixed general-purpose probabilistic decision model, on 1,800 class-balanced accelerometer windows from WISDM, UCI341, and PAMAP2. Jev receives no labeled examples, retrieval context, or HAR-specific parameter updates. We evaluate three deterministic sensor representations and compare 5,400 Jev decisions with a generative baseline and three supervised HAR models. Jev remains far below supervised recognition, with its strongest representation reaching macro-F1 of 0.038, 0.118, and 0.089 across the three datasets, compared with 0.686 to 0.907 for the supervised models. More numerical features do not improve Jev. Instead, they reduce recognition on all three datasets, while augmenting the same numerical evidence with a deterministic semantic rendering partially recovers performance, although the experiment does not isolate semantics from the accompanying serialization and redundancy changes. Jev is fast and inexpensive to query, but its probabilities are not reliably calibrated for recognition. A post-hoc fusion analysis finds a small improvement on WISDM that does not replicate on UCI341 or PAMAP2. These results show that training-free sensor decisions depend not only on the information available in the signal, but also on whether the model can use the representation through which that information is exposed. The sensor-to-model interface should therefore be treated as part of the model evaluation rather than as a neutral preprocessing step.

B2B advertising targets a viewer's professional attributes (employer size and industry, function, seniority) and has obtained them by matching identities across sites. Safari and Firefox block third-party cookies, Google retired the Privacy Sandbox cohort APIs in 2025, and reverse-IP firmographics decay under remote work. We present SIF (Sealed Inference Frame), which infers coarse professional cohorts on the device and emits only a locally differentially private, taxonomy-coded label into the OpenRTB bid stream, with no cross-site identifier. It rests on a property of the web platform we make precise: a navigated cross-origin iframe is the only way third-party code obtains a policy it controls, so inference runs in WebAssembly even where the publisher's CSP forbids it, and a nested worker served with default-src 'none' gives the model no network. Even a malicious model leaks at most about 5 bits per site per week. Labels pass through a memoised k-ary randomised response keyed to the publisher's first-party identifier, which gives $\varepsilon$-local differential privacy, defeats averaging, and links requests no better than the identifier already sent. An org-conditional k-anonymity rule suppresses cells, more strictly on corporate networks than at home. Cohorts ride OpenRTB user.data in a LinkedIn-aligned taxonomy, and attribution uses LinkedIn's click-scoped li_fat_id without bridging identities. We report a crawl of CSP deployment on 7,969 top sites and 431 B2B publishers, Heavy-Ad budgets, closed-form privacy-utility trade-offs, a re-identification simulation, and an assessment of which attributes are predictable at all: company type and size are, seniority largely is not. On-device is a design property, not a consent exemption.

Electronic health records provide irregular observations of latent patient states that evolve continuously over time. Recent autoregressive models condition on clinical histories to forecast future events as sequences of discrete observations. Conversely, multi-marginal flow matching provides a continuous-time formulation, but using multiple observations to supervise training paths does not itself give the learned dynamics access to preceding patient history. We introduce EHRFlow, a multi-marginal flow-matching framework that conditions on encoded patient history, thereby allowing future dynamics to depend on the patient's prior clinical trajectory. Our proposed framework accommodates irregular observation times and supports forecasting at arbitrary horizons. Across controlled synthetic benchmarks, EHRFlow improves clinical-code forecasting and latent-state recovery. On real-world clinical datasets comprising more than one million patients, including an independent external validation cohort, EHRFlow improves horizon-averaged top-5 clinical-code accuracy over autoregressive and history-independent flow-matching baselines. Finally, in a controlled counterfactual simulation, conditional guidance approximates the known effect of an antihypertensive intervention without training a task-specific outcome model.

Functional Kolmogorov-Arnold Networks (FunKAN) achieve state-of-the-art accuracy on MRI Gibbs artifact removal and anatomical segmentation, but their 11.6 M parameters and 8.7 GFLOPs are too large for edge medical devices. We present FunKANLite, a two-stage, hardware-aware compression of FunKAN for point-of-care use. FunKANLite-TR reduces the spatial prior and replaces the ResBlock offset predictor with a depthwise-separable block. It has 1.9x fewer parameters than FunKAN and no loss in accuracy. We then distill FunKANLite-TR into FunKANLite-ST, which lowers the Hermite basis rank, factorizes the spatial prior into a low-rank form, and halves the filter widths. FunKANLite-ST has 5.6x fewer parameters and 3.7x fewer GFLOPs than FunKAN. It stays within 1.4 percentage points IoU of FunKAN on BUSI, GlaS, and CVC-ClinicDB, and reaches 33.95 dB PSNR on IXI. On an NVIDIA Jetson Orin Nano and a Raspberry Pi 5, FunKANLite-ST reduces energy per inference by up to 68% and raises throughput by 2.9x.

Modern AI architectures used to solve visual reasoning tasks typically rely heavily on global connectivity and synchronization. As biological systems demonstrate, though, sophisticated computation can be performed in a more decentralized fashion. In this work, we test the reasoning capabilities of Neural Cellular Automata (NCAs), networks of recurrent cells that use strictly local connectivity and asynchronous updates. NCAs have been extensively studied in artificial life experiments, but it is unclear whether they can perform complex multi-step reasoning. We show that NCAs produce spatio-temporal dynamics capable of solving challenging visual reasoning tasks, including large mazes, Sudoku, and ARC-AGI-1. Furthermore, we provide evidence that NCAs generalize out-of-distribution when running with larger grids, longer rollouts, or parallel trials; and that the latter can be made more efficient via pruning of redundant trajectories. We find that these generalization capabilities depend on training with sample replay and stochastic perturbations, and that stochasticity remains beneficial at test time. Finally, we show that NCAs are robust reasoners capable of dynamically modulating compute to recover efficiently from damage, and that they can scale to solve reasoning in raw pixel space.

Learned rotations play an important role in enabling low-bit weight and activation quantization of large language models by smoothing outliers in the activation distribution. State-of-the-art approaches include gradient-based procedures such as SpinQuant and computationally friendlier gradient-free approaches such as DartQuant, but both remain hard to scale to the largest architectures. To address the computational bottlenecks in gradient-free rotation learning, we introduce two ideas for efficiency, (i) a data selection procedure which reduces the required number of calibration data points, and (ii) an exact reduction of the associated optimization on this reduced calibration set. Our data selection procedure exploits the geometric structure of the convex hull of the activations. Using this idea, we show that a carefully selected calibration set with several orders of magnitude fewer activations than state-of-the-art rotation-based methods can match their performance in low-bit quantization settings. Under this extreme data efficiency, the selected activations span an $r$-dimensional subspace with $r<d$, making optimization over a $d\times d$ rotation equivalent to optimizing a $d\times r$ matrix on the Stiefel manifold. We solve this reduced problem using an efficient ADMM algorithm that iteratively employs thin matrix updates at every step, hence the name ThinQuant. For Llama-3-70B with W4A4KV4 quantization, ThinQuant completes the entire rotation calibration in under 12 minutes and achieves a WikiText-2 perplexity of 5.63, compared with 7.55 for DartQuant, which requires 111 minutes. Unlike SpinQuant and DartQuant, ThinQuant also scales to Llama-3.1-405B on a single H200 GPU, completing rotation calibration in just over 2 hours and achieving WikiText-2 perplexity of 2.97 at W4A4, compared with 3.48 for GPTAQ+QuaRoT.

Securing modern AI systems against backdoor attacks remains an open challenge and requires fundamentally principled estimates of the adversary's budget – the poison fraction $\pi$ and trigger strength $\alpha$ needed to construct successful yet stealthy attacks. Motivated by recent empirical evidence that poisoning large language models can require a nearly constant number of malicious samples even as clean datasets grow, we derive an exact closed-form analysis of a quadratic neuron trained on a poisoned Gaussian mixture. We show, perhaps counterintuitively, that the same feature-learning dynamics that make neural networks powerful can also make them more vulnerable to backdoors. Specifically, with clean accuracy preserved to first order, $O(\pi)$, we demonstrate that lazy learning imposes the inverse-square-root scaling $\alpha \propto \pi^{-1/2}$ for a successful attack, while feature learning induces a quadratic detector whose loss margin scales as $O(\alpha^4)$, improving the attack budget to $\alpha \propto \pi^{-1/4}$. Consequently, nonlinear feature learning substantially reduces the trigger strength required at small poison fractions, thereby in a sense making feature learners more backdoor vulnerable. These results provide a theoretical mechanism consistent with large-scale empirical observations and demonstrate that security audits based on linear heuristics can systematically underestimate backdoor vulnerability in the widely adopted feature-learning regimes.

We study how to build more capable general-purpose agents by extending large language models (LLMs) with native typed decision-making. We introduce Dyad, an architecture that augments a pretrained LLM with an environment-conditioned action encoder that embeds each candidate action description in parallel, then scores these embeddings against the LLM's internal state to yield a distribution over typed actions. By factorizing decision-making into representations of the evolving interaction state and environment-specific action semantics, Dyad introduces an inductive bias for learning reusable representations while keeping action scoring efficient even as the action space grows. We investigate two complementary reinforcement learning settings driven by environment interaction. With the LLM frozen, training the action encoder alone achieves consistent gains across four unseen environments, enabling modular adaptation without modifying any LLM parameters. Jointly optimizing both components outperforms conventional RL post-training across diverse interactive tasks and model scales, including a 3.80% average absolute gain on ALFWorld with a 9B model, while improving general knowledge, reasoning, and coding.