Papers

The manifold hypothesis (MH) is often used to explain how machine learning can overcome the curse of dimensionality. However, the MH is only applicable in regimes where the training data provides a sufficiently dense sample of the underlying low-dimensional data manifold, or where such a low-dimensional manifold is conceivably present. We describe the regimes where the MH is not applicable as sparse. In this paper, we demonstrate that models succeed in the sparse regime by exploiting a highly structured local geometry, a property we formalize as normal alignment. We prove that normal-aligned classifiers – whose input-output Jacobians are rank-one and align perfectly with the training data – minimize the training objective under norm constraints and achieve maximal local robustness under a non-zero Jacobian constraint. For continuous piecewise-affine deep networks, normal alignment manifests geometrically as centroid alignment within the network's induced power diagram partition and results from the feature-learning regime. Motivated by these theoretical insights, we introduce GrokAlign, a regularization strategy that actively induces normal alignment. We demonstrate that GrokAlign significantly accelerates the training dynamics of deep networks relevant to the grokking phenomenon. Furthermore, we apply the principle of normal alignment to Recursive Feature Machines (RFMs) to introduce Recursive Feature Alignment Machines (RFAMs). We show that RFAMs exhibit greater adversarial robustness compared to RFMs when trained on tabular data.

Coherent, continuous spatial representations are critical for synthesizing physical and perceptual phenomena into a single representational space. Radial basis kernels provide a path forward for this type of distributed representation. In this work, we aim to characterize and analyze common radial basis kernels realizable in the neurally-plausible framework of spatial semantic pointers. Further, we analyze previous radial basis kernel work based on grid cell-like representations and demonstrate that such representations are both capable of and optimal for realizing radial basis kernels.

The realtime analysis and secure transmission of electrocardiogram ECG signals are critical for accurate diagnosis and safeguarding patient privacy in telemedicine applications This study presents a novel realtime ECG monitoring system that employs a learnable key generator LKG derived from each patients own ECG signal characteristics to dynamically produce unique encryption keys These keys determine the parameters r and x0 of a logistic map used for chaotic encryption The system securely encrypts realtime ECG data immediately after acquisition ensuring confidential transmission and storage in the cloud For remote clinical access the encrypted data is downloaded and decrypted on the doctors side using the matching key generated at the source or securely stored in the cloud This approach eliminates the need for traditional key exchange and substantially raises the cost of exhaustive key search in practice through persegment biometric key refresh and combined permutation and XOR diffusion supported by minentropy evaluation Compared to statickey methods the learnable biometric key design offers greater unpredictability and individualization A comprehensive set of security assessments including Shannon entropy 7678 bits correlation and autocorrelation disruption histogram statistics NIST SP 80022 frequency testing plaintextkey sensitivity avalanche effect FFTbased spectral flatness and robustness to noise and occlusion confirms the methods strength Reconstruction fidelity MSE approximately 5x106 PSNR greater than 52 dB MAE approximately 0002 demonstrates nearlossless decryption and preserved diagnostic features Encryption latency remains low preserving realtime performance.

Debugging CUDA programs has long been challenging because failures often arise from subtle interactions among hardware behavior, compiler decisions, memory hierarchy, and asynchronous execution. More importantly, with the rapid expansion of GPU usage across scientific computing, machine learning, graphics, and systems workloads, CUDA debugging has become more challenging than ever. Current evaluations of LLM-based CUDA programming largely miss this setting: a model can pass correctness tests with repair by degeneration, simplifying the CUDA code into a safer but slower program that abandons the original optimization structure. We introduce CUDABEAVER, a benchmark for CUDA debugging from real failing workspaces produced during LLM-based CUDA generation. Each task provides the broken candidate, native build/test commands, raw error evidence, and a single editable file. CUDABEAVER evaluates whether a fixer truly repairs the failing CUDA code or merely finds a slower test-passing replacement, reporting results by failure category, debugging trajectory, stagnation mode, and performance preservation. We further propose pass@k(M,C,A), a protocol-conditional CUDA debugging metric by making the fixer M, corpus C, and protocol axes Aexplicit. Using this metric across 213 tasks and seven frontier LLMs, we show that protocol-aware evaluation gives a more faithful view of CUDA debugging ability: when performance-loss tolerance is high, fixers appear much stronger, but even a minor stricter performance requirement can sharply reduce measured success, shifting scores by up to 40 percentage points.

Recovering continuous-time dynamics from discrete observations is difficult because local supervision loses fidelity as the observation interval grows. We replace local supervision with a global structural constraint: any flow representing autonomous dynamics must satisfy the semi-group property under time translation. We train a time-conditioned secant velocity field whose deviation from this property—which we call Symmetry Rupture—serves two roles: as a training regularizer it confines the hypothesis space to flows that compose consistently across temporal scales; as an inference oracle it guides an adaptive solver to select the largest step size that preserves internal consistency, without relying on local truncation error estimates. On the diffusion-reaction and Navier-Stokes benchmarks, our method achieves the lowest rollout RMSE among all evaluated methods using $\leq 3$ function evaluations per step. On the near-conservative shallow water benchmark, our method matches or exceeds competing flow-based methods at far lower computational cost; the strongest Neural ODE baseline achieves lower absolute RMSE but at the expense of $>100$ function evaluations per step. In the direct auto-regressive setting, our solver maintains stable long-horizon rollouts where flow-matching baselines diverge and Neural ODE requires up to $12\times$ more function evaluations. The core contribution is a model that internalizes temporal consistency, removing the need for high-order external solvers to compensate for structural bias.

This paper studies the role of sinks and diagonal patterns as attention switch and anti-oversmoothing mechanisms. We analyze geometric conditions under which sinks can be represented, showing a necessary alignment between the embedding of the sink and all other embeddings. Next, we refine the current understanding of the role of sinks in oversmoothing prevention: we specify the conditions under which dense attention is expected to smooth more than sparse attention, and empirically verify that such conditions are often satisfied in practice. We further prove an equivalence between sinks and hard attention switch, in which the output of the attention is identically 0. Finally, we relax the hard attention switch by allowing token self-communication: we provide a quantitative comparison of the costs of representing sinks vs.\ diagonal patterns, showing why sinks are favored in pretrained transformers. The introduction and analysis of diagonal patterns and the generalization of the attention switch close the gap between what oversmoothing prevention requires and what sinks provide, while also establishing when and why attention layers act like MLPs if token communication is not necessary.

Convolutional architectures have emerged as powerful alternatives to Transformers for sequence modeling. The primary advantage is that they offer improved theoretical sequence length complexity by leveraging the Fast Fourier Transform (FFT). However, this theoretical improvement does not always meaningfully land in practice. One critical obstacle is that applying standard FFTs is not amenable to the large-scale training pipeline wherein data is packed from different sources into a single sequence for hardware efficiency. Indeed, standard FFT algorithms are not easily amenable to document packing. Existing workarounds suffer from severe inefficiencies, crippling the practical performance of convolutional architectures. We close this gap with RubiConv, a novel algorithm for performing hardware-efficient, boundary-respecting convolutions on packed sequences. Extensive experiments show that RubiConv achieves significant speedups over both attention and standard FFT-based baselines. This work makes the theoretical efficiency of long convolutional models a practical reality for large-scale, real-world data packing.

Imitation learning is effective for training agents when expert demonstrations are available, but collecting demonstrations for every complex task in an environment is costly. We study the long-horizon, goal-conditioned setting where a fixed demonstration dataset contains useful behavior, but not complete examples for every task the agent must solve. Existing imitation learning methods can learn strong policies from demonstrations, but when solving long-horizon tasks, small errors accumulate over long primitive-action trajectories and make zero-shot adaptation to new tasks unreliable. We introduce Zero-shot Agents from Latent Topologies (ZALT), an imitation-learning method that solves unseen start-goal tasks beyond those demonstrated during training. ZALT identifies latent hub states where trajectories converge or diverge, learns policies and a dynamics model over hub-to-hub transitions, and plans over the hub topology to complete new tasks. This topology makes demonstrated behaviors explicitly composable while compressing long tasks into shorter sequences of abstract transitions – combined, these enable ZALT to perform zero-shot adaptation. In a complex 3D maze environment, ZALT achieves 55% zero-shot success on unseen tasks, compared to 6% for the strongest baseline.

Posterior-seeking training need not produce the best predictive distribution in its family under model misspecification. For a last-layer weight belief, we locate the restriction responsible for this gap and quantify its cost. At the last layer the negative ELBO decomposes exactly into the belief's predictive log scores, the prior KL, and a binding: a sum of per-observation Jensen gaps that constrains predictive dispersion through posterior covariance. Removing the binding gives the established KL-regularised form of predictive-score training; a parameterisation-dependent factor-normaliser form performs similarly at the last layer. At fixed predictive mean, the leading restricted direction is the variance at each input, and its local value is the classical variance-component score evaluated at the plug-in predictor. This yields both a diagnostic and a price: in the flat-leverage limit, the Gaussian cost is the Jensen gap of a representable residual-variance profile, whereas a likelihood without a dispersion parameter pays the full expressible improvement over the plug-in. Controlled experiments find typically tenths of a nat for Gaussian regression but units to tens of nats for overdispersed Poisson counts. In an exploratory, prospectively specified validation on 28 count sets, 27 unused elsewhere, a screen computed from the point-prediction head alone agreed with the measured outcome on 24 of 25 decided sets (an always-open rule: 18), and its price proxy, once fixed, was within a factor of three of the predictive-score gain of the fitted free profile on all 10 open sets tested afterwards. Score training changes the inferential target: it yields classical per-input overdispersed heads whose covariance represents residual dispersion, not posterior or epistemic uncertainty.

In LLM agents with persistent memory, whether a model stores an injected instruction and whether it later acts on it are separable safety properties. We call this the injection-execution dissociation. We evaluate 29 frontier endpoints with the malicious document in the retrieval corpus (N=10 each, four topped up to N=40; an earlier bug-affected collection is superseded). Execution is common: 7 endpoints exfiltrate in at least 9 of 10 runs and 11 more in some but fewer. Zero exfiltration is weak evidence of resistance: GPT-5.4, GPT-5.6-sol and Claude Haiku 4.5 never exfiltrate in 40 draft-only runs each but send almost no email, and in an exploratory follow-up that supplies the content and asks for a send they exfiltrate in 20 of 20, 5 of 20 and 6 of 20 runs. Anthropic's Opus and Sonnet models mostly store no forwarding directive and exfiltrate in at most 1 of 10 draft-only runs. We test six defences at four architectural layers in a 5,040-run factorial on nine open-weight models (N=40 per cell); these models store the instruction in every run, so it tests execution only. Input filters never see the payload, which arrives through retrieval. Retrieval-layer classifiers see it but, at the small operating points tested, do not flag it; frontier judges did. Only Memory Sandbox, which removes the recall tool, blocks the attack (0% ASR for eight of nine models). It is bypassed when a model rebuilds the instruction from the retrieval corpus: three Bedrock-served models do so (32.5-97.5% ASR), and a causal-mediation study on two OpenAI models attributes every success under the sandbox to this path. Stored but dormant instructions may create a cross-model risk; transfer was not demonstrated. Defences should be evaluated at both stages.

Reinforcement learning with verifiable rewards (RLVR) generates hundreds of thousands of tokens per training step, with rollout generation dominating the computational cost. The overall token budget can be controlled along two main dimensions: (i) deciding which prompts to allocate rollouts to, and (ii) deciding how long each rollout should be. Prior work has generally controlled only one of these dimensions at a time. We show that jointly tuning both decisions under a shared compute budget improves both reasoning quality and wall-clock training time. We instantiate this view as DUal-controlled tokEn allocaTion (DUET), a computationally efficient layer over GRPO that uses a lightweight pre-rollout surrogate of prompt informativeness to set how many rollouts each prompt receives, and a marker-gated abort rule with importance reweighting to set when to stop them. On Qwen3-1.7B trained on MATH, DUET outperforms full-budget GRPO and the other three budget-aware baseline methods. DUET's advantage further generalizes to other benchmarks across math and coding, and is on par with the best baseline on the scientific Q&A domain, while also achieving a $1.62\times$ wall-clock speedup. More notably, using only 50% of the token budget, DUET still outperforms all baseline methods at their full budget, achieving an even higher $2.51\times$ speedup over full-budget GRPO. We verify the high performance of DUET on other backbone LLMs, including Qwen3-4B to 14B and Llama-3.2-3B-Instruct. Notably, the gap between DUET and the strongest baseline widens as the budget tightens, contrary to the usual pattern in which efficient methods trade off quality as compute decreases. More broadly, these results suggest that DUET budget-aware control strategies are valuable not only for accelerating training, but also for improving the quality of the learning signal.

Adversarial purification with diffusion models seeks to project adversarial examples back toward the data manifold, but balancing semantic preservation and robustness against adaptive attacks remains challenging. Recent work shows that standard diffusion purification can fail under adaptive evaluation, while test-time score-based optimization is more resilient. Existing optimization defenses, however, typically rely on a single diffusion noise regime or treat timesteps uniformly, overlooking the distinct roles of coarse and fine denoising scales. We propose Temporal Adversarial Rectification Optimization (TARO), an inference-time purification method that builds a temporally guided score prior from multiple denoising views along the diffusion trajectory. TARO forms a coarse-to-fine residual target: high-noise experts provide globally smoothed structure with reduced adversarial sensitivity, while low-noise experts restore image-specific, class-relevant details. A guidance strength controls this temporal correction, allowing TARO to balance robust global rectification with semantic preservation. Empirically, TARO improves robust accuracy across datasets and adaptive threat models in a zero-shot setting, while remaining compatible with complementary adversarial-likelihood objectives for further robustness gains.

We introduce a meshfree exterior calculus (MEEC) for learning structure-preserving descriptions of physics on point clouds, and use it to build MEEC-Net, a data-efficient surrogate that transfers across resolutions, geometries, and physical parameters. MEEC equips an $\varepsilon$-ball graph with virtual node and edge measures via a single sparse Schur complement solve; the resulting complex satisfies discrete conservation exactly, is end-to-end differentiable in the point positions, and exposes a direct geometry-to-physics link without the mesh-generation step required by conventional structure-preserving discretizations. MEEC-Net learns unknown physics as a shared edge-wise flux law in an SO($d$)-invariant local frame, so the same kernel produces compatible fluxes on any point cloud whose features lie in the training range. We prove a solution-error bound that splits into discretization and kernel-approximation terms which is independent of problem geometry, explaining the observed transfer from very few examples. We show that single-solution training transfers to unseen geometries, boundary conditions, and physical parameters. On five canonical PDE benchmarks MEEC-Net achieves 1-2 orders of magnitude lower out-of-distribution error than baseline neural-operator approaches. On the SimJEB structural-bracket benchmark it achieves competitive error while using substantially fewer training geometries.

Post-deployment monitoring of healthcare AI requires statistically valid, label-efficient methods, but gold-standard labels from clinician chart review are expensive. Prediction-powered inference (PPI) and active statistical inference (ASI) reduce label cost by combining a small labeled sample with abundant model predictions, but both are restricted to a single predictor, a poor fit for modern clinical pipelines that have multiple predictors of differing cost and accuracy available at inference time. We propose Active Multiple-Prediction-Powered Inference (AM-PPI), which routes each instance to a cost-appropriate predictor subset, samples gold-standard labels in proportion to the chosen subset's residual uncertainty, and reweights predictions to minimize estimator variance, all under a single deployment-time budget. AM-PPI generalizes ASI to leverage multiple predictors and extends Multiple-PPI from global per-predictor allocation to per-instance adaptive routing. We derive closed-form Karush-Kuhn-Tucker (KKT) conditions for all three decisions and prove, via biconvexity and strong duality, that the resulting fixed point is a global optimum despite the joint problem being non-jointly-convex. We establish asymptotic normality with valid coverage, minimum-variance unbiasedness within the linear-prediction augmented inverse propensity weighted (AIPW) class, and a closed-form criterion identifying when multiple predictors help. On synthetic data and three healthcare monitoring tasks, AM-PPI produces 10 to 40 percent narrower confidence intervals (CIs) than single-predictor ASI in the budget regime where routing matters, and matches the better baseline elsewhere.

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.

We formulate flow matching on probability measures over probability measures. Directly from the Wasserstein-on-Wasserstein (WoW) continuity equation, we show that admissible nested outer and inner couplings induce a valid metameasure flow and non-local velocity field. Validity requires neither optimal plans nor a smooth manifold. Practically, this opens up the design space of multisample flow matching: exact Wasserstein plans are distinguished by energy and straightness, while sliced and lazy linear plans yield scalable alternatives. Point-cloud experiments, including controlled Bures-Wasserstein and Euclidean minibatch baselines, demonstrate the resulting geometry-cost trade-off.

We present a data-adaptive method for parameter-efficient fine-tuning of large neural networks. Standard low-rank adaptation methods improve efficiency by restricting each layer update to a fixed low-rank form, but this static parameterization can be too rigid when the appropriate correction depends on the input and on the evolving depth-wise computation of the network. Our approach replaces a purely layer-local adapter with a shared queryable memory of low-rank update atoms. For each block of layers, the model forms a query from the current low-rank state and a running summary of previous blocks, uses this query to retrieve a content-dependent combination of shared update components via attention, and applies the resulting routed operator within the low-rank bottleneck. In this way, the method retains the efficiency and scalability of low-rank adaptation while allowing the effective update to vary across inputs and to share reusable structure across layers. The resulting architecture provides a principled middle ground between static LoRA-style updates and fully generated parameter updates: it remains compact and parameter-efficient while supporting dynamic, context-sensitive adaptation. Further, we incorporate instruction-regularization by augmenting routing logits with a language-induced prior over update atoms, thereby biasing the selection of low-rank transformations toward semantically relevant directions without generating unconstrained parameter updates. Experiments on noisy non-linear regression tasks and LLM fine-tuning suggest that this queryable update-memory formulation can improve final test performance and training stability compared to standard low-rank adaptation, while using a comparable number of trainable parameters.

Designing model-free algorithms for distributionally robust reinforcement learning (DRRL) poses fundamental challenges. The robust Bellman operator is nonlinear in the transition kernel, which makes one-sample Bellman updates biased, while the adversarial optimization underlying robustness makes robust evaluation computationally demanding. To address these difficulties, we consider the natural small-ambiguity regime under Kullback–Leibler ambiguity sets and propose an approximate DRRL framework based on a first-order expansion of the relevant robust functional. This yields an approximate robust Bellman equation that removes the adversarial optimization while remaining first-order accurate in the ambiguity radius. To learn the fixed point of this approximate equation, we propose Mean-Variance Stochastic Approximation (MVSA), a model-free algorithm that uses only one-sample updates. This is achieved via a lifted stochastic approximation dynamics and a two-time-scale design. We then prove convergence and a central limit theorem for MVSA: its main iterate satisfies a central limit theorem at the canonical $n^{-1/2}$ scale, with explicitly characterized asymptotic covariances. Finally, we validate our theoretical findings with a numerical experiment.

Jurisprudence, the study of how judges should properly decide cases, and alignment, the science of getting AI models to conform to human values, share a fundamental structure. These seemingly distant fields both seek to predict and shape how decisions by powerful actors, in one case judges and in the other increasingly powerful artificial intelligences, will be made in the unknown future. And they use similar tools of the specification and interpretation of language to try to accomplish those goals. The great debates of jurisprudence, about what the law is and what it should be, can provide insight into alignment, and lessons from what does and does not work in alignment can help make progress in jurisprudence. This essay puts the two fields directly into conversation. Drawing on leading accounts of jurisprudence, particularly Dworkin's principle-oriented interpretivism and Sunstein's positivist account of law as analogical reasoning, and on cutting-edge alignment approaches, namely Constitutional AI and case-based reasoning, it illustrates the value of a more sophisticated legally-inspired approach to the interplay of rules and cases in finetuning alignment and points to ways that AI can provide a better understanding of how the law works and how it can be improved by the introduction of AI. AI systems and the law should operate to empower people to act in the world, helping to expand their capabilities and the extent to which they are able to achieve their goals. As AI continues to improve in capacity, and as the constraints that legal theory places on human judges seem be coming undone, the conversation between these two fields will become increasingly essential and may help point to a better version of both.

Physics-informed neural networks (PINNs) provide a flexible framework for solving forward and inverse problems governed by partial differential equations (PDEs), but standard PINN training typically relies on soft penalty formulations that combine PDE residuals, data mismatch, and initial/boundary conditions using manually chosen weights. This often leads to ill-conditioning, sensitivity to loss weights, and poor constraint satisfaction. In this work, we reformulate PINN training as an equality-constrained optimization problem and propose a novel Adaptive Momentum Feedback Linearization Optimization for Hard Constrained PINN (AdamFLIP). The key idea is to view the constraint residuals as the output of a controlled dynamical system and to compute the Lagrange multiplier as a feedback input that locally drives these residuals toward stable linear contraction dynamics. AdamFLIP then applies Adam-style first- and second-moment adaptation to the resulting feedback-linearized Lagrangian gradient, combining principled constraint handling with the scalability and robustness of adaptive neural-network optimization. We test AdamFLIP on a range of benchmark forward and inverse PDE problem, and it consistently outperforms both the standard soft-constrained PINN and state-of-the-art constrained optimizers. Specifically, on the Navier–Stokes equations benchmark, AdamFLIP \textbf{reduces relative $L_2$ error by more than two thirds} for the predicted solution compared to the next best method. Our AdamFLIP framework provides an effective and computationally scalable hard constraint optimization method for PINN training.

How do LLMs learn in-context? Is it by pattern-matching recent tokens, or by inferring latent structure? We probe this question using a toy graph random-walk across two competing graph structures. This task's answer is, in principle, decidable: either the model tracks global topology, or it copies local transitions. We present two lines of evidence that neither account alone is sufficient. First, reconstructing the internal representation structure via PCA reveals that at intermediate mixture ratios, both graph topologies are encoded in orthogonal principal subspaces simultaneously. This pattern is difficult to reconcile with purely local transition copying. Second, residual-stream activation patching and graph-difference steering causally intervene on this graph-family signal: late-layer patching almost fully transfers the clean graph preference, while linear steering moves predictions in the intended direction and fails under norm-matched and label-shuffled controls. Taken together, our findings are most consistent with a dual-mechanism account in which genuine structure inference and induction circuits operate in parallel.

Dynamics of interacting systems in engineering, society, and nature often evolve over latent networks that govern which entities can interact. We study the problem of inferring these networks from event-based observations, which arise naturally in finance, seismology, and neuroscience. While there is substantial algorithmic work addressing this important problem, theoretical results are scarce. In this paper we ask the following fundamental question: what is the minimum time that one must observe the dynamics in order to exactly recover the underlying network, as a function of the number $d$ of interacting entities? For a class of stationary Hawkes processes with sparse, weak interactions, we prove that an observation time of order $\log d$ is sufficient and necessary. For the upper bound we construct a two-stage estimator that uses clipped and binned event data for screening, followed by a least-squares refinement, and apply concentration bounds derived from the Poisson cluster representation. For the lower bound we combine Fano's inequality with Jacod's Girsanov formula for point processes on a suitable subclass of networks.

In this work, we show that Latent Flow-Matching (LFM) models are robust to different types of perturbations, including data reduction and model capacity shrinkage. We characterize this stability by these models' tendency to generate similar outputs under identical noise seeds. We provide a perspective relating this phenomenon to flow matching theory, which indicates that this stability is inherent to the FM objective. We further exploit this stability to derive practical algorithms for more efficient training and inference. Concretely, first, we show that by training LFM models on significantly reduced datasets, performance is preserved, and in compute-constrained regimes, the model converges faster while maintaining quality. This yields multiple advantages, including savings in the training time due to faster convergence, and alleviating annotation effort when training conditional models. Second, LFM stability under architectural shrinkage gives rise to a two-model coarse-to-fine approach, one using a light-weight architecture for the first phase of the FM trajectory, and one with higher capacity for the second, thereby reducing the inference cost substantially. To determine which samples are informative, we introduce three sample-scoring criteria and evaluate them under standard metrics for generative models. Our results are thoroughly evaluated on multiple datasets, demonstrating the practical advantage of this stability, including data savings and a more than two-fold inference speedup while generating comparable outputs.

Practical diffusion sampling requires simulating a reverse-time ODE or SDE with a limited number of denoising steps, making the choice of sampling parameters crucial for minimizing discretization error. Non-asymptotic convergence bounds characterize sampling complexity, but their worst-case constants can obscure target geometry and thereby limit guidance on parameter optimization. Rather than bounding the error, we derive asymptotically exact small-stepsize expansions of Euler-Maruyama weak and Frechet errors for general smooth reverse diffusions, with explicit formulas for Gaussian data. These formulas provide tractable objectives for optimizing diffusion parameters, including the noise and rescaling schedules and the stochasticity coefficient, according to the target's covariance spectrum. In particular, our theory predicts lower optimal stochasticity at smaller step budgets, shows how to adapt the rescaling coefficient to the data power spectrum, and motivates a new family of effective noise schedules. A perturbative extension to Gaussian scale mixtures (GSMs) quantifies how departures from Gaussianity shift the optimal parameters. Finally, experiments on different real image datasets show that FID-optimal parameters agree with the qualitative theoretical predictions.

Cooperative multi-agent reinforcement learning agents that act on partial local observations face a fundamental information bottleneck: the knowledge needed to select jointly optimal actions is scattered across the team, yet each agent must commit to a decision without access to its teammates' observations, intentions, or chosen actions. Existing methods either ignore this bottleneck, compress it into a scalar mixing signal, or route around it with learned communication channels. Framing action coordination as a problem of structured information integration among agents, we propose structured agent coordination via holistic information integration, or SACHI, in which graph transformer convolutions over an inter-agent coordination graph enrich each agent's representation with receiver-sensitive, content-dependent signals from teammates prior to action selection. We evaluate SACHI across five cooperative tasks spanning spatial, communicative, and adversarial coordination challenges against twelve baselines. SACHI consistently matches or outperforms the best baseline on every task, and rigorous aggregate statistical analyses, including normalized metrics with bootstrap confidence intervals, Friedman ranking, and performance profiling, confirm that this advantage is statistically significant, robust across environments, and not attributable to increased model capacity. Parameter-matched ablations further trace the source of the gains to a single architectural property: the degree of content-dependence in the message-passing operator.

Sequence prediction methods for linear dynamical systems with long memory, i.e. marginally stable systems, typically achieve regret that grows linearly with the hidden dimension of the underlying generative model. While many methods have been developed to address this regime with varying success, we show that simply using the second-order Vovk-Azoury-Warmuth (VAW) algorithm to learn a short autoregressive-with-inputs (ARX) model achieves astoundingly strong results: for bounded sequential data from a marginally-stable linear dynamical system with spectra in the complex disk except for angular wedge of width $\delta$ around the negative real axis, this algorithm achieves dimension-free regret $O\left( \delta^{-4} \log^2 T \right)$. These bounds are state-of-the-art to our knowledge. The key components for our result come from 1) using the theory of “Universal Sequence Preconditioning” (USP) \cite{marsdenuniversal} to prove the existence of an optimal setting of autoregressive coefficients, 2) the application of VAW which takes better advantage of the memory compression provided by USP, and 3) the analysis of Faber polynomials on circular sectors to extend these results to systems with complex spectra.

This paper proposes a time-warping transfer learning method, a technique for temporally rescaling the learned dynamics of a recurrent neural network (RNN) with a Long Short-Term Memory (LSTM) layer to enable task transfer across fuel moisture classes. Fuel moisture content (FMC) is divided into idealized classes based on characteristic lag time. Large quantities of real-time data are available for 10h fuels from sensors on weather stations, but observations of other fuel classes are sparse in space and time. We use transfer learning to adapt an RNN pretrained on 10h FMC to predict FMC for 1h, 100h, and 1000h fuels. We validate this method using data from a landmark field study conducted in Oklahoma that was used to calibrate the state-of-the-art Nelson fuel moisture model.

Reinforcement learning has become a powerful paradigm for improving the capability of intelligent systems, but its practical deployment faces two central challenges. First, reinforcement learning must scale efficiently in distributed environments where communication bandwidth is limited and computation is heterogeneous across agents. Second, as reinforcement learning is increasingly used in post-training large language models and autonomous agents, the optimized policies must also be aligned with human preferences and satisfy safety requirements such as privacy-aware information disclosure. This dissertation addresses both challenges through four complementary contributions spanning federated optimization, preference alignment, and contextual safety. The first part of the dissertation studies scalable reinforcement learning in federated settings. The second part of the dissertation studies trustworthy reinforcement learning for large language models. Together, these contributions advance reinforcement learning along two complementary dimensions. On the one hand, they make reinforcement learning more scalable through communication-efficient and asynchronous federated optimization. On the other hand, they make reinforcement learning more trustworthy by improving alignment with human preferences and by reducing contextually inappropriate information disclosure in language-based intelligent systems. As a whole, this dissertation argues that the next generation of intelligent systems will require both efficient optimization and trustworthy behavior, and that reinforcement learning provides a unifying framework for addressing both goals.

In many practical applications it is important to build symmetries into neural network architectures. Consider the important case of permutation symmetry on point clouds consisting of $n$ points in $d$ dimensions. In this case the network learns a function on a set of $n$ points in $\mathbb{R}^d$, and a natural paradigm for constructing invariant networks is Janossy pooling, which generalizes the popular Deep Sets architecture. We study the universality of this approach, in particular the important question of how large the embedding dimension must be to guarantee universality of this architecture. Specifically, using a novel technique, we prove new lower bounds on the required size of this embedding dimension. For Deep Sets, this gives the correct minimal dimension up to a constant factor for all $d > 1$. For $k$-ary Janossy pooling, we prove the first non-trivial lower bound on the required embedding dimension when $k > 1$.

Debates about large language model post-training often treat supervised fine-tuning (SFT) as imitation and reinforcement learning (RL) as discovery. But this distinction is too coarse. What matters is whether a training procedure increases the probability of behaviors the pretrained model could already produce, or whether it changes what the model can practically reach. We argue that post-training research should distinguish between capability elicitation and capability creation. We make this distinction operational by introducing the notion of accessible support: the set of behaviors that a model can practically produce under finite budgets. Post-training that reweights behaviors within this support is capability elicitation; whereas changing the support itself corresponds to capability creation. We develop this argument through a free-energy view of post-training. SFT and RL can both be seen as reweighting a pretrained reference distribution, only with different external signals. Demonstration signals define low-energy behavior for SFT, and reward signals define low-energy behavior for RL. When the update remains close to the base model, the main effect is local reweighting, not capability creation. Within this framework, the central question is no longer whether post-training is framed as SFT or RL, but whether it reweights behaviors already within reach, or instead expands the model's reachable behavioral space through search, interaction, tool use, or the incorporation of new information.