Papers

Bayesian optimization is widely used for hyperparameter optimization when model evaluations are expensive; however, noisy acquisition estimates can lead to unstable decisions. We identify acquisition estimation noise as a failure mode that was previously overlooked: even when the surrogate model and acquisition target are correctly specified, finite-sample Monte Carlo error can perturb acquisition values. This can, in turn, flip candidate rankings and lead to suboptimal BO decisions. As a remedy, we aim at variance reduction and propose an orthogonal acquisition estimator that subtracts an optimally weighted score-function control variate, which yields an acquisition residual orthogonal to posterior score directions and which thus reduces Monte Carlo variance. We further introduce OrthoBO: a Bayesian optimization framework that combines our orthogonal acquisition estimator with ensemble surrogates and an outer log transformation. We show theoretically that our estimator preserves the target, leads to variance reduction, and improves pairwise ranking stability. We further verify the theoretical properties of OrthoBO through numerical experiments where our framework reduces acquisition estimation variance, stabilizes candidate rankings, and achieves strong performance. We also demonstrate the downstream utility of OrthoBO in hyperparameter optimization for neural network training and fine-tuning.

Channel state information (CSI)-based human activity recognition (HAR) is vulnerable to performance degradation under domain shifts across varying physical environments. Continual learning (CL) offers a principled way to learn new domains sequentially while preserving past knowledge, but existing CL solutions for CSI-based HAR scale poorly with accumulating domains, rely on a large replay buffer, or incur linearly growing inference cost. In this letter, we propose Scene-Adaptive Mixture of Experts with Clustered Specialists (SAMoE-C), which formulates cross-domain CSI-based HAR as a mixture-of-experts system that enables scene-specific adaptation, via an attention-based semantic router that activates only selected experts for each input. Moreover, we develop a novel training protocol, which requires only a tiny replay buffer for stabilizing domain discrimination of the router. Experimental results on a four-scene CSI dataset demonstrate that SAMoE-C approaches the state-of-the-art accuracy, while maintaining a significantly lower inference cost. By jointly combining modular experts, selective activation with router and a lightweight training protocol, SAMoE-C enables scalable cross-domain CSI-based HAR deployment with low training overhead and high computational efficiency in real-world settings.

Federated learning with heterogeneous clients remains a significant challenge for deep learning, primarily due to client drift arising from inconsistent local updates. Existing federated optimization methods typically address this issue through objective-level regularization or update-correction mechanisms. Recent studies, however, suggest that Transformer-based architectures may be inherently more robust than conventional models under heterogeneous federated training. Motivated by this observation, we investigate how different parameter components within the attention mechanism influence federated optimization. Specifically, we decompose the attention module into a query/key block, which determines the attention kernel, and a value block, which performs semantic transformation under the induced kernel. Based on this perspective, we propose FedFrozen, a two-stage federated optimization framework that first performs full-model warm-up training and then freezes the query/key block while continuing to optimize the value block. Under a linear-attention formulation, we show that the warm-up stage can be interpreted as an inexact descent procedure on a regularized kernel-profile objective, while the frozen stage reduces to a restricted value-block optimization problem under a fixed attention kernel. Our analysis further reveals an explicit trade-off that governs the choice of warm-up length. Simulations validate the predicted bias-drift behavior, and real-data experiments demonstrate that FedFrozen improves both the stability and effectiveness of Transformer models in heterogeneous federated learning.

Concept Bottleneck Models (CBMs) have become a popular approach to enable interpretability in neural networks by constraining classifier inputs to a set of human-understandable concepts. While effective, current models embed concepts in flat Euclidean space, treating them as independent, orthogonal dimensions. Concepts, however, are highly structured and organized in semantic hierarchies. To resolve this mismatch, we propose Hyperbolic Concept Bottleneck Models (HypCBM), a post-hoc framework that grounds the bottleneck in this structure by reformulating concept activation as asymmetric geometric containment in hyperbolic space. Rather than treating entailment cones as a pre-training penalty, we show they encode a natural test-time activation signal: the margin of inclusion within a concept's entailment cone yields sparse, hierarchy-aware activations without any additional supervision or learned modules. We further introduce an adaptive scaling law for hierarchically faithful interventions, propagating user corrections coherently through the concept tree. Empirically, HypCBM rivals post-hoc Euclidean models trained on 20$\times$ more data in sparse regimes required for human interpretability, with stronger hierarchical consistency and improved robustness to input corruptions.

Traditional multi-population models, such as the Li-Lee framework, rely on the assumption of mean-reverting country-specific deviations. However, recent data from high-longevity clusters suggest a systemic break in this paradigm. We identify a stationarity paradox where mortality residuals in countries like Sweden and West Germany exhibit persistent unit roots, leading to a systematic mispricing of longevity risk in linear models. To address these non-linearities, we propose Hybrid-Lift, a neural-actuarial framework that combines Hierarchical LSTM networks with a Mean-Bias Correction (MBC) anchoring mechanism. Positioned as a governance-friendly model challenger rather than a replacement of classical approaches, the framework exhibits selective superiority on out-of-sample validation (2012-2020): it outperforms Li-Lee by 17.40% in Sweden and 12.57% in West Germany, while remaining comparable for near-linear regimes such as Switzerland and Japan. We complement the predictive model with an integrated governance suite comprising SHAP-based cross-country influence mapping, a dual uncertainty framework for regulatory capital calibration (Swiss ES 99.0% of +1.153 years), and a reverse stress test identifying the critical shock threshold for solvency buffer exhaustion. This research provides evidence that neural networks, when properly anchored by actuarial principles, can serve as effective model challengers for longevity risk management under the SST and Solvency II standards.

The use of content features, particularly textual and linguistic for fake news detection is under-researched, despite empirical evidence showing the features could contribute to differentiating real and fake news. To this end, this study investigates a selection of content features such as word bigrams, part of speech distribution etc. to improve fake news detection. We performed a series of experiments on a new dataset gathered during the COVID-19 pandemic and using Decision Tree, K-Nearest Neighbor, Logistic Regression, Support Vector Machine and Random Forest. Random Forest yielded the best results, followed closely by Support Vector Machine, across all setups. In general, both the textual and linguistic features were found to improve fake news detection when used separately, however, combining them into a single model did not improve the detection significantly. Differences were also noted between the use of bigrams and part of speech tags. The study shows that textual and linguistic features can be used successfully in detecting fake news using the traditional machine learning approach as opposed to deep learning.

Subgraph pattern detection is a fundamental graph mining task central to anti-money laundering (AML) in finance. However, state-of-the-art graph neural network (GNN)-based methods assume centralized graph access, impractical when transaction data cannot be shared across institutions. In practice, graphs are distributed across clients, where local GNN training diverges from centralized learning, creating representation and adjoint equivalence gaps that induce a parameter-gradient gap. We formalize this as a structural observability problem, where AML patterns crossing subgraph boundaries become locally unidentifiable. To bridge these gaps, we propose a layer-wise exchange framework: clients synchronize node representations each forward pass layer and adjoints each backward pass, without exposing raw features or labels. Under a shared-parameter assumption, we prove node representations and parameter gradients are equivalent to those of centralized GNN training. Experiments on directed multigraphs with money-laundering patterns show layer-wise exchange performs best with per-step synchronization and remains effective under practical per-epoch synchronization. Our results show layer-wise exchange and parameter synchronization to be complementary, supporting distributed pattern detection in privacy-constrained financial settings.

Pixel-space diffusion has re-emerged as a promising alternative to latent-space generation because it avoids the representation bottleneck introduced by VAEs. Yet most existing methods still treat image generation as a frequency-homogeneous process, overlooking the distinct roles and learning dynamics of low- and high-frequency components. To address this, we propose FREPix, a FREquency-heterogeneous flow matching framework for Pixel-space image generation. FREPix explicitly decomposes generation into low- and high-frequency components, assigns them separate transport paths, predicts them with a factorized network, and trains them with a frequency-aware objective. In this way, coarse-to-fine generation becomes an explicit design principle rather than an implicit behavior. On ImageNet class-to-image generation, FREPix achieves competitive results among pixel-space generation models, reaching 1.91 FID at $256\times256$ and 2.38 FID at $512\times512$, with particularly strong performance in the early stages of training and in the low-NFE regime.

We introduce E = T*H/(O+B), a dimensionless control parameter that predicts whether Mixture-of-Experts (MoE) models will develop a healthy expert ecology or collapse into dead experts. E combines four hyperparameters – routing temperature T, routing entropy weight H, oracle weight O, and balance weight B – into a single quantity. Through 12 controlled experiments (8 vision, 4 language) totaling over 11,000 training epochs, we establish that E >= 0.5 alone is sufficient to guarantee zero dead experts, removing the necessity for handcrafted load-balancing auxiliary losses. We validate this cross-modally on CIFAR-10, CIFAR-100, TinyImageNet-200, WikiText-2, and WikiText-103. Six additional findings emerge: (1) dead experts can resuscitate – triggered by balance loss driving router re-exploration; (2) ortho toxicity is dataset-dependent, not universal; (3) task complexity shifts the critical E threshold; (4) model overfitting is decoupled from expert ecological health; (5) three-tier MoE spontaneously collapses into a two-tier functional structure; (6) ecological structure is temperature-invariant across a 50x range. We propose that E serves as a unified diagnostic for MoE training, analogous to the Reynolds number in fluid dynamics.

Prior-Fitted Networks (PFNs) amortize Bayesian prediction by meta-learning over a synthetic task prior, but their standard output is a posterior predictive distribution over noisy observations. For sequential decision-making, such as active learning and Bayesian optimization, acquisition should prioritize epistemic uncertainty about the latent signal rather than irreducible aleatoric observation noise. We show that this epistemic–aleatoric split is not identifiable in general from the posterior predictive distribution alone, even when that distribution is known exactly. We then exploit a distinctive advantage of PFNs: because the synthetic data-generating process is under our control, each task can contain an explicit latent signal and noise function, and the generator can provide query-level labels for both the noiseless target and the observation-noise variance. We use these labels to train a decoupled PFN with separate latent-signal and aleatoric heads. The observation-level predictive is induced by convolving the latent signal distribution with the learned noise model. Empirically, epistemic-only acquisition mitigates the failure mode of total-variance exploration in noisy and heteroscedastic settings. In matched comparisons, decoupled models usually improve over tuned observation-level baselines, with the clearest gains in HPO; in broader sweeps, a decoupled model obtains the best average rank in both HPO and synthetic BO.

Gradient-based attribution methods are model-faithful and scalable, but Integrated Gradients (IG) can be brittle because explanations depend on heuristic baselines, straight-line paths, discretization, and saturation. We propose Fisher–Rao Integrated Gradients (FRInGe), which defines both the reference and interpolation schedule in predictive distribution space. FRInGe replaces input baselines with a maximum-entropy predictive reference and follows a Fisher-Rao geodesic on the probability simplex. The corresponding input-space trajectory is realized through the pullback Fisher metric and stabilized by KL and Euclidean trust regions; attributions are obtained by integrating input gradients along this trajectory. Across six ImageNet architectures, FRInGe most clearly improves calibration-oriented attribution metrics, especially MAS scores, while remaining competitive on perturbation AUC and infidelity.

Semi-structured sparsity provides a practical path to accelerate large language models (LLMs) with native hardware support, but post-training semi-structured pruning often suffers from substantial quality degradation due to strong structural coupling. Existing methods rely on large-scale sparse retraining to recover accuracy, resulting in high computational cost. We propose SparseForge, a post-training framework that improves recovery efficiency by directly optimizing the sparsity mask rather than scaling up retraining tokens. SparseForge combines Hessian-aware importance estimation with progressive annealing of soft masks into hardware-executable structured sparsity, enabling stable and efficient sparse recovery. On LLaMA-2-7B under 2:4 sparsity, SparseForge achieves 57.27% average zero-shot accuracy with only $\textbf{5B}$ retraining tokens, surpassing the dense model's 56.43% accuracy and approaching the 57.52% result of a state-of-the-art method using $\textbf{40B}$ tokens. Such improvements on the accuracy-efficiency trade-off from SparseForge are shown to be consistent across model families.

Modern deep learning architectures increasingly contend with sophisticated signals that are natively infinite-dimensional, such as time series, probability distributions, or operators, and are defined over irregular domains. Yet, a unified learning theory for these settings has been lacking. To start addressing this gap, we introduce a novel convolutional learning framework for possibly infinite-dimensional signals supported on a manifold. Namely, we use the connection Laplacian associated with a Hilbert bundle as a convolutional operator, and we derive filters and neural networks, dubbed as HilbNets. We make HilbNets and, more generally, the convolution operation, implementable via a two-stage sampling procedure. First, we show that sampling the manifold induces a Hilbert Cellular Sheaf, a generalized graph structure with Hilbert feature spaces and edge-wise coupling rules, and we prove that its sheaf Laplacian converges in probability to the underlying connection Laplacian as the sampling density increases. Notably, this result is a generalization to the infinite-dimensional bundle setting of the Belkin & Niyogi \cite{BELKIN20081289} convergence result for the graph Laplacian to the manifold Laplacian, a theoretical cornerstone of geometric learning methods. Second, we discretize the signals and prove that the discretized (implementable) HilbNets converge to the underlying continuous architectures and are transferable across different samplings of the same bundle, providing consistency for learning. Finally, we validate our framework on synthetic and real-world tasks. Overall, our results broaden the scope of geometric learning as a whole by lifting classical Laplacian-based frameworks to settings where the signal at each point lives in its own Hilbert space.

These lecture notes introduce some topics of classical statistical physics, particularly those that are relevant for neural networks and deep learning. Statistical physics is treated as a branch of probability theory or statistics, with the goal of making concepts such as phase transitions and the renormalization group accessible to readers without prior knowledge of physics. We introduce the Boltzmann-Gibbs distribution and the thermodynamic potentials on a finite configuration space, notably for Ising spins and spin-glass models on a lattice, and then define phase transitions as discontinuities that arise in the limit that the number of lattice points goes to infinity. We further introduce Hopfield networks and Boltzmann machines, which are governed by the same energy function as spin-glass models, and discuss the learning algorithm for restricted Boltzmann machines. In this algorithm hidden neurons are integrated out as in the renormalization group. Finally, modern deep learning is introduced, whose early developments were in part motivated by restricted Boltzmann machines in that they carry many layers of hidden neurons. A description of large language models is given.

World model-based policy evaluation is a practical proxy for testing real-world robot control by rolling out candidate actions in action-conditioned video diffusion models. As these models increasingly adopt latent diffusion modeling (LDM), choosing the right latent space becomes critical. While the status quo uses autoencoding latent spaces like VAEs that are primarily trained for pixel reconstruction, recent work suggests benefits from pretrained encoders with representation-aligned semantic latent spaces. We systematically evaluate these latent spaces for action-conditioned LDM by comparing six reconstruction and semantic encoders to train world model variants under a fixed protocol on BridgeV2 dataset, and show effective world model training in high-dimensional representation spaces with and without dimension compression. We then propose three axes to assess robotic world model performance: visual fidelity, planning and downstream policy performance, and latent representation quality. Our results show visual fidelity alone is insufficient for world model selection. While reconstruction encoders like VAE and Cosmos achieve strong pixel-level scores, semantic encoders such as V-JEPA 2.1 (strongest overall on policy), Web-DINO, and SigLIP 2 generally excel across the other two axes at all model scales. Our study advocates semantic latent space as stronger foundation for policy-relevant robotics diffusion world models.

On-policy distillation (OPD) trains a student on its own trajectories with token-level teacher feedback and often outperforms off-policy distillation and standard reinforcement learning. However, we find that its standard advantage weighted policy gradient suffers from three structural weaknesses, including high variance updates, vanishing gradients in zero-advantage regions, and exploration bottlenecks when corrective signals are insufficient. We therefore propose Asymmetric On-Policy Distillation (AOPD), which replaces ineffective negative reinforcement with localized divergence minimization in non-positive advantage regions while preserving positive reinforcement learning. Experiments on mathematical reasoning benchmarks show that AOPD consistently outperforms standard OPD, with average gains of 4.09 / 8.34 under strong / weak initialization, respectively. AOPD also maintains higher policy entropy during training and better capability retention during sequential tool-use adaptation.

This position paper argues that, in debiased machine learning, balancing functions should be derived from the Neyman orthogonal score, not chosen only as functions of covariates. Covariate balancing is effective when the regression error entering the score can be represented by functions of covariates alone, and it is the natural finite-dimensional approximation for targets such as ATT counterfactual means. For ATE estimation under treatment effect heterogeneity, however, the score error generally contains treatment-specific components because the outcome regression is a function of the full regressor $X=(D,Z)$. In that case, balancing common functions of $Z$ can leave the treatment-specific component unbalanced. We therefore advocate regressor balancing, implemented by Riesz regression with basis functions of $X$, as the general balancing principle for DML. The position is not that covariate balancing is invalid, but that covariate balancing should be understood as the special case that is appropriate when the score-relevant regression error is a function of covariates alone.

Estimating causal effects from observational data requires identifying valid adjustment sets. This task is especially challenging in realistic settings where latent confounding and feedback loops are present. Existing approaches typically assume acyclicity or rely on global causal structure learning, limiting applicability and computational efficiency. In this work, we study a local, data-driven method for covariate selection based on conditional independence information under a pre-treatment assumption. While this method is known to be sound and complete in acyclic causal models, its validity in the presence of cycles has remained unclear. Our main contribution is to show that these guarantees extend to cyclic causal models. In particular, our result relies on the invariance of conditional independence assertions under $\sigma$-acyclification. These findings establish a unified, cycle-agnostic perspective on covariate selection, which allows for causal effect estimation using any consistent estimator. Therefore, showing that the method applies across cyclic and acyclic settings without modification. Empirically, we validate this on extensive synthetic data, showing reliable performance in cyclic causal models.

We introduce MinMax Recurrent Neural Cascades (MinMax RNCs), a class of recurrent neural networks built from a novel form of recurrence over the MinMax algebra. We show that MinMax RNCs enjoy key properties that are difficult to obtain simultaneously: strong formal expressivity, efficient evaluation, stable dynamics, and non-vanishing state gradients. First, their formal expressivity corresponds to the regular languages, arguably the maximal expressivity for finite-memory systems. Second, in addition to evaluation in recurrent form, they also admit parallel-scan evaluation with logarithmic depth and linear work in the input length. Third, their states and activations are uniformly bounded for all sequence lengths. Fourth, their loss gradients exist almost everywhere and are uniformly bounded for all sequence lengths. Fifth, they do not exhibit vanishing state gradients: the gradient of a state with respect to a past state can retain norm one independently of the temporal distance between the states. Empirically, we find that these theoretical properties translate into strong practical performance. MinMax RNCs solve the considered synthetic tasks perfectly, generalise to long sequences, and outperform the recurrent baselines considered in our experiments. We also train a 112M-parameter MinMax RNC for next-token prediction, obtaining competitive performance for its size and providing initial evidence that MinMax recurrence can scale to real-world sequence-modelling tasks.

The topology of a classifier's decision regions determines how inputs with the same predicted label can be connected and deformed without changing that prediction. Prior empirical work constructed paths between same-label images within a single region, but did not examine whether loops bound surfaces within that region. We investigate this question using adaptive quadrilateral meshes with targeted repair of off-label interior vertices, while holding the same-label boundary loop fixed. A finite-resolution acceptance criterion distinguishes completed constructions from those left unresolved at the refinement ceiling. Across the pretrained classifiers studied, every tested loop admits an accepted filling. Construction effort varies by orders of magnitude within classes and is greater for mean-score-adjusted randomly initialised classifiers than for trained classifiers. An analytic control with a known hole leaves winding loops unresolved at the tested hole radii at or above the resolution threshold. These results provide empirical evidence consistent with simply connected decision regions at the tested resolution.

We study Nash equilibrium learning in partially observable Markov games (POMGs), a multi-agent reinforcement learning framework in which agents cannot fully observe the underlying state. Prior work in this setting relies on centralization or information sharing, and suffers from sample and computational complexity that scales exponentially in the number of players. We focus on a subclass of POMGs with independent state transitions, where agents remain coupled through their rewards, and assume that the underlying fully observed Markov game is a Markov potential game. For this class, we present an independent learning algorithm in which players, observing only their own actions and observations and without communication, jointly converge to an approximate Nash equilibrium. Due to partial observability, optimal policies may in general depend on the full action-observation history. Under a filter stability assumption, we show that policies based on finite history windows provide sufficient approximation guarantees. This enables us to approximate the POMG by a surrogate Markov game that is near-potential, leading to quasi-polynomial sample and computational complexity for independent Nash equilibrium learning in the underlying POMG.

Policy-gradient methods need a scalar reward, yet for many tasks no formula captures what we want, while judging which of two outcomes is better is easy. We study reinforcement learning that receives only such comparisons - no reward function, no reward model - in multi-step tasks. Groups of trajectories start from a shared state, and each action is credited by how its reward-to-go compares with those of other trajectories at the same time step. We show the expected update is policy gradient with reward-to-go replaced by its win rate against the population still running at that time, weighted by the chance an opponent is still running; in deterministic environments the update ranks actions by true value and the optimal policy is stable, whereas under stochastic outcomes it prefers actions that win more often rather than those with higher mean. Label noise near ties or from position bias leaves the expected update unchanged; only noise that also flips clear comparisons halves the signal. A critic trained from the same comparisons removes the state confound without changing the expected update. In a pre-registered study over six environments with ten seeds per claim, comparisons cost no detectable performance relative to scalar returns at the same granularity, learning a reward model from the same comparisons is not better, and ten percent of comparisons suffice when spent on a few time points with all pairs. With real LLM judges, comparisons are 7-10 points more accurate than pointwise scores, and the gap grows with the stronger judge. Trained end-to-end with an LLM judge on a requirement the hand-written reward omits, policies satisfy it in all five seeds but land far less often than with oracle comparisons, because the judge is near chance on the near-tied failures dominating early training. Comparisons do not close the gap to PPO with scalar rewards on HalfCheetah.

Finite-sample analyses of deep Q-learning typically treat replayed data as independent, even though it is sampled from temporally dependent state-action trajectories. We study the Deep Q-networks (DQN) algorithm under explicit dependence by modelling the minibatches used for updating the network as $\tau$-mixing. We show that this assumption holds under certain dependence conditions on the underlying trajectories and the mechanism used to sample minibatches. Building on this observation, we extend statistical analyses of DQN with fully connected ReLU architectures to dependent data. We formulate each update as a nonparametric regression problem with $\tau$-mixing observations and derive finite-sample risk bounds under this dependence structure. Our results show that temporal dependence leads to a degradation in the statistical rate by inducing an additional dimensionality penalty in the rate exponent, reflecting the reduced effective sample size of $\tau$-mixing data. Moreover, we derive the sample complexity of DQN under $tau$-mixing from these risk bounds. Finally, we empirically demonstrate on standard Gymnasium environments that the independence assumption is systematically violated and that replay sampling yields approximately exponentially decaying correlations, supporting our theoretical framework.

Despite extensive research into mitigating distribution shifts, many existing algorithms yield inconsistent performance, often failing to outperform baseline Empirical Risk Minimization (ERM) across diverse scenarios and necessitating newer algorithms which can handle scenarios where existing algorithms currently underperform. Furthermore, high algorithmic complexity frequently limits interpretability and offers only an indirect means of addressing spurious correlations. We propose eXplaining to Learn (eX2L): an interpretable, explanation-based framework that decorrelates confounding features from a classifier's latent representations during training. eX2L achieves this by penalizing the similarity between Grad-CAM activation maps generated by a primary label classifier and those from a concurrently trained confounder classifier. On the rigorous Spawrious Many-to-Many Hard Challenge synthetic data benchmark, eX2L achieves an average accuracy (AA) of 82.24% +/- 3.87% and a worst-group accuracy (WGA) of 66.31% +/- 8.73%, outperforming the current state-of-the-art (SOTA) by 5.49% and 10.90%, respectively. Beyond its competitive performance, eX2L demonstrates that functional domain invariance can be enforced by explicitly decoupling label and nuisance attributes at the group level.

Real-world datasets are inherently heterogeneous, yet how per-class structural differences and sampling imbalance shape the training dynamics of diffusion models-and potentially exacerbate disparities-remains poorly understood. While models typically transition from an initial phase of generalization to memorizing the training set, existing theory assumes homogeneous data, leaving open how class imbalance and heterogeneity reshape these dynamics. In this work, we develop a high-dimensional analytical framework to study class-dependent learning in score-based diffusion models. Analyzing a random-features model trained on Gaussian mixtures, we derive the feature-covariance spectrum to characterize per-class generalization and memorization times. We reveal the explicit hierarchy governing these dynamics: class variance is the primary determinant of learning order-consistently favoring higher-variance classes-while centroid geometry plays a secondary role. Sampling imbalance acts as a modulator that can reverse this ordering and, under strong imbalance, forces minority classes to acquire distinct, delayed speciation times during backward diffusion. Together, these results suggest that diffusion models can memorize some classes while others remain insufficiently learned. We validate our theoretical predictions empirically using U-Net models trained on Fashion MNIST.

Diffusion language models (DLMs) have recently emerged as competitive alternatives to autoregressive (AR) language models, yet differences in their activation dynamics remain poorly understood. We characterize these dynamics in LLaDA-8B and identify a striking layer-collapse property: a few early layers exhibit highly similar, collapsed activation patterns dominated by a single large super-outlier persisting over a long token range. Despite its apparent redundancy, this outlier is critical: pruning it causes outputs to degrade into repetitive random token loops. Paradoxically, layers in LLaDA contain more redundant representations overall, with redundancy most pronounced in earlier layers – the reverse of AR models, where deeper layers grow redundant due to undertraining. Our analysis indicates that layer collapse in DLMs is not driven by undertraining but by overtraining: a dominant outlier becomes an indispensable information carrier while remaining representations collapse into redundant structure. These findings have strong practical implications, verified through controlled pre-training experiments. DLMs are surprisingly robust to compression: LLaDA under 3-bit GPTQ quantization drops only -1.8% on GSM8K, whereas Llama-3.1-8B drops -64.7%. Optimal sparsity allocation also reverses between families: at 50% average sparsity, allocating more to early layers in LLaDA yields +8.4% over the reverse strategy, while the same allocation costs Llama -8.4%. Our findings reveal that the DLM training objective fundamentally reshapes layer dynamics relative to AR models, with direct consequences for compression and deployment. Code: github.com/Conzel/super-outlier-dlm.

We introduce a new generative modeling framework, Flow Matching with Arbitrary Auxiliary Paths (AuxPath-FM), which generalizes conditional flow matching by incorporating an auxiliary variable drawn from an arbitrary distribution into the probability path. Unlike prior methods that restrict auxiliary components to Gaussian noise, AuxPath-FM allows the variable $\eta$ to follow any distribution, producing trajectories of the form $X_t = a(t)X_1 + b(t)X_0 + c(t)\eta$. We theoretically demonstrate that this construction preserves the continuity equation and maintains a training objective consistent with the marginal formulation. This flexibility enables the design of diverse probability paths using various priors, including Gaussian, Uniform, Laplace, and discrete Rademacher distributions, each offering unique geometric properties for generative flows. Furthermore, our framework allows for specialized tasks such as label-guided generation by encoding structured semantic information into the auxiliary distribution. Overall, AuxPath-FM provides a principled and general foundation for probability path design, offering both theoretical generality and practical flexibility for diverse generative modeling tasks.

This paper presents a preliminary analysis of the ability of Chronos foundation model to process and internally represent frequency domain information. Foundation models that process time-series data offer practitioners a unified architecture capable of learning generic temporal representations across diverse tasks and domains, reducing the need for task-specific feature engineering and enabling transfer across signal modalities. Despite their growing adoption, the extent to which such models encode fundamental signal properties remains insufficiently characterised. We address this gap by analysing Chronos under controlled conditions, starting from the simplest class of signals: discrete sinusoids generated at fixed frequencies. Using lightweight online minimum description length probes applied to the decoder architecture, we test for the presence and separability of frequency information in the model's internal representations. The results provide insight into how frequential content is captured across the frequency spectrum and highlight regimes in which representation quality may degrade or require particular care. These findings offer practical guidance for users of Chronos in signal processing and information fusion contexts, and contribute to ongoing efforts to improve the interpretability and evaluation of foundation models for temporal data.

This work studies the robust evaluation of iterative stochastic purification defenses under white-box adversarial attacks. Our key technical insight is that gradient checkpointing makes exact end-to-end gradient computation through long purification trajectories practical by trading additional recomputation for substantially lower memory usage. This enables full-gradient adaptive attacks against diffusion- and Langevin-based purification defenses, where prior evaluations often resort to approximate backpropagation due to memory constraints. These approximations can weaken the attack signal and risk overestimating robustness. In parallel, stochasticity in iterative purification is frequently under-controlled, even though different purification trajectories can substantially change reported robustness metrics. Building on this insight, we introduce a memory-efficient full-gradient evaluation framework for stochastic purification defenses. The framework combines checkpointed backpropagation with evaluation protocols that control stochastic variability, thereby reducing memory bottlenecks while preserving exact gradients. We evaluate diffusion-based purification and Langevin sampling with Energy-Based Models (EBMs), demonstrating that full-gradient attacks uncover vulnerabilities missed by approximate-gradient evaluations. Our framework yields stronger state-of-the-art $\ell_{\infty}$ and $\ell_{2}$ white-box attacks and further supports probing out-of-distribution robustness. Overall, our results show that exact-gradient evaluation is essential for reliable benchmarking of iterative stochastic defenses.

Order-Agnostic autoregressive models have demonstrated strong performance in deep generative modeling, yet their use in settings with incomplete data remains largely unexplored. In this work, we reinterpret them through the lens of missing data. First, we show that their standard training procedure on fully observed data implicitly performs imputation under a missing completely at random mechanism, resulting in robust out-of-sample imputation performance in settings with high missingness. Second, we introduce the first principled framework for training them directly on incomplete datasets under general missingness mechanisms. Third, we leverage their amortized conditional density estimation to perform active information acquisition, i.e., sequentially selecting the most informative missing variables for downstream prediction or inference. Across a suite of real-world benchmarks, our Missingness-Aware Order-Agnostic Autoregressive Model (MO-ARM) consistently outperforms established imputation baselines.