machine learning platform developers

Abstract: Hardware-accelerated spiking neural networks (SNNs) implemented on memristive crossbar arrays offer significant gains in energy efficiency, while device variability and analog noise are sometimes assumed to hinder information extraction. However, whether these hardware effects improve, preserve, or degrade privacy after deployment remains largely unexplored. Therefore, we ask whether model-inversion leakage characterized on an SNN remains representative after the same model is mapped to a non-ideal memristive substrate, and how hardware non-idealities and reliability faults affecting synapses, neuron thresholds, and spike processing reshape that leakage. To study this problem, we introduce Spiking Domain Model Inversion (SDMI), an inversion method that combines Expectation over Transformation (EOT) with surrogate gradients to optimize inputs through stochastic analog behavior and non-differentiable spiking dynamics. We use SDMI consistently across the SNN model, its memristive SNN, and the same memristive SNN under additional hardware perturbations. Across four datasets, reconstruction leakage does not vary monotonically with model accuracy or hardware variability. At comparable accuracy levels, neuron-threshold perturbations increase reconstruction fidelity, measured by SSIM, by up to 51.8%, whereas spike-level perturbations primarily degrade utility without consistently improving reconstruction. Weight noise, in contrast, reduces leakage under the same inversion procedure. These results show that hardware variability is not inherently privacy-preserving: model inversion leakage in memristive SNNs depends on the hardware state and cannot be inferred from model accuracy alone.

Abstract: Continuous diffusion language models involve two coupled design choices: how to corrupt continuous token representations and how to invert that corruption at generation time. Recent work has explored language generation on the unit sphere through geodesic and von Mises--Fisher (vMF) paths, raising the question of which forward process to use on this geometry and how much that choice matters relative to inference. We introduce Spherical Brownian Diffusion, using Brownian motion as the canonical isotropic diffusion on the sphere. Its heat semigroup provides a principled forward process, an additive dimension-free noise clock, and an exact ancestral bridge, while a clean-token posterior learned with cross-entropy supports multiple samplers. To separate corruption from inference, we derive high-dimensional approximation rates between Brownian and moment-matched vMF corruption and use this relationship for controlled substitutions. On Sudoku, Brownian and moment-matched vMF give nearly identical results across four samplers and five budgets under otherwise identical settings. In contrast, with the corruption process fixed, sampler choice produces a 25.0-point spread at a sampler budget of 180 calls. Using the ancestral bridge, we split this gain into a memory effect from re-noising the state, present on Sudoku and GSM8K, and a token-selection effect on GSM8K and largest at small budgets. Across reasoning tasks, categorical refinement improves performance, whereas unconditional text generation benefits from stochastic updates that preserve diversity.

Abstract: Training-free few-shot adaptation of vision-language models typically constructs a classifier from CLIP's final projected image representation, leaving intermediate representations unused. Combining them is not straightforward. Class scores from different representations differ in spread by orders of magnitude for the same query. Direct averaging allocates influence by numerical scale rather than predictive information. We propose Layer-Ensemble Gaussian Discriminant Analysis (LE-GDA), a method that fits regularized shared-covariance discriminants independently to intermediate and final representations extracted in one frozen encoder pass. LE-GDA standardizes each discriminant's class scores per query, selects a small representation multiset on labeled validation data, and combines it with the zero-shot classifier without gradient-based optimization. Across 11 datasets, four CLIP backbones, five shot counts, and three independently sampled K-shot training sets, LE-GDA achieves the highest mean accuracy of all evaluated methods on every backbone, raising the 16-shot mean from 78.01% to 79.54% over a matched GDA-CLIP (final-feature) baseline and improving that baseline for every backbone-shot combination. At one shot, LE-GDA retains a 1.20-point mean gain over the baseline even though covariance estimation is unavailable and per-query standardization is neutral. A control with richer per-class structure inside a single representation does not reproduce the gain, indicating that the benefit comes from combining representations. These results show that intermediate CLIP representations carry information beyond the final embedding and can improve training-free few-shot adaptation. Code and evaluation resources are available at https://github.com/pavodi-NM/le-gda.

Abstract: Domain generalization (DG) aims to improve the generalization of the model to any unseen domains. Targeting this goal, existing DG methods mainly focus on learning domain-invariant features or prompts from the given source domains. However, these methods rely on source-domain data, and collecting and annotating large-scale single- or multi-source datasets is often infeasible, limiting the practical applicability of DG. To tackle this challenge, we propose {\underline{Do}}main \underline{P}rompt Learning (DoP), a method built on a large vision-language model (e.g., CLIP) for zero-shot domain generalization, which recognizes unseen classes in unseen target domains without access to source-domain data. DoP essentially learns the target domain knowledge via prompt both in text and visual modalities to overcome the domain shift in text-visual modalities under the test-time adaptation formulation. Specifically, DoP first takes text domain learning (TDL) to simulate various distribution shifts by learning diverse style tokens implicitly in the text space with inter-style diversity loss and intra-style consistency loss. Subsequently, DoP employs visual domain learning (VDL) to further refine the target-domain-relevant style token generated by TDL by making use of the test visual data with self-supervised learning. Extensive experiments on popular DG benchmark datasets (e.g., PACS, VLCS, OfficeHome, and DomainNet) and ImageNet show that our DoP consistently outperforms the state-of-the-art DG approaches. Codes are available at: https://github.com/shiming-chen/DoP.

Abstract: AI control protocols aim to get useful work from models that may be misaligned by monitoring their behaviour. One approach uses a second copy of the same model as the monitor. This avoids relying on a weaker reviewer the model could outsmart, but risks collusion: the model, now monitoring itself, may cover up its own misbehaviour. One remedy is to fine-tune the monitor on whatever examples of misbehaviour the control team has caught. We ask whether this training also removes collusion that the control team has not yet found. We fine-tune Llama-3.3-70B and Qwen2.5-72B to install five policy-monitor behaviour pairs, correct the monitor on 300 honest reviews of one caught collusion, and measure collusion on the other four in unseen prompt formats. Averaged over the three seeds of each family, correction reduces held-out collusion from 91% to 14-20% under three of the four methods we compare (individual seeds 10-32%), while a control fine-tuned on the same number of unrelated examples leaves 56-93%. Family-average accuracy on control items changes by at most five points and capability is preserved, apart from a drop in mathematical accuracy on one Llama seed. Under the tested training configuration, methods with similar caught-task accuracy differ in held-out transfer, from 58% to 83% average reduction, so the caught behaviour alone does not reveal how far a correction reaches. Catching one collusion lets the control team reduce, but not remove, collusion it has not caught, by retraining the monitor alone.

Abstract: Token merging is an effective approach to reducing KV-cache memory consumption in autoregressive video generation by clustering temporally redundant tokens and retaining only a small set of representatives. Existing methods, however, typically apply a uniform merging ratio across all attention layers and heads. We observe that different layers and heads exhibit substantially different yet jointly correlated sensitivities to token merging, and that these sensitivity patterns remain consistent across prompts. Based on this observation, we introduce CALM, a Calibrated Allocation for Layer-Head Merging that uses offline calibration to improve online cache compression. Exhaustively evaluating all possible layer-head allocations is prohibitively expensive. To enable efficient calibration, we show that attention-output error strongly correlates with end-to-end video quality and can therefore serve as a lightweight proxy for allocation risk assessment. We evaluate different allocation strategies using this proxy and construct a fixed layer-head risk map offline. Given a cache budget at inference time, CALM uses dynamic programming to determine the best allocation of merging ratios. Experiments across three different autoregressive video models demonstrate that CALM consistently outperforms uniform token merging under the same memory budget, by reducing the LPIPS to original videos by around 21% with higher video-quality scores.

Abstract: Data selection plays a central role in modern large language model training, as the quality of training data strongly influences downstream performance. A common approach is to curate a development set that encodes the desired skills (e.g., reasoning or multilinguality) and select data that are similar to them. Existing methods compute similarity using cosine distance over either the model's hidden states or gradients. However, we observe that this max-similarity strategy induces a mode-seeking bias: the selected data clusters tightly around the development samples. To address this, we propose ICLHerd, a data selection framework that leverages the von Mises-Fisher kernel to balance mode-seeking with mode-averaging. We further introduce in-context learning as a validation proxy, enabling efficient hyperparameter search over the balance trade-off. Results show that ICLHerd outperforms existing LLM data selection algorithms, highlighting the importance of balancing mode-seeking and mode-averaging.

Abstract: Neural Posterior Estimation (NPE) fits an amortized conditional density estimator to a simulated training set of complete-data realizations. NPE's training objective, log-likelihood, favors posterior approximations that assign arbitrarily high density to the training points, and typical families of amortized conditional density estimators (e.g., normalizing flows, Gaussian encoders) are flexible enough to oblige. The resulting policy of extreme overconfidence translates poorly to held-out data. Early stopping and other regularizers can prevent this degeneracy, but they are limited to selecting posterior approximations along an unfavorable optimization trajectory set by in-sample residuals, where the posterior mean is underfit while the posterior variance is overfit. To avoid this trade-off, we propose Cross-Fitting NPE (CF-NPE), which uses sample splitting to separate conditional-mean estimation from the estimation of the posterior's higher moments. Using $K$-fold cross-fitting, CF-NPE fits $K$ posterior-mean regressors and then fits a conditional density estimator to the resulting out-of-fold residuals. In experiments with synthetic data and in a cosmological application, CF-NPE achieves higher held-out log-likelihood than standard NPE and typically improves calibration. CF-NPE is most useful in high-dimensional, simulation-limited settings.

Abstract: Most vision-language models (VLMs) treat the vision encoder as a static feature extractor: it encodes an image once, independently of the question, and leaves subsequent reasoning to the language model. Yet different questions may require different visual operations and processing depths. We present AdaLoop, a framework for question-conditioned visual recurrence in VLMs. AdaLoop adapts visual computation to the question and reuses the encoder's transformations across multiple iterations, with the pretrained vision weights frozen. We also introduce VisRO, a procedural dataset and generation pipeline for training and evaluating vision-centric reasoning. On six VisRO tasks, AdaLoop outperforms standard fine-tuning and question-conditioned single-pass baselines with matched training data and optimizer updates. These gains hold across three VLM families on both in-distribution and out-of-distribution examples. Our findings suggest that pretrained vision encoders can contribute to reasoning through task-directed, iterative visual computation.

Abstract: What makes a deep network train, and what shape is its loss surface? We answer both with one construction at two scales: a polynomial activation f (z) = z + λz 3 , clamped and projected onto the unit sphere, plus an unnormalized identity path. As a deep stack it trains to one hundred layers on MNIST with no normalization machinery; as a small model its square loss is a polynomial system we solve exactly. The small scale explains the large one: it prescribes how to design, initialize and train the stack.

Abstract: Recent work has explored visual reasoning through image generation, where models solve visual-reasoning problems by directly generating solution images. Intuitively, iterative samplers such as flow matching are well suited to this task, since each sampling step may resemble one reasoning step. However, in this work, from a circuit complexity perspective, we show that this intuition is misleading. We prove that as long as the model outputs satisfy certain mild regularity conditions, the trajectory of the vanilla flow-matching sampler can be simulated by a constant number of sequential rounds of parallel model evaluations. Consequently, additional sampling steps cannot increase the expressive power of a constant-depth backbone model under the standard complexity-theoretic model of transformers. We then reveal that self-conditioning, which conditions the model on the previous step's estimate of the final image, plays a previously unrecognized role in removing this bottleneck. Controlled experiments on diverse visual-reasoning tasks show that flow matching with self-conditioning achieves higher accuracy and greater reasoning depth than vanilla flow matching, and its performance improves with more sampling steps. Theoretically, we prove that with self-conditioning, a constant-depth backbone can achieve expressive power equivalent to that of space-bounded Turing machines. Our study suggests a design principle for future visual reasoning research: carry more information between sampling steps, so that additional steps translate into deeper reasoning.

Abstract: Diffusion language models reconstruct masked tokens in parallel, offering an alternative to autoregressive generation. Yet their Transformer backbones typically carry visible evidence and evolving masked predictions through a single residual stream. We equip masked diffusion models with multiple residual streams that persist through depth and adaptive routing between them, then investigate how they support these complementary roles. On the pretraining task, two streams reduce reconstruction perplexity by 5.71% on validation data and 5.13% on average across external corpora compared with standard residual connections. We observe that, without predefined roles, visible and masked positions learn distinct routing patterns. Causal interventions show that this distinction organizes read and write routing, while token-specific write strength remains important within each role. Guided by these findings, a model trained with role-conditioned routing recovers half the validation gain from adding a second stream. State interventions explain the benefit of the additional stream: it preserves early token representations for later reuse while evidence accumulates at masked positions. This mechanism recurs at a larger model scale, showing how persistent streams support context preservation alongside prediction refinement.

Abstract: Random walk (RW)-based algorithms are widely used in decentralized learning due to their low communication overhead and scalability. While most existing methods rely on a single token, multiple tokens can explore the network in parallel and exchange information upon meeting. However, the fundamental benefits of such meeting-triggered interactions remain unclear. In this work, we investigate multi-token RW learning under two information exchange policies: \textsc{Average-Gradient} and \textsc{Average-Model}, which average stochastic gradients and updated models, respectively, when tokens meet. We establish iteration complexity lower bounds for non-convex, convex, and strongly convex objectives on complete graphs, which apply to both exchange policies. Compared with single-token RW learning, the scaling factor in these bounds is reduced by a factor of $m^2$ for non-convex and convex objectives, and by an exponentially in $m$ for strongly convex objectives, where $m$ denotes the number of tokens. Numerical experiments reveal distinct behaviors of the two policies under two performance criteria: mean-token loss (the average loss across all tokens) and best-token loss (the minimum loss across all tokens). \textsc{Average-Model} achieves lower mean-token loss and is relatively insensitive to the network-to-token ratio $N/m$, where $N$ is the number of nodes. In contrast, \textsc{Average-Gradient} achieves lower best-token loss but exhibits a trade-off between parallel exploration and gradient interference as $N/m$ decreases.

Abstract: Many widely used activation functions, including ReLU, GELU, Swish, and Sigmoid, produce predominantly positive outputs, resulting in a nonzero post-activation mean that we term activation bias. Despite its ubiquity, the effect of activation bias on learning dynamics remains poorly understood. Using controlled models, we show that activation bias accelerates initial loss reduction but is accompanied by overshooting, non-smooth trajectories, and a steeper loss landscape. This early advantage also coincides with the model's use of task-relevant information encoded in the activation bias component. In contrast, reducing activation bias slows initial learning but ultimately yields a lower final loss. Experiments on larger vision and language models exhibit a similar pattern: smaller activation bias is associated with better late-stage performance. Thus, this seemingly minor property can substantially redirect learning and warrants explicit consideration in future activation function design.

Abstract: Diffusion feature, i.e., intermediate activations extracted from diffusion model backbones, has emerged as a promising approach for unifying generation and vision understanding (discrimination). Despite the advancement of diffusion backbones into Diffusion Transformers (DiT), it has been found in previous literature that DiT models tend to yield weaker features despite their stronger generation capability. This poses a threat to good generation-understanding unification. In this work, we show via controlled from-scratch training that long skip connections (LSCs) affect feature quality: a DiT with LSCs can produce better features. Furthermore, through detailed diagnosis, we reveal the mechanism causing this effect. Specifically, the input to a diffusion backbone contains both the latent image and the diffusion noise, where the diffusion noise, though appearing in the form of noises, actually reflects the information of the diffusion trajectory. To make the final prediction of a noisy variable ($v$ or $ε$), the diffusion backbone needs both signals. However, carrying both signals through the main backbone effectively "pollutes" the image signal with noises, leading to poor feature quality. In comparison, LSCs can shortcut diffusion noises directly to where they are needed, leaving much weaker noise signals in the main backbone, thus improving feature quality. To name the broader impact, this reveals that LSCs are not a mere trick but closely related to the diffusion objective. The code is at https://github.com/Darkbblue/dit-lsc.

Abstract: Flow matching has emerged as a scalable paradigm for training high-quality generative models, but sampling from the learned probability flow requires many network evaluations. Distillation can reduce this cost to one or a few evaluations; however, one-step generation often sacrifices quality, making few-step generation the practical operating regime. Existing few-step methods perform their iterative computation along the probability flow and therefore require a fixed, manually chosen timestep discretization. This discretization is often chosen heuristically and is expensive to tune; it may also be restrictive when refinement difficulty differs across samples or spatial locations. We introduce data-space iteration, a few-step generation framework that removes flow discretization altogether. Starting from noise, a shared generator directly refines its prediction in data space, with every iteration trained to produce the best sample permitted by its capacity. Our formulation integrates with distribution matching distillation (DMD) with minimal changes, enabling a controlled comparison between iteration methods under matched training settings. On class-conditional ImageNet 256x256, data-space iteration outperforms standard discretization baselines and matches or improves upon variants selected through schedule search, without requiring schedule-specific training. These results show that data-space iteration provides a simple and effective alternative to discretized flow-space iteration for fast generation.

Abstract: Existing model marketplaces typically trade and select individual models as indivisible units, limiting their ability to exploit complementarities among heterogeneous experts. This paper proposes an MoE-based model market framework that lifts Mixture-of-Experts from a model-level learning architecture to a market-level coordination mechanism. In this framework, brokers use gating networks to coordinate multiple heterogeneous experts and deliver a composite model service. We formalize the market participants, service workflow, expert cost structure, and a welfare objective that combines predictive utility with heterogeneous execution costs. We then derive a cost-aware gating mechanism and market-aware training objective, and introduce a cost-adjusted revenue allocation rule that distributes residual revenue according to realized expert participation and execution cost. We also establish basic theoretical properties of the allocation rule, including budget balance, participation monotonicity, and cost sensitivity. Experiments over five random seeds on fifteen tabular and image benchmarks use independently trained and frozen neural and tree-based experts together with latency-derived execution costs. MoE Market achieves the highest mean welfare on all fifteen datasets and a lower mean expected cost than Standard MoE in every case, while maintaining competitive predictive performance. The allocation experiments further demonstrate systematic sensitivity to expert participation and cost, together with substantially lower computational overhead than exact Shapley allocation. These results suggest that MoE can serve as a market-level coordination principle for collaborative, cost-aware, and economically grounded model marketplaces.

Abstract: Autonomous driving capabilities depend strongly on the distribution of scenarios encountered during training. Existing methods commonly construct or dynamically adapt training scenario distributions using surrogate criteria such as realism, difficulty, or risk. However, these predefined surrogates may misrepresent training value, leading to inefficient use of training resources. To address this limitation, we propose a Lamarckian evolutionary framework that replaces surrogate-based guidance with competition among candidate distributions. We formulate training strategy discovery as a multi-stage bilevel optimization problem and use Lamarckian evolution algorithm to approximate its solution. At the outer level, Darwinian crossover, mutation, and selection explore the scenario distribution space; at the inner level, policy learning acquires new capabilities, and Lamarckian inheritance transfers them to subsequent stages, allowing scenario distributions and policy capabilities to co-evolve. The resulting evolutionary trajectories reveal recurring stage-wise regularities among high-value distributions, characterized by capability accumulation through stage-wise challenge rotation. We further distill these regularities into a lightweight, reusable Lamarckian Training Strategy. Experiments show that, compared with the baseline, the complete framework reduces performance loss by up to 25.07%, while the lightweight strategy still achieves a 19.13% reduction. These results demonstrate that evolutionary competition can both discover effective training strategies and reveal reusable stage-wise patterns in how the value of training distributions changes with policy capability. Code is available on GitHub.

Abstract: Hessian spectra at trained models in deep learning exhibit a persistent pattern: eigenvalues organize into distinct clusters, including a large bulk near zero and a few isolated outliers. This paper shows that a natural account of these spectral phenomena emerges when the original setting is understood as a departure from a nearby, otherwise hidden, highly symmetric reference. Modifications, including changes to the architecture, data distribution, or parameter metric, expose a nearby reference configuration whose Hessian exhibits rich invariances-ones not accounted for by weight symmetries. There, symmetry enables a precise description of the spectra, forcing high-dimensional kernels and eigenvalues of large multiplicity. Returning to the original configuration breaks the Hessian symmetry and thereby produces the observed hierarchy of clusters and outliers. The framework is developed in some generality, with a detailed analysis of three-layer ReLU networks and applications to convolutional, graph, and transformer models, as well as to the NTK. The same mechanism is further shown to yield analogous spectral structures in layerwise Hessians and the Gauss-Newton matrix.

Abstract: Diffusion models generate data through a continuum of denoising problems, and are widely observed to reveal coarse structure before fine detail. Yet, this intuition is mostly empirical and qualitative. We introduce feature information dynamics, an information-theoretic framework for localizing when a feature is generated during diffusion. Using the I-MMSE identity, we connect the rate of feature mutual information change to a gap between optimal unconditional and feature-conditional denoising losses, yielding practical estimators for feature information density. We further develop a chained decomposition that separates shared from incremental information in a feature hierarchy. We use this framework first to quantitatively confirm spectral autoregression in pixel diffusion, and then to extend the analysis beyond frequency: under a class $\to$ mask $\to$ Canny conditioning chain, the per-feature information densities differ across pixel, SDVAE, VAVAE, and RAE, exposing fundamental differences between these representations and suggesting that ordered generation could be beneficial for training diffusion models. Our code is available at https://github.com/AI4Science-WestlakeU/feature-information-dynamics.

Abstract: Matrix completion underlies problems from tabular imputation to causal inference, yet existing tabular foundation models treat it as entry-by-entry prediction, repeating context for every target and discarding the matrix's two-dimensional structure. We introduce MatrixFormer, a pre-trained matrix-native transformer that predicts a full distribution for every missing entry in a single forward pass. MatrixFormer is trained entirely on synthetic low-rank and latent-factor matrices under diverse missingness patterns. Applied zero-shot and with the same model weights, MatrixFormer achieves competitive performance on causal inference panel-data tasks, language-model benchmark-score completion, tabular imputation, and recommendation systems matrix completion. These results position MatrixFormer as a general-purpose foundation model for matrix completion.

Abstract: Neural Architecture Search (NAS) has proved to be a strong alternative to manual network design, but applying it to 3D medical image segmentation is limited by two well-known costs, large annotation budgets and multi-GPU clusters. Thus, this paper introduces Active-DiNTS (Active Differentiable Network Topology Search), an approach that embeds pool-based Active Learning (AL) into a bi-level differentiable topology search to perform architecture discovery and label curation jointly. At each query round, unlabeled MRI volumes are ranked by one of three uncertainty signals (Entropy, Variance, or Standard Deviation), and only the top-ranked volumes are sent to an oracle for annotation. The new labels feed two interlocked stages. Network weights are updated in an outer loop, while the macro/micro topology of a U-Net-style backbone is refined in an inner loop. Three AL regimes (weights-only, topology-only, joint) expose the speed-accuracy trade-off. Evaluations on the Medical Segmentation Decathlon (MSD) Task01 BrainTumour benchmark showed that Active-DiNTS surpasses DiNTS, C2FNAS, and nnU-Net in Dice, with gains of about 10 percentage points on Edema and 5 points on Non-Enhancing core, on a single GPU and using a fraction of the labeled volumes. The discovered architectures are denser and more FLOP-heavy than prior baselines, but remain competitive in trainable parameters and peak memory; the fastest search variant finishes in under 0.25 GPU-days, over 27x faster than the eight-GPU DiNTS search. Together, these results indicate that pairing differentiable NAS with active data acquisition is a practical recipe for accurate 3D segmentation under realistic constraints.

Abstract: A decade after its inception, Inception crop has become the standard crop-based data augmentation method for training deep vision models. Not only is its practice of uniformly sampling crop scale and aspect ratio widely adopted, but also its lower and upper bounds, with the scale lower bound being the sole exception that is sometimes tuned. It is therefore surprising that the standard implementation in the TensorFlow / JAX ecosystem samples crop scale with probability density function $f(A) \propto \frac{1}{\sqrt{A}}$ unlike the PyTorch counterpart, which follows the original description. Motivated by this discovery, we train 522 ViT-S/16 models on the ImageNet-1k dataset with various training budgets and crop scale distributions. We reach $78.78\pm0.09$ top-1 val. accuracy with 90 epochs of training budget and find that 1. Higher training budget requires stronger augmentation; 2. Lower tail of the distribution of the crop scale determines the augmentation strength of Inception crop; 3. Models trained with higher training budget exhibit sparser saliency, regardless of the crop scale distribution or weight decay. Based on 2. we revisit the performance of Beta crop, whose softer cutoff allows it to optimize model performance across training budgets with less compromise. We replicate 1. and 3. with Scion optimizer in addition to AdamW, suggesting that the results may be general.

Abstract: Semantic Knowledge-driven Contrastive Learning (SKCL) uses a language model to decide which classes are related, and pulls each image towards the prototypes of its semantic neighbours. On CIFAR-100-LT (beta = 100) it reports 54.02% top-1 accuracy, 2.01 points above Balanced Contrastive Learning (BCL), the method it builds on. The code and the class descriptions are not public. I reimplement SKCL, BCL and ConCutMix in one framework, check it against the public baseline code, and run every configuration with three seeds. The two baselines reproduce within 1.5 points, but SKCL built on BCL, as the paper describes it, ends up 0.56 points below BCL. To find out why, I add SKCL to the authors' own ConCutMix code. Trained for the paper's 300 epochs, it reaches 53.69, only 0.33 below the published number. At the same budget, however, the semantic graph adds just 0.23 points over ConCutMix, while training ConCutMix for 100 more epochs adds 1.04. Together with ConCutMix's published lead over BCL (1.15), this explains the claimed gain. A BCL model that never sees the graph already shares 41.2% of the graph's top-2 neighbours with its own most-confused classes (2.0% by chance), which shows why the graph adds so little on these benchmarks. I also test several changes to SKCL. Combining it with the CutMix branch improves it by 1.06 points, and an adaptive version of the graph improves it slightly (+0.34 and +0.28 in two codebases), although these gains are within seed noise.

Abstract: Quasi-Newton (QN) methods have long been among the most effective methods for large-scale unconstrained convex optimization. Two obstacles have limited their use in deep learning: non-convexity and enormous parameter sizes. We introduce SoftServe, a family of QN methods designed to overcome these obstacles without line searches or ad hoc curvature corrections. SoftServe derives positivedefinite curvature estimates from the variational objective of Berglund et al. (2025), even in the presence of negative curvature. We develop diagonal and Kroneckerfactored variants that preserve positive definiteness by construction and scale to massive neural networks. Finally, SoftServe relies on the stable coupled Newton-Schulz iteration for the required matrix operations, replacing costly matrix decompositions with GPU-friendly matrix multiplications. SoftServe excels on problems that are severely ill-conditioned, including tasks such as recurrent networks, deep autoencoders, physics-informed neural networks, and a 136M-parameter physics-informed diffusion model, often achieving lower losses than established baselines including Adam, Muon, and SOAP.

Abstract: Curvature is often treated as an intrinsic property of a representation, although its empirical effect also depends on coordinate scale, learned logit temperature, and numerical safeguards. We study this interaction in hZACH-ViT, a compact Vision Transformer with Euclidean, Poincare, and spherical prototype heads. The backbone architecture, seed-specific initialization, 50-per-class training subset, and optimization protocol are matched across three MedMNIST datasets and five seeds. At the fixed comparison curvature $c=1$, Poincare has the lowest class-macro PGD attack-success rate in all 12 dataset-budget cells and under a stronger CE+DLR multi-restart attack on all three datasets, but it also has the lowest clean MacroF1. An end-to-end curvature intervention changes the interpretation. Reducing Poincare curvature to $c=0.1$ improves clean MacroF1 in every one of the 15 paired seed-dataset comparisons and removes hard boundary clipping, yet on OrganAMNIST it increases strong attack success from $89.7\%$ to $99.3\%$ (paired difference $+9.57$ points; 95\% hierarchical bootstrap CI $[+5.52,+14.03]$). At $c=1$, $40$-$47\%$ of clean Poincare features are hard-clipped, the radial Jacobian of the inherited map is nearly zero, and dimensionless attack trajectories are unusually long and inefficient. The spherical head provides a control: its curvature change is an exact scale gauge to floating-point precision and produces much smaller attack differences. These results do not establish intrinsic hyperbolic robustness. They identify an implementation-sensitive regime in which curvature, scale, and proximity to the Poincare boundary jointly organize clean recognition and adversarial representation motion.

Abstract: Private Evolution (PE) is a differentially private algorithm for synthetic data generation. While it can be viewed as a Wasserstein learning algorithm, it performs much better in practice than worst-case Wasserstein analyses would predict. We recast PE as generative model-augmented Wasserstein learning. We show theoretically that when we take into account the use of a generative model that is able to capture something about the true distribution, then we can obtain much better performance bounds. For example, if the generator gives samples in the same low-dimensional space as the distribution, then sample complexity depends on intrinsic, not ambient, dimension. We also show that standard variants of PE can fail to converge on simple well-clustered instances, and propose a new geometry-aware version of PE with provable convergence on such instances. Experimentally, we show that our new algorithm is competitive with standard baselines and can improve recall.

Abstract: 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.

Abstract: Contextualization, the core operation of language modeling, transmits information across words to build sentence-specific word representations. Prior works mainly study contextualization, focusing on individual words and attention heads as a growing discrete dictionary, lacking a global view of their general behavior. Therefore, we propose a general principle: Globally, we find and estimate that different words carry different amounts of information, and less-informative words tend to absorb more contextual information. Specifically, these low-information words do not absorb contextual words uniformly, and finer-grained selectivity enables more precise routing to promote information transmission between matched words. Moreover, to find what mechanism causes such processing, we reinterpret attention heads as channels gated by their singular vectors and find that: (1) these singular vectors point to the hidden states of more informative words, allowing such words to write their information to others more strongly to act as information sources, and vice versa; and (2) these singular vectors can be viewed equally as hidden state features, enabling automated interpretation of attention heads beyond prior heuristic head discovery, also embedding heads into a continuous space rather than treating them as discrete, independent dictionary entries.

Abstract: Hyperbolic space can embed intrinsic hierarchies in data with low distortion due to the exponential growth of volume with distance from the origin. However, current Lorentz transformer blocks are formulated in ambient coordinates, where numerical errors increase at large radii due to instability in the Lorentzian inner product, leading many models to limit the radius to avoid this issue. This prevents us from utilizing the regions of hyperbolic space that motivate the geometry. To address this, we revisit the components of the transformer block in polar coordinates, where hyperbolic operations such as distance calculation and attention centroids can be computed without the numerical cancellation in their ambient-coordinate formulations. Specifically, we propose a polar fully connected layer that separately maps an embedding's direction and radius, allowing the radius of a feature to be learned rather than determined by the norm of a linear map. We additionally introduce horospherical shifts as relative positional encodings whose query-key distances grow logarithmically with the token gap. Finally, we reformulate the residual connection as average radius Lorentz boosts. Combining these components, we develop a fully hyperbolic transformer that substantially improves performance over Euclidean and hyperbolic baselines on standard vision tasks. We further evaluate our model on ImageNet and demonstrate its ability to generalize to other datasets using pre-trained weights, similar to Euclidean counterparts.

Abstract: Automatic fact-checking poses a significant challenge in Arabic natural language processing due to the scarcity of datasets and resources. In this manuscript, we introduce Arafa, a new large-scale dataset for fact-checking in Modern Standard Arabic, constructed through an automated framework leveraging large language models (LLMs). The dataset was constructed through a three-step pipeline: (1) claim generation from Arabic Wikipedia pages with supporting textual evidence, (2) claim mutation to generate challenging counterfactual claims with refuting evidence, and (3) an automatic validation step to validate that the generated claims are either supported or refuted by their accompanying evidence, or if the evidence does not provide enough information to judge the validity of the claims. The resulting dataset comprises 181,976 claim-evidence pairs labeled as supported, refuted, or not enough information. Human evaluation carried out on a test sample from the dataset demonstrated strong inter-annotator agreement (kappa = 0.89) using Cohen's Kappa for supported claims and (kappa = 0.94) for refuted claims. Automatic validation based on a human-evaluated sample achieved 86% accuracy for supported claims and 88% for refuted ones. To showcase Arafa's value as a resource for automatic Arabic fact-checking, four open-source transformer-based models were fine-tuned using Arafa, with the top-performing model achieving a Macro F1-score of 77% on the test data. In addition to Arafa being the first large-scale dataset for Arabic fact-checking, our framework presents a scalable approach for developing similar resources for other low-resource languages.

Abstract: Classifier-free guidance (CFG) improves conditional generation in Flow Matching, but strong guidance can distort the generated distribution and reduce diversity. We provide a geometric account of this behavior by viewing Flow Matching as a time-varying gradient flow and characterizing how CFG reshapes its underlying potential. This view explains how stronger alignment can be accompanied by mean displacement and trajectory concentration, and motivates controlling guidance through the model-implied terminal posterior mean. We therefore propose Posterior-Mean-Capped CFG (PMC-CFG), a training-free, per-sample method that adaptively retains the strongest feasible guidance without additional network evaluations. Experiments on synthetic and large-scale image-generation benchmarks show that PMC-CFG limits guidance-induced distortion and concentration while improving the alignment--diversity trade-off, with particularly strong benefits when nominal guidance is large.

Abstract: Contrastive self-supervised learning has achieved strong performance by learning representations from multiple augmented views of the same image. However, most existing methods construct positive pairs using independently sampled stochastic augmentations, which may alter semantic content and ignore the intrinsic geometry of the data distribution. In this work, we propose OTCLR, an optimal transport-aware framework for contrastive learning representations that generates geometry-consistent positive samples. Instead of directly contrasting two randomly augmented views, we construct intermediate views between the original image and its augmented variants through entropic optimal-transport displacement interpolation. These transport-interpolated samples serve as positive views that better preserve image structure while explicitly modeling spatial distributional geometry. To further promote smooth representation learning, we evaluate auxiliary Sinkhorn regularization terms that encourage transport-interpolated views to remain consistent with their endpoint images. The proposed method can be incorporated into standard contrastive learning pipelines without modifying the encoder architecture. Experiments on multiple benchmark datasets show that our approach improves representation quality and transfer learning performance compared with conventional augmentation-based contrastive learning baselines.

Abstract: Clustering is a fundamental task in data analysis, typically addressed through centroid-based methods such as K-means. In this work, we present a general framework for multi-domain clustering via measure quantization: given samples from multiple domains, we learn a shared set of cluster prototypes by minimizing a probability metric, such as the Sinkhorn divergence or the Maximum Mean Discrepancy, between each domain's probability measure and the measure of prototypes. Data points are then assigned to clusters either via nearest centroid, or via optimal transport, a collaborative strategy that couples all samples within a domain. A mini-batch optimization strategy makes both fitting and assignment scalable, reducing memory and computational cost while preserving clustering performance. Experimental results on 5 multi-domain benchmarks spanning image, audio and sensor data show that our Sinkhorn-based method consistently outperforms classical and multi-domain clustering baselines, and that this advantage persists when scaling to hundreds of thousands of samples.

Abstract: Vision Transformers (ViTs) are central to most modern vision models, yet obtaining input attributions that are fine-grained, faithful, and stable remains challenging. Layer-wise Relevance Propagation (LRP) has been adapted to transformer attention, but in ViTs it often produces noisy, unfaithful explanations. We show that the missing ingredient is the treatment of residual connections: cancellation effects in residual pathways lead to attribution explosion. Moreover, we find that these cancellations are substantially stronger in ViTs than in language transformers. To address this issue, we introduce Residual-aware Layer-wise Relevance Propagation (ResLRP), a simple extension of LRP whose propagation rules explicitly account for cancellations in residual branches, are exactly conservative, and provably bound relevance explosion. Causal channel-wise interventions confirm that residual cancellation, not a generic regularization effect, drives the instability. ResLRP substantially improves attribution quality across faithfulness and localization, evaluated on ViT architectures spanning supervised, self-supervised, contrastive, hierarchical, and multimodal families, as well as on the ground-truth-controlled FunnyBirds benchmark. The largest gains arise in modern Vision Language Models (VLMs), with +27-29% localization and up to 3.4x faithfulness scores. Beyond benchmarks, ResLRP localizes Sparse Autoencoder (SAE) features in input space, and our residual amplification measure serves as an architecture-level diagnostic predicting where attribution degrades.

Abstract: Semantic watermarking schemes such as Tree-Ring hide a detectable pattern in the initial noise latent of a diffusion model. Recent work shows these watermarks are not only removable but forgeable: an attacker who never sees the watermarking key can still produce images the genuine detector accepts. We reproduce the Reprompt forgery attack of Müller et al. against Tree-Ring on Stable Diffusion XL, using the authors' released code, on free-tier dual T4 GPUs with 14.6 GB of usable memory per device, substantially less per-GPU memory than the A40 hardware used in the original study. The attack reproduces. Over six trials of three arms we detect genuine images 6/6, clean images 0/6, and forged images 5/6, at 325-332 s per attack. Three further results came out of running it under constraint. The released detector computes a non-central $χ^2$ statistic and hands back only its CDF, so we recovered the discarded statistic; our recovery matches the released detector exactly, and two natural scores built from it separate the forged arm from the clean null at AUC 0.861 and 0.972 on the same eighteen observations. Running SDXL in half precision requires patching the pipeline's direct autoencoder calls, and a controlled probe confirms the patched path leaves the detector statistic unchanged. Finally, we report a prediction we made from reading the detector source that our measurements then contradicted. The notebook, the pinned fork and every measurement artifact are released with the paper.

Abstract: Federated learning over IEEE~802.11 shares the wireless channel among clients that send model updates. We use ns-3 to measure the frame-delivery ratio and saturation throughput for different client densities and offered loads. A separate FedAvg trainer uses the frame-delivery ratio as a first-order proxy for the update-admission probability and uses an equation to estimate communication time. The method does not simulate the delivery of a complete model update or measure end-to-end training time. Across 720 evaluated runs with two datasets, two data partitions, six client densities, six offered loads, and five seeds, all runs reached their predefined target accuracy within the round budget. Rounds-to-target changed little with offered load, while communication time-to-target increased by about two orders of magnitude across the client-density range. A Bianchi-anchored estimator produced a mean absolute percentage error from $2.3\%$ to $10.2\%$ on held-out configurations. This error is measured against communication time constructed from the same round-duration equation, not against independently measured completion time. We also compare uniform participation with persistent heterogeneous participation. The study does not detect a statistically distinguishable excluded-class accuracy gap over five seeds, but the confidence intervals are wide. The results apply only to the evaluated configurations and do not provide a general convergence or fairness guarantee.

Abstract: The learning-rate schedule is a consequential choice in training deep networks, yet the policies in common use are heuristic, and published comparisons are hard to read, because architecture, dataset, and budget tend to vary alongside the schedule. We study BrachistoneLR, a schedule built by mapping the vertical coordinate of the brachistochrone, the curve of fastest descent under gravity, onto the range between a peak and a floor rate. Expanding the definition shows it to be cosine annealing with the half-period set to E - 1 instead of E, the configuration a standard implementation gives when its period argument is one less than the number of epochs. The rate therefore reaches its floor at the last epoch trained rather than one epoch later, and we show this difference decays as E^-2, making it a short-horizon effect. We then benchmark six schedules over 72 runs on three image classification datasets (MNIST, Fashion-MNIST, CIFAR-10) and four architecture families (fully connected, convolutional, recurrent, residual), fixing the optimizer, data pipeline, and evaluation protocol so that only the schedule varies. Schedules that fall smoothly from peak to floor beat the constant rate and calendar-based decay by margins that grow with task difficulty, reaching 2.5 points of dataset mean on CIFAR-10. Within that leading group, BrachistoneLR, cosine annealing, and warmup-cosine lie within 0.06 accuracy points and 0.17 of a mean rank, which one seed per configuration cannot separate. BrachistoneLR is best on both residual networks and has the highest CIFAR-10 mean, and it sets no milestones, decay factor, warmup length, or restart period. We conclude that the shape of a schedule matters more than its parameterization, that the choice of whether to use a smooth schedule matters more than the choice among them, and that the terminal-rate distinction is worth attention only over short horizons.