arXiv subjects

Daniel Soudry

Publications and source records attributed to Daniel Soudry.

At least 19 recordsLinked to original sources

arXiv2026

Convergence of Continual Learning in Homogeneous Deep Networks

We characterize weakly regularized continual classification in homogeneous models as sequential projections onto task margin sets. This result generalizes prior analyses restricted to either stationary (single-task) deep models or continual linear models. We show that global convergence generally fails, even for simple models linear in data but nonlinear in parameters. Nevertheless, by leveraging results from nonconvex projection theory, we identify regularity properties of homogeneous deep networks that guarantee local linear convergence under random and cyclic task sequences. Finally, we extend our analysis to continual regression, unifying the framework for homogeneous models.

cs.LG↗

arXiv2026

Workspace Optimization: How to Train Your Agent

Modern agents built on frontier language models often cannot adapt their weights. What, then, remains trainable? We argue it is the agent's \emph{workspace}, the structured external substrate it reads, writes, and tests; we call its evolution workspace optimization. Workspace optimization targets hard multi-turn environments where a frontier model has strong priors but cannot solve the task in a single shot, so the agent must learn through interaction. We propose a principled way to evolve the workspace, mirroring the structure of weight-space training: artifacts in place of parameters, evidence in place of data, counterexamples in place of losses, and textual feedback in place of gradients. We instantiate the idea in DreamTeam, a multi-agent harness for ARC-AGI-3 whose roles build an executable world model, plan, hypothesize, probe, strategize, and route failures. On the current 25-game ARC-AGI-3 public set under the official scoring protocol and averaged over two independent runs, DreamTeam improves the SOTA protocol-matched agent's score from 36% to 38.4%, while using 31% fewer environment actions per game.

cs.AI↗

arXiv2026

Retrieval from Within: An Intrinsic Capability of Attention-Based Models

Retrieval-augmented generation (RAG) typically treats retrieval and generation as separate systems. We ask whether an attention-based encoder-decoder can instead retrieve directly from its own internal representations. We introduce INTRA (INTrinsic Retrieval via Attention), a framework where decoder attention queries score pre-encoded evidence chunks that are then directly reused as context for generation. By construction, INTRA unifies retrieval and generation, eliminating the retriever-generator mismatch typical of RAG pipelines. This design also amortizes context encoding by reusing precomputed encoder states across queries. On question-answering benchmarks, INTRA outperforms strong engineered retrieval pipelines on both evidence recall and end-to-end answer quality. Our results demonstrate that attention-based models already possess a retrieval mechanism that can be elicited, rather than added as an external module.

cs.LG↗

arXiv2026

Normalized Architectures are Natively 4-Bit

Training large language models at 4-bit precision is critical for efficiency. We show that nGPT, an architecture that constrains weights and hidden representations to the unit hypersphere, is inherently more robust to low-precision arithmetic. This removes the need for interventions-such as applying random Hadamard transforms and performing per-tensor scaling calculations-to preserve model quality, and it enables stable end-to-end NVFP4 training. We validate this approach on both a 1.2B dense model and hybrid (Mamba-Transformer) MoE models of up to 3B/30B parameters. We trace this robustness to the dot product: while quantization noise remains largely uncorrelated in both standard and normalized architectures, the signal behaves differently. In nGPT, the hypersphere constraint enhances weak positive correlations among the element-wise products, leading to a constructive accumulation of the signal across the hidden dimension while the noise continues to average out. This yields a higher effective signal-to-noise ratio and a flatter loss landscape, with the effect strengthening as the hidden dimension grows, suggesting increasing advantages at scale. A reference implementation is available at https://github.com/anonymous452026/ngpt-nvfp4

cs.LG↗

arXiv2026

Optimal L2 Regularization in High-dimensional Continual Linear Regression

We study generalization in an overparameterized continual linear regression setting, where a model is trained with L2 (isotropic) regularization across a sequence of tasks. We derive a closed-form expression for the expected generalization loss in the high-dimensional regime that holds for arbitrary linear teachers. We demonstrate that isotropic regularization mitigates label noise under both single-teacher and multiple i.i.d. teacher settings, whereas prior work accommodating multiple teachers either did not employ regularization or used memory-demanding methods. Furthermore, we prove that the optimal fixed regularization strength scales nearly linearly with the number of tasks $T$, specifically as $T/\ln T$. To our knowledge, this is the first such result in theoretical continual learning. Finally, we validate our theoretical findings through experiments on linear regression and neural networks, illustrating how this scaling law affects generalization and offering a practical recipe for the design of continual learning systems.

cs.LG↗

arXiv2025

Block Sparse Flash Attention

Modern large language models increasingly require long contexts for reasoning and multi-document tasks, but attention's quadratic complexity creates a severe computational bottleneck. We present Block-Sparse FlashAttention (BSFA), a drop-in replacement that accelerates long-context inference while preserving model quality. Unlike methods that predict importance before computing scores, BSFA computes exact query-key similarities to select the top-k most important value blocks for each query. By comparing per-block maximum scores against calibrated thresholds, we skip approximately 50% of the computation and memory transfers for pruned blocks. Our training-free approach requires only a one-time threshold calibration on a small dataset to learn the per-layer and per-head attention score distributions. We provide a CUDA kernel implementation that can be used as a drop-in replacement for FlashAttention. On Llama-3.1-8B, BSFA achieves up to 1.10x speedup on real-world reasoning benchmarks and up to 1.24x for needle-in-a-haystack retrieval tasks while maintaining above 99% baseline accuracy, with certain configurations even improving accuracy by focusing on the most relevant content, substantially outperforming existing sparse attention methods. The implementation is available at https://github.com/Danielohayon/Block-Sparse-Flash-Attention

cs.LG↗

arXiv2025

Alias-Free ViT: Fractional Shift Invariance via Linear Attention

Transformers have emerged as a competitive alternative to convnets in vision tasks, yet they lack the architectural inductive bias of convnets, which may hinder their potential performance. Specifically, Vision Transformers (ViTs) are not translation-invariant and are more sensitive to minor image translations than standard convnets. Previous studies have shown, however, that convnets are also not perfectly shift-invariant, due to aliasing in downsampling and nonlinear layers. Consequently, anti-aliasing approaches have been proposed to certify convnets' translation robustness. Building on this line of work, we propose an Alias-Free ViT, which combines two main components. First, it uses alias-free downsampling and nonlinearities. Second, it uses linear cross-covariance attention that is shift-equivariant to both integer and fractional translations, enabling a shift-invariant global representation. Our model maintains competitive performance in image classification and outperforms similar-sized models in terms of robustness to adversarial translations.

cs.CV↗

arXiv2025

Are Greedy Task Orderings Better Than Random in Continual Linear Regression?

We analyze task orderings in continual learning for linear regression, assuming joint realizability of training data. We focus on orderings that greedily maximize dissimilarity between consecutive tasks, a concept briefly explored in prior work but still surrounded by open questions. Using tools from the Kaczmarz method literature, we formalize such orderings and develop geometric and algebraic intuitions around them. Empirically, we demonstrate that greedy orderings converge faster than random ones in terms of the average loss across tasks, both for linear regression with random data and for linear probing on CIFAR-100 classification tasks. Analytically, in a high-rank regression setting, we prove a loss bound for greedy orderings analogous to that of random ones. However, under general rank, we establish a repetition-dependent separation. Specifically, while prior work showed that for random orderings, with or without replacement, the average loss after $k$ iterations is bounded by $\mathcal{O}(1/\sqrt{k})$, we prove that single-pass greedy orderings may fail catastrophically, whereas those allowing repetition converge at rate $\mathcal{O}(1/\sqrt[3]{k})$. Overall, we reveal nuances within and between greedy and random orderings.

cs.LG↗

arXiv2025

Tensor-Parallelism with Partially Synchronized Activations

Training and inference of Large Language Models (LLMs) with tensor-parallelism requires substantial communication to synchronize activations. Our findings suggest that with a few minor adjustments to current practices, LLMs can be trained without fully synchronizing activations, reducing bandwidth demands. We name this "Communication-Aware Architecture for Tensor-parallelism" (CAAT-Net). We train a 7B parameter CAAT-Net model and show that tensor-parallel communication can be reduced by up to 50% with no significant drop in pretraining accuracy across nearly all evaluated benchmarks. We also experiment with smaller 130M and 1.1B models to show the robustness and scalability of our method. We find that, in some scenarios, validation loss can even improve when reducing communication. Finally, we demonstrate how CAAT-Net accelerates both training and inference workloads across various settings and model sizes.

cs.LG↗

arXiv2025

When Diffusion Models Memorize: Inductive Biases in Probability Flow of Minimum-Norm Shallow Neural Nets

While diffusion models generate high-quality images via probability flow, the theoretical understanding of this process remains incomplete. A key question is when probability flow converges to training samples or more general points on the data manifold. We analyze this by studying the probability flow of shallow ReLU neural network denoisers trained with minimal $\ell^2$ norm. For intuition, we introduce a simpler score flow and show that for orthogonal datasets, both flows follow similar trajectories, converging to a training point or a sum of training points. However, early stopping by the diffusion time scheduler allows probability flow to reach more general manifold points. This reflects the tendency of diffusion models to both memorize training samples and generate novel points that combine aspects of multiple samples, motivating our study of such behavior in simplified settings. We extend these results to obtuse simplex data and, through simulations in the orthogonal case, confirm that probability flow converges to a training point, a sum of training points, or a manifold point. Moreover, memorization decreases when the number of training samples grows, as fewer samples accumulate near training points.

stat.ML↗

arXiv2025

Optimal Rates in Continual Linear Regression via Increasing Regularization

We study realizable continual linear regression under random task orderings, a common setting for developing continual learning theory. In this setup, the worst-case expected loss after $k$ learning iterations admits a lower bound of $\Omega(1/k)$. However, prior work using an unregularized scheme has only established an upper bound of $O(1/k^{1/4})$, leaving a significant gap. Our paper proves that this gap can be narrowed, or even closed, using two frequently used regularization schemes: (1) explicit isotropic $\ell_2$ regularization, and (2) implicit regularization via finite step budgets. We show that these approaches, which are used in practice to mitigate forgetting, reduce to stochastic gradient descent (SGD) on carefully defined surrogate losses. Through this lens, we identify a fixed regularization strength that yields a near-optimal rate of $O(\log k / k)$. Moreover, formalizing and analyzing a generalized variant of SGD for time-varying functions, we derive an increasing regularization strength schedule that provably achieves an optimal rate of $O(1/k)$. This suggests that schedules that increase the regularization coefficient or decrease the number of steps per task are beneficial, at least in the worst case.

cs.LG↗

arXiv2025

Temperature is All You Need for Generalization in Langevin Dynamics and other Markov Processes

We analyze the generalization gap (gap between the training and test errors) when training a potentially over-parametrized model using a Markovian stochastic training algorithm, initialized from some distribution $\theta_0 \sim p_0$. We focus on Langevin dynamics with a positive temperature $\beta^{-1}$, i.e. gradient descent on a training loss $L$ with infinitesimal step size, perturbed with $\beta^{-1}$-variances Gaussian noise, and lightly regularized or bounded. There, we bound the generalization gap, at any time during training, by $\sqrt{(\beta\mathbb{E} L (\theta_0) + \log(1/\delta))/N}$ with probability $1-\delta$ over the dataset, where $N$ is the sample size, and $\mathbb{E} L (\theta_0) =O(1)$ with standard initialization scaling. In contrast to previous guarantees, we have no dependence on either training time or reliance on mixing, nor a dependence on dimensionality, gradient norms, or any other properties of the loss or model. This guarantee follows from a general analysis of any Markov process-based training that has a Gibbs-style stationary distribution. The proof is surprisingly simple, once we observe that the marginal distribution divergence from initialization remains bounded, as implied by a generalized second law of thermodynamics.

cs.LG↗

arXiv2025

FP4 All the Way: Fully Quantized Training of LLMs

We demonstrate, for the first time, fully quantized training (FQT) of large language models (LLMs) using predominantly 4-bit floating-point (FP4) precision for weights, activations, and gradients on datasets up to 200 billion tokens. We extensively investigate key design choices for FP4, including block sizes, scaling formats, and rounding methods. Our analysis shows that the NVFP4 format, where each block of 16 FP4 values (E2M1) shares a scale represented in E4M3, provides optimal results. We use stochastic rounding for backward and update passes and round-to-nearest for the forward pass to enhance stability. Additionally, we identify a theoretical and empirical threshold for effective quantized training: when the gradient norm falls below approximately $\sqrt{3}$ times the quantization noise, quantized training becomes less effective. Leveraging these insights, we successfully train a 7-billion-parameter model on 256 Intel Gaudi2 accelerators. The resulting FP4-trained model achieves downstream task performance comparable to a standard BF16 baseline, confirming that FP4 training is a practical and highly efficient approach for large-scale LLM training. A reference implementation is supplied in https://github.com/Anonymous1252022/fp4-all-the-way .

cs.LG↗

arXiv2025

PLUMAGE: Probabilistic Low rank Unbiased Min Variance Gradient Estimator for Efficient Large Model Training

Accelerator memory and networking constraints have emerged as dominant bottlenecks when training large language models LLMs with billions of parameters. Existing low rank gradient estimators such as GaLoRE and FLORA compress gradients and optimizer tensors by projecting weight gradients onto a rank r subspace, enabling LLM training on consumer hardware. Yet, these methods are either biased or subject to high estimator variance. Moreover, the optimizer state based on the first and second moments estimates expressed in the previous subspace becomes misaligned whenever the projection is updated, leading to instabilities during training. We propose PLUMAGE: Probabilistic Low rank Unbiased Minimum vAriance Gradient Estimator. PLUMAGE is a drop in replacement for existing low rank gradient estimators. It does not introduce new hyperparameters beyond the chosen rank r and the update interval. In addition, we resolve optimizer state misalignment issues to prevent spurious weight updates and enhance training stability. We empirically demonstrate that PLUMAGE shrinks the full rank optimization's gap over the pre training evaluation loss by 33% on average across models and the average training loss across the GLUE benchmark by 28% within a similar computational and memory footprint as GaloRE.

cs.LG↗

arXiv2025

From Continual Learning to SGD and Back: Better Rates for Continual Linear Models

We study the common continual learning setup where an overparameterized model is sequentially fitted to a set of jointly realizable tasks. We analyze forgetting, defined as the loss on previously seen tasks, after $k$ iterations. For continual linear models, we prove that fitting a task is equivalent to a single stochastic gradient descent (SGD) step on a modified objective. We develop novel last-iterate SGD upper bounds in the realizable least squares setup and leverage them to derive new results for continual learning. Focusing on random orderings over $T$ tasks, we establish universal forgetting rates, whereas existing rates depend on problem dimensionality or complexity and become prohibitive in highly overparameterized regimes. In continual regression with replacement, we improve the best existing rate from $O((d-\bar{r})/k)$ to $O(\min(1/\sqrt[4]{k}, \sqrt{(d-\bar{r})}/k, \sqrt{T\bar{r}}/k))$, where $d$ is the dimensionality and $\bar{r}$ the average task rank. Furthermore, we establish the first rate for random task orderings without replacement. The resulting rate $O(\min(1/\sqrt[4]{T},\, (d-\bar{r})/T))$ shows that randomization alone, without task repetition, prevents catastrophic forgetting in sufficiently long task sequences. Finally, we prove a matching $O(1/\sqrt[4]{k})$ forgetting rate for continual linear classification on separable data. Our universal rates extend to broader methods, such as block Kaczmarz and POCS, illuminating their loss convergence under i.i.d. and single-pass orderings.

cs.LG↗

arXiv2024

The Implicit Bias of Gradient Descent on Separable Multiclass Data

Implicit bias describes the phenomenon where optimization-based training algorithms, without explicit regularization, show a preference for simple estimators even when more complex estimators have equal objective values. Multiple works have developed the theory of implicit bias for binary classification under the assumption that the loss satisfies an exponential tail property. However, there is a noticeable gap in analysis for multiclass classification, with only a handful of results which themselves are restricted to the cross-entropy loss. In this work, we employ the framework of Permutation Equivariant and Relative Margin-based (PERM) losses [Wang and Scott, 2024] to introduce a multiclass extension of the exponential tail property. This class of losses includes not only cross-entropy but also other losses. Using this framework, we extend the implicit bias result of Soudry et al. [2018] to multiclass classification. Furthermore, our proof techniques closely mirror those of the binary case, thus illustrating the power of the PERM framework for bridging the binary-multiclass gap.

cs.LG↗

arXiv2024

Provable Tempered Overfitting of Minimal Nets and Typical Nets

We study the overfitting behavior of fully connected deep Neural Networks (NNs) with binary weights fitted to perfectly classify a noisy training set. We consider interpolation using both the smallest NN (having the minimal number of weights) and a random interpolating NN. For both learning rules, we prove overfitting is tempered. Our analysis rests on a new bound on the size of a threshold circuit consistent with a partial function. To the best of our knowledge, ours are the first theoretical results on benign or tempered overfitting that: (1) apply to deep NNs, and (2) do not require a very high or very low input dimension.

cs.LG↗

arXiv2024

Foldable SuperNets: Scalable Merging of Transformers with Different Initializations and Tasks

Recent methods aim to merge neural networks (NNs) with identical architectures trained on different tasks into a single multi-task model. While most works focus on the simpler setup of merging NNs initialized from a common pre-trained network, we target the harder problem of merging large transformers trained on different tasks from distinct initializations. We show that traditional merging methods fail catastrophically in this setup, while Knowledge Distillation (KD) achieves much better results, though at a higher cost. However, KD is data-inefficient, as it does not exploit the original models' weights. To solve this, we introduce "Foldable SuperNet Merge" (FS-Merge), which trains a SuperNet containing the original models (with frozen weights) using a feature reconstruction objective. After training, the SuperNet is folded back to the size of a single original model. FS-Merge is simple, data-efficient, has a computational cost comparable to KD, and is proven to have superior expressiveness compared to traditional merging methods on MLP models. It achieves SOTA results when tested on MLPs and transformers across various sizes, tasks, modalities, and distribution shifts, especially in low-data scenarios.

cs.LG↗