arXiv subjects

Ivan Oseledets

Publications and source records attributed to Ivan Oseledets.

At least 37 records · Page 2Linked to original sources

arXiv2026

IDLM: Inverse-distilled Diffusion Language Models

Diffusion Language Models (DLMs) have recently achieved strong results in text generation. However, their multi-step sampling leads to slow inference, limiting practical use. To address this, we extend Inverse Distillation, a technique originally developed to accelerate continuous diffusion models, to the discrete setting. Nonetheless, this extension introduces both theoretical and practical challenges. From a theoretical perspective, the inverse distillation objective lacks uniqueness guarantees, which may lead to suboptimal solutions. From a practical standpoint, backpropagation in the discrete space is non-trivial and often unstable. To overcome these challenges, we first provide a theoretical result demonstrating that our inverse formulation admits a unique solution, thereby ensuring valid optimization. We then introduce gradient-stable relaxations to support effective training. As a result, experiments on multiple DLMs show that our method, Inverse-distilled Diffusion Language Models (IDLM), reduces the number of inference steps by 4x-64x, while preserving the teacher model's generation quality. We provide the code, model checkpoints, and video tutorials on the project page: https://david-cripto.github.io/idlm-project-page

cs.LG↗

arXiv2026

OCC-RAG: Optimal Cognitive Core for Faithful Question Answering

Recent progress in the development of language models has been defined by scale, with each generation absorbing more of the world's knowledge into its weights. However, many practical applications benefit more from robust reasoning than from extensive parametric knowledge. In this setting, task-specialized small language models (SLMs) offer a principled design choice. We introduce Optimal Cognitive Core (OCC), a family of SLMs built around this premise. As a variant of OCC, we present OCC-RAG, optimized for faithful question answering (QA) grounded in the provided context. This task directly aligns with the OCC design approach, requiring multi-hop reasoning over supplied passages while ignoring memorized knowledge. To train OCC-RAG, we implement a novel pipeline for synthesizing multi-context, multi-hop QA data at scale, producing a corpus of over three million examples targeting multi-hop reasoning, strict context faithfulness, and calibrated abstention. We release OCC-RAG-0.6B and OCC-RAG-1.7B, both mid-trained on this corpus. The models produce structured reasoning traces with source citations grounded in literal quotes from the context. Through OCC-RAG, we demonstrate that compact, task-specialized SLMs can match or exceed general-purpose models 2 -- 6x their size across multi-hop reasoning (HotpotQA, MuSiQue, TAT-QA), faithfulness (ConFiQA), and refusal (MuSiQue-Un) benchmarks.

cs.CL↗

arXiv2026

Harnessing non-adversarial robustness in large language models

The work presents an approach for addressing the challenge of robustness in Large Language Models (LLMs) to alterations and potential errors caused by semantically similar but textually different prompts. Recent works have shown that these kinds of prompt variations can significantly impact the performance of LLMs on tasks. The central question is: can LLMs' robustness to semantically-neutral prompt alterations be acquired without expensive retraining of the entire model? We address this question both theoretically and through experiments. Our theoretical analysis reveals a crucial factor impacting model robustness - a systematic expected shift or perturbation-induced bias in neural network module outputs. Motivated by this analysis, we show that robustness can be achieved via a simple fine-tuning process: debiasing for robustness. We identify conditions when debiasing helps and when it does not, and demonstrate, through both theory and extensive experiments, that debiasing for robustness may indeed be a quick and efficient tool to enhance robustness and provide certification against random prompt perturbations.

cs.AI↗

arXiv2026

Sinc Kolmogorov-Arnold network and its application for solving PDEs with singularities

In this paper, we propose to use Sinc interpolation in the context of Kolmogorov-Arnold Networks, neural networks with learnable activation functions, which recently gained attention as alternatives to Multilayer Perceptron. Many different function representations have already been tried, but we show that Sinc interpolation proposes a viable alternative, since it is known in numerical analysis to effectively represent both smooth functions and functions with singularities. This is important not only for function approximation but also for solving the partial differential equations with physics-informed neural networks. Through a series of experiments, we show that SincKANs provide better results in almost all of the examples we have considered.

cs.LG↗

arXiv2026

Message-Passing GNNs Fail to Approximate Sparse Triangular Factorizations

Graph Neural Networks (GNNs) have been proposed as a tool for learning sparse matrix preconditioners, which are key components in accelerating linear solvers. We present theoretical and empirical evidence that message-passing GNNs are fundamentally incapable of approximating sparse triangular factorizations for classes of matrices for which high-quality preconditioners exist but require non-local dependencies. To illustrate this, we construct a set of baselines using both synthetic matrices and real-world examples from the SuiteSparse collection. Across a range of GNN architectures, including Graph Attention Networks and Graph Transformers, we observe low cosine similarity ($\leq0.7$ in key cases) between predicted and reference factors. Our theoretical and empirical results suggest that architectural innovations beyond message-passing are necessary for applying GNNs to scientific computing tasks such as matrix factorization. Moreover, experiments demonstrate that overcoming non-locality alone is insufficient. Tailored architectures are necessary to capture the required dependencies since even a completely non-local Global Graph Transformer fails to match the proposed baselines.

cs.LG↗

arXiv2026

TUBE: Tangent Upper Bound on Evidence for Discrete Diffusion Language Models

Log-likelihood is a standard metric for evaluating generative models. Unfortunately, in contrast to autoregressive models (ARMs), discrete diffusion models generally do not admit exact computation of this quantity. Existing evaluations, therefore, rely on the evidence lower bound (ELBO), leaving unclear how much higher the true value may be. We address this by introducing the Tangent Upper Bound on Evidence (TUBE), a variational upper bound on log-likelihood that admits an unbiased Monte Carlo estimator. Our TUBE extends across latent-variable models, including masked diffusion models (MDMs), any-order ARMs (AO-ARMs), and block variants of both. Applied to block MDMs and block AO-ARMs, TUBE reveals our key empirical finding that these models lie strictly below the exact ARM baseline, showing that ARMs still dominate in likelihood.

cs.LG↗

arXiv2026

Logit-KL Flow Matching: Non-Autoregressive Text Generation via Sampling-Hybrid Inference

Non-autoregressive (NAR) language models offer notable efficiency in text generation by circumventing the sequential bottleneck of autoregressive decoding. However, accurately modeling dependencies in discrete sequences remains challenging in this paradigm. In this work, we advance the field of NAR generation by applying conditional flow matching (CFM) methods grounded in geometrically principled interpolation, specifically leveraging Kullback-Leibler (KL) divergence geodesics, which correspond to linear interpolation in logit space. We rigorously establish that maximizing conditional likelihood in this setting precisely recovers the flow matching velocity field, supplying the theoretical justification for this approach in sequence modeling. To address practical performance gaps of basic inference, we propose a novel empirical sampling strategy that iteratively denoises and re-noises, along with a hybrid scheme that integrates our sampling method with basic procedure. Across unconditional and conditional text and code infilling, the approach improves perplexity and downstream metrics over prior NAR baselines under matched settings.

cs.CL↗

arXiv2026

Low-rank surrogate modeling and stochastic zero-order optimization for training of neural networks with black-box layers

The growing demand for energy-efficient, high-performance AI systems has led to increased attention on alternative computing platforms (e.g., photonic, neuromorphic) due to their potential to accelerate learning and inference. However, integrating such physical components into deep learning pipelines remains challenging, as physical devices often offer limited expressiveness, and their non-differentiable nature renders on-device backpropagation difficult or infeasible. This motivates the development of hybrid architectures that combine digital neural networks with reconfigurable physical layers, which effectively behave as black boxes. In this work, we present a framework for the end-to-end training of such hybrid networks. This framework integrates stochastic zeroth-order optimization for updating the physical layer's internal parameters with a dynamic low-rank surrogate model that enables gradient propagation through the physical layer. A key component of our approach is the implicit projector-splitting integrator algorithm, which updates the lightweight surrogate model after each forward pass with minimal hardware queries, thereby avoiding costly full matrix reconstruction. We demonstrate our method across diverse deep learning tasks, including: computer vision, audio classification, and language modeling. Notably, across all modalities, the proposed approach achieves near-digital baseline accuracy and consistently enables effective end-to-end training of hybrid models incorporating various non-differentiable physical components (spatial light modulators, microring resonators, and Mach-Zehnder interferometers). This work bridges hardware-aware deep learning and gradient-free optimization, thereby offering a practical pathway for integrating non-differentiable physical components into scalable, end-to-end trainable AI systems.

cs.LG↗

arXiv2026

Deep Learning for Subspace Regression

It is often possible to perform reduced order modelling by specifying linear subspace which accurately captures the dynamics of the system. This approach becomes especially appealing when linear subspace explicitly depends on parameters of the problem. A practical way to apply such a scheme is to compute subspaces for a selected set of parameters in the computationally demanding offline stage and in the online stage approximate subspace for unknown parameters by interpolation. For realistic problems the space of parameters is high dimensional, which renders classical interpolation strategies infeasible or unreliable. We propose to relax the interpolation problem to regression, introduce several loss functions suitable for subspace data, and use a neural network as an approximation to high-dimensional target function. To further simplify a learning problem we introduce redundancy: in place of predicting subspace of a given dimension we predict larger subspace. We show theoretically that this strategy decreases the complexity of the mapping for elliptic eigenproblems with constant coefficients and makes the mapping smoother for general smooth function on the Grassmann manifold. Empirical results also show that accuracy significantly improves when larger-than-required subspaces are predicted. With the set of numerical illustrations we demonstrate that subspace regression can be useful for a range of tasks including parametric eigenproblems, deflation techniques, relaxation methods, optimal control and solution of parametric partial differential equations.

cs.LG↗

arXiv2026

Listener-Rewarded Thinking in VLMs for Image Preferences

Training robust and generalizable reward models for human visual preferences is essential for aligning text-to-image and text-to-video generative models with human intent. However, current reward models often fail to generalize, and supervised fine-tuning leads to memorization, demanding complex annotation pipelines. While reinforcement learning (RL), specifically Group Relative Policy Optimization (GRPO), improves generalization, we uncover a key failure mode: a significant drop in reasoning accuracy occurs when a model's reasoning trace contradicts that of an independent, frozen vision-language model ("listener") evaluating the same output. To address this, we introduce a listener-augmented GRPO framework. Here, the listener re-evaluates the reasoner's chain-of-thought to provide a dense, calibrated confidence score, shaping the RL reward signal. This encourages the reasoner not only to answer correctly, but to produce explanations that are persuasive to an independent model. Our listener-shaped reward scheme achieves best accuracy on the ImageReward benchmark (67.4%), significantly improves out-of-distribution (OOD) performance on a large-scale human preference dataset (1.2M votes, up to +6% over naive reasoner), and reduces reasoning contradictions compared to strong GRPO and SFT baselines. These results demonstrate that listener-based rewards provide a scalable, data-efficient path to aligning vision-language models with nuanced human preferences. We will release our reasoning model here: https://huggingface.co/alexgambashidze/qwen2.5vl_image_preference_reasoner.

cs.CV↗

arXiv2026

Geological Field Restoration through the Lens of Image Inpainting

We study an ill-posed problem of geological field reconstruction under limited observations. Engineers often have to deal with the problem of reconstructing the subsurface geological field from sparse measurements such as exploration well data. Inspired by image inpainting, we model this partially observed spatial field as a multidimensional tensor and recover missing values by enforcing a global low rank structure together with spatial smoothness. We solve the resulting optimization via the Alternating Direction Method of Multipliers. On the SPE10 model 2 benchmark, this deterministic approach yields consistently lower relative squared error than ordinary kriging across various sampling densities and produces visually coherent reconstructions.

cs.CV↗

arXiv2026

OrthoFuse: Training-free Riemannian Fusion of Orthogonal Style-Concept Adapters for Diffusion Models

In a rapidly growing field of model training there is a constant practical interest in parameter-efficient fine-tuning and various techniques that use a small amount of training data to adapt the model to a narrow task. However, there is an open question: how to combine several adapters tuned for different tasks into one which is able to yield adequate results on both tasks? Specifically, merging subject and style adapters for generative models remains unresolved. In this paper we seek to show that in the case of orthogonal fine-tuning (OFT), we can use structured orthogonal parametrization and its geometric properties to get the formulas for training-free adapter merging. In particular, we derive the structure of the manifold formed by the recently proposed Group-and-Shuffle ($\mathcal{GS}$) orthogonal matrices, and obtain efficient formulas for the geodesics approximation between two points. Additionally, we propose a $\text{spectra restoration}$ transform that restores spectral properties of the merged adapter for higher-quality fusion. We conduct experiments in subject-driven generation tasks showing that our technique to merge two $\mathcal{GS}$ orthogonal matrices is capable of uniting concept and style features of different adapters. To the best of our knowledge, this is the first training-free method for merging multiplicative orthogonal adapters. Code is available via the $\href{https://github.com/ControlGenAI/OrthoFuse}{link}$.

cs.CV↗

arXiv2026

Marchuk: Efficient Global Weather Forecasting from Mid-Range to Sub-Seasonal Scales via Flow Matching

Accurate subseasonal weather forecasting remains a major challenge due to the inherently chaotic nature of the atmosphere, which limits the predictive skill of conventional models beyond the mid-range horizon (approximately 15 days). In this work, we present \textit{Marchuk}, a generative latent flow-matching model for global weather forecasting spanning mid-range to subseasonal timescales, with prediction horizons of up to 30 days. Marchuk conditions on current-day weather maps and autoregressively predicts subsequent days' weather maps within the learned latent space. We replace rotary positional encodings (RoPE) with trainable positional embeddings and extend the temporal context window, which together enhance the model's ability to represent and propagate long-range temporal dependencies during latent forecasting. Marchuk offers two key advantages: high computational efficiency and strong predictive performance. Despite its compact architecture of only 276 million parameters, the model achieves performance comparable to LaDCast, a substantially larger model with 1.6 billion parameters, while operating at significantly higher inference speeds. We open-source our inference code and model at: https://v-gen-ai.github.io/Marchuk/

cs.LG↗

arXiv2026

Speech-to-LaTeX: New Models and Datasets for Converting Spoken Equations and Sentences

Conversion of spoken mathematical expressions is a challenging task that involves transcribing speech into a strictly structured symbolic representation while addressing the ambiguity inherent in the pronunciation of equations. Although significant progress has been achieved in automatic speech recognition (ASR) and language models (LM), the problem of converting spoken mathematics into LaTeX remains underexplored. This task directly applies to educational and research domains, such as lecture transcription or note creation. Based on ASR post-correction, prior work requires 2 transcriptions, focuses only on isolated equations, has a limited test set, and provides neither training data nor multilingual coverage. To address these issues, we present the first fully open-source large-scale dataset, comprising over 66,000 human-annotated audio samples of mathematical equations and sentences in English and Russian, drawn from diverse scientific domains. In addition to the ASR post-correction models and few-shot prompting, we apply audio language models, demonstrating comparable character error rate (CER) results on the MathSpeech benchmark (28% vs. 30%) for the equations conversion. In contrast, on the proposed S2L-equations benchmark, our models outperform the MathSpeech model by a substantial margin of more than 36 percentage points, even after accounting for LaTeX formatting artifacts (27% vs. 64%). We establish the first benchmark for mathematical sentence recognition (S2L-sentences) and achieve an equation CER of 40%. This work lays the groundwork for future advances in multimodal AI, with a particular focus on mathematical content recognition.

cs.CV↗

arXiv2026

Astral: training physics-informed neural networks with error majorants

The primal approach to physics-informed learning is a residual minimization. We argue that residual is, at best, an indirect measure of the error of approximate solution and propose to train with error majorant instead. Since error majorant provides a direct upper bound on error, one can reliably estimate how close PiNN is to the exact solution and stop the optimization process when the desired accuracy is reached. We call loss function associated with error majorant \textbf{Astral}: neur\textbf{A}l a po\textbf{ST}erio\textbf{R}i function\textbf{A}l \textbf{L}oss. To compare Astral and residual loss functions, we illustrate how error majorants can be derived for various PDEs and conduct experiments with diffusion equations (including anisotropic and in the L-shaped domain), convection-diffusion equation, temporal discretization of Maxwell's equation, magnetostatics and nonlinear elastoplasticity problems. The results indicate that Astral loss is competitive to the residual loss, typically leading to faster convergence and lower error. The main benefit of using Astral loss comes from its ability to estimate error, which is impossible with other loss functions. Our experiments indicate that the error estimate obtained with Astral loss is usually tight enough, e.g., for a highly anisotropic equation, on average, Astral overestimates error by a factor of $1.5$, and for convection-diffusion by a factor of $1.7$. We further demonstrate that Astral loss is better correlated with error than residual and is a more reliable predictor of the error value. Moreover, unlike residual, the error indicator obtained from Astral loss has a superb spatial correlation with error. Backed with the empirical and theoretical results, we argue that one can productively use Astral loss to perform reliable error analysis and approximate PDE solutions with accuracy similar to standard residual-based techniques.

physics.comp-ph↗

arXiv2026

Spread them Apart: Towards Robust Watermarking of Generated Content

Generative models that can produce realistic images have improved significantly in recent years. The quality of the generated content has increased drastically, so sometimes it is very difficult to distinguish between the real images and the generated ones. Such an improvement comes at a price of ethical concerns about the usage of the generative models: the users of generative models can improperly claim ownership of the generated content protected by a license. In this paper, we propose an approach to embed watermarks into the generated content to allow future detection of the generated content and identification of the user who generated it. The watermark is embedded during the inference of the model, so the proposed approach does not require the retraining of the latter. We prove that watermarks embedded are guaranteed to be robust against additive perturbations of a bounded magnitude. We apply our method to watermark diffusion models and show that it matches state-of-the-art watermarking schemes in terms of robustness to different types of synthetic watermark removal attacks.

cs.CV↗

arXiv2026

Locally Subspace-Informed Neural Operators for Efficient Multiscale PDE Solving

Neural operators (NOs) struggle with high-contrast multiscale partial differential equations (PDEs), where fine-scale heterogeneities cause large errors. To address this, we use the Generalized Multiscale Finite Element Method (GMsFEM) that constructs localized spectral basis functions on coarse grids. This approach efficiently captures dominant multiscale features while solving heterogeneous PDEs accurately at reduced computational cost. However, computing these basis functions is computationally expensive. This gap motivates our core idea: to use a NO to learn the subspace itself - rather than individual basis functions - by employing a subspace-informed loss. On standard multiscale benchmarks - namely a linear elliptic diffusion problem and the nonlinear, steady-state Richards equation - our hybrid method cuts solution error by approximately $60\%$ compared with standalone NOs and reduces basis-construction time by about $60$ times relative to classical GMsFEM, while remaining independent of forcing terms and boundary conditions. The result fuses multiscale finite-element robustness with NO speed, yielding a practical solver for heterogeneous PDEs.

math.NA↗

arXiv2026

The Rogue Scalpel: Activation Steering Compromises LLM Safety

Activation steering is a promising technique for controlling LLM behavior by adding semantically meaningful vectors directly into a model's hidden states during inference. It is often framed as a precise, interpretable, and potentially safer alternative to fine-tuning. We demonstrate the opposite: steering systematically breaks model alignment safeguards, making it comply with harmful requests. Through extensive experiments on different model families, we show that even steering in a random direction can increase the probability of harmful compliance from 0% to 1-13%. Alarmingly, steering benign features from a sparse autoencoder (SAE), a common source of interpretable directions, demonstrates a comparable harmful potential. Finally, we show that combining 20 randomly sampled vectors that jailbreak a single prompt creates a universal attack, significantly increasing harmful compliance on unseen requests. These results challenge the paradigm of safety through interpretability, showing that precise control over model internals does not guarantee precise control over model behavior.

cs.LG↗