On the Spatial Structure of Mixture-of-Experts in Transformers

A common assumption is that MoE routers primarily leverage semantic features for expert selection. However, our study challenges this notion by demonstrating that positional token information also plays a crucial role in routing decisions. Through extensive empirical analysis, we provide evidence supporting this hypothesis, develop a phenomenological explanation of the observed behavior, and discuss practical implications for MoE-based architectures.

Test-Time Reasoning Through Visual Human Preferences with VLMs and Soft Rewards

Can Visual Language Models (VLMs) effectively capture human visual preferences? This work addresses this question by training VLMs to think about preferences at test time, employing reinforcement learning methods inspired by DeepSeek R1 and OpenAI O1. Using datasets such as ImageReward and Human Preference Score v2 (HPSv2), our models achieve accuracies of 64.9% on the ImageReward test set (trained on ImageReward official split) and 65.4% on HPSv2 (trained on approximately 25% of its data). These results match traditional encoder-based models while providing transparent reasoning and enhanced generalization. This approach allows to use not only rich VLM world knowledge, but also its potential to think, yielding interpretable outcomes that help decision-making processes. By demonstrating that human visual preferences reasonable by current VLMs, we introduce efficient soft-reward strategies for image ranking, outperforming simplistic selection or scoring methods. This reasoning capability enables VLMs to rank arbitrary images-regardless of aspect ratio or complexity-thereby potentially amplifying the effectiveness of visual Preference Optimization. By reducing the need for extensive markup while improving reward generalization and explainability, our findings can be a strong mile-stone that will enhance text-to-vision models even further.

I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders

Large Language Models (LLMs) have achieved remarkable success in natural language processing. Recent advances have led to the developing of a new class of reasoning LLMs; for example, open-source DeepSeek-R1 has achieved state-of-the-art performance by integrating deep thinking and complex reasoning. Despite these impressive capabilities, the internal reasoning mechanisms of such models remain unexplored. In this work, we employ Sparse Autoencoders (SAEs), a method to learn a sparse decomposition of latent representations of a neural network into interpretable features, to identify features that drive reasoning in the DeepSeek-R1 series of models. First, we propose an approach to extract candidate ''reasoning features'' from SAE representations. We validate these features through empirical analysis and interpretability methods, demonstrating their direct correlation with the model's reasoning abilities. Crucially, we demonstrate that steering these features systematically enhances reasoning performance, offering the first mechanistic account of reasoning in LLMs. Code available at https://github.com/AIRI-Institute/SAE-Reasoning

Exploring specialization and sensitivity of convolutional neural networks in the context of simultaneous image augmentations

Drawing parallels with the way biological networks are studied, we adapt the treatment--control paradigm to explainable artificial intelligence research and enrich it through multi-parametric input alterations. In this study, we propose a framework for investigating the internal inference impacted by input data augmentations. The internal changes in network operation are reflected in activation changes measured by variance, which can be decomposed into components related to each augmentation, employing Sobol indices and Shapley values. These quantities enable one to visualize sensitivity to different variables and use them for guided masking of activations. In addition, we introduce a way of single-class sensitivity analysis where the candidates are filtered according to their matching to prediction bias generated by targeted damaging of the activations. Relying on the observed parallels, we assume that the developed framework can potentially be transferred to studying biological neural networks in complex environments.

Combining Flow Matching and Transformers for Efficient Solution of Bayesian Inverse Problems

Solving Bayesian inverse problems efficiently remains a significant challenge due to the complexity of posterior distributions and the computational cost of traditional sampling methods. Given a series of observations and the forward model, we want to recover the distribution of the parameters, conditioned on observed experimental data. We show, that combining Conditional Flow Mathching (CFM) with transformer-based architecture, we can efficiently sample from such kind of distribution, conditioned on variable number of observations.

M3DA: Benchmark for Unsupervised Domain Adaptation in 3D Medical Image Segmentation

Domain shift presents a significant challenge in applying Deep Learning to the segmentation of 3D medical images from sources like Magnetic Resonance Imaging (MRI) and Computed Tomography (CT). Although numerous Domain Adaptation methods have been developed to address this issue, they are often evaluated under impractical data shift scenarios. Specifically, the medical imaging datasets used are often either private, too small for robust training and evaluation, or limited to single or synthetic tasks. To overcome these limitations, we introduce a M3DA /"mEd@/ benchmark comprising four publicly available, multiclass segmentation datasets. We have designed eight domain pairs featuring diverse and practically relevant distribution shifts. These include inter-modality shifts between MRI and CT and intra-modality shifts among various MRI acquisition parameters, different CT radiation doses, and presence/absence of contrast enhancement in images. Within the proposed benchmark, we evaluate more than ten existing domain adaptation methods. Our results show that none of them can consistently close the performance gap between the domains. For instance, the most effective method reduces the performance gap by about 62% across the tasks. This highlights the need for developing novel domain adaptation algorithms to enhance the robustness and scalability of deep learning models in medical imaging. We made our M3DA benchmark publicly available: https://github.com/BorisShirokikh/M3DA.

LLM-Microscope: Uncovering the Hidden Role of Punctuation in Context Memory of Transformers

We introduce methods to quantify how Large Language Models (LLMs) encode and store contextual information, revealing that tokens often seen as minor (e.g., determiners, punctuation) carry surprisingly high context. Notably, removing these tokens -- especially stopwords, articles, and commas -- consistently degrades performance on MMLU and BABILong-4k, even if removing only irrelevant tokens. Our analysis also shows a strong correlation between contextualization and linearity, where linearity measures how closely the transformation from one layer's embeddings to the next can be approximated by a single linear mapping. These findings underscore the hidden importance of filler tokens in maintaining context. For further exploration, we present LLM-Microscope, an open-source toolkit that assesses token-level nonlinearity, evaluates contextual memory, visualizes intermediate layer contributions (via an adapted Logit Lens), and measures the intrinsic dimensionality of representations. This toolkit illuminates how seemingly trivial tokens can be critical for long-range understanding.

FLAME: Flexible LLM-Assisted Moderation Engine

The rapid advancement of Large Language Models (LLMs) has introduced significant challenges in moderating user-model interactions. While LLMs demonstrate remarkable capabilities, they remain vulnerable to adversarial attacks, particularly ``jailbreaking'' techniques that bypass content safety measures. Current content moderation systems, which primarily rely on input prompt filtering, have proven insufficient, with techniques like Best-of-N (BoN) jailbreaking achieving success rates of 80% or more against popular LLMs. In this paper, we introduce Flexible LLM-Assisted Moderation Engine (FLAME): a new approach that shifts the focus from input filtering to output moderation. Unlike traditional circuit-breaking methods that analyze user queries, FLAME evaluates model responses, offering several key advantages: (1) computational efficiency in both training and inference, (2) enhanced resistance to BoN jailbreaking attacks, and (3) flexibility in defining and updating safety criteria through customizable topic filtering. Our experiments demonstrate that FLAME significantly outperforms current moderation systems. For example, FLAME reduces attack success rate in GPT-4o-mini and DeepSeek-v3 by a factor of ~9, while maintaining low computational overhead. We provide comprehensive evaluation on various LLMs and analyze the engine's efficiency against the state-of-the-art jailbreaking. This work contributes to the development of more robust and adaptable content moderation systems for LLMs.

Screener: Self-supervised Pathology Segmentation Model for 3D Medical Images

Accurate segmentation of all pathological findings in 3D medical images remains a significant challenge, as supervised models are limited to detecting only the few pathology classes annotated in existing datasets. To address this, we frame pathology segmentation as an unsupervised visual anomaly segmentation (UVAS) problem, leveraging the inherent rarity of pathological patterns compared to healthy ones. We enhance the existing density-based UVAS framework with two key innovations: (1) dense self-supervised learning (SSL) for feature extraction, eliminating the need for supervised pre-training, and (2) learned, masking-invariant dense features as conditioning variables, replacing hand-crafted positional encodings. Trained on over 30,000 unlabeled 3D CT volumes, our model, Screener, outperforms existing UVAS methods on four large-scale test datasets comprising 1,820 scans with diverse pathologies. Code and pre-trained models will be made publicly available.

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.