Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs.

Nikita Afonin, Nikita Andriianov, Vahagn Hovhannisyan, Nikhil Bageshpura, Kyle Liu, Kevin Zhu, Sunishchal Dev, Ashwinee Panda, Oleg Rogov, Elena Tutubalina, Alexander Panchenko, Mikhail Seleznyov. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Speaker recognition technology is applied to various tasks, from personal virtual assistants to secure access systems. However, the robustness of these systems against adversarial attacks, particularly to additive perturbations, remains a significant challenge. In this paper, we pioneer applying robustness certification techniques to speaker recognition, initially developed for the image domain. Our work covers this gap by transferring and improving randomized smoothing certification techniques against norm-bounded additive perturbations for classification and few-shot learning tasks to speaker recognition. We demonstrate the effectiveness of these methods on VoxCeleb 1 and 2 datasets for several models. We expect this work to improve the robustness of voice biometrics and accelerate the research of certification methods in the audio domain.

While Deep Neural Networks demonstrate remarkable performance in practical tasks, they are vulnerable to membership inference attacks aimed at identifying whether a certain object belongs to the training dataset. To conduct a membership inference attack on a target model, an adversary has to train a set of shadow models and conduct a statistical test to determine the membership status of the particular input object. Usually, shadow models are trained without taking into account the target model; we argue that utilizing the predictions of the target model can guide the training process of the shadow model. To improve the efficiency of shadow model-based membership inference attacks, we propose GLiRA, a knowledge distillation-guided approach to membership inference attacks. We observe that the knowledge distillation significantly improves the efficiency of a likelihood ratio membership inference attack when the architecture of the target model is both known and unknown to an attacker. We evaluate the proposed method across multiple image classification datasets and model architectures and demonstrate that knowledge distillation-guided likelihood ratio attack outperforms the current state-of-the-art membership inference attacks in the majority of experimental settings.

Image corruption by motion artifacts is an ingrained problem in Magnetic Resonance Imaging (MRI). In this work, we propose a neural network-based regularization term to enhance Autofocusing, a classic optimization-based method to remove motion artifacts. The method takes the best of both worlds: the optimization-based routine iteratively executes the blind demotion and deep learning-based prior penalizes for unrealistic restorations and speeds up the convergence. We validate the method on three models of motion trajectories, using synthetic and real noisy data. The method proves resilient to noise and anatomic structure variation, outperforming the state-of-the-art demotion methods.

Peripheral Difficult Venous Access (PDVA) is a commonplace problem in clinical practice which results in repetitive punctures, damaged veins, and significant discomfort to the patients. Nowadays, the poor visibility of subcutaneous vasculature in the visible part of the light spectrum is overcome by near-infrared (NIR) imaging and a returned projection of the recognized vasculature back to the arm of the patient. We introduce the first “smart” engine to govern the components of such imagers in a mixed reality setting. Namely, a closed-loop hardware system that optimizes cross-talk between the virtual mask generated from the NIR measurement and the projected augmenting image is proposed. Such real-virtual image translation is accomplished in several steps. First, the NIR vein segmentation task is solved using U-Net-based network architecture and the Frangi vesselness filter. The generated mask is then transformed and translated into the visible domain by a projector that adjusts for distortions and misalignment with the true vasculature using the paradigm of Reinforcement Learning (RL). We propose a new class of mixed reality reward functions that guarantees proper alignment of the projected image regardless of angle, translation, and scale offsets between the NIR measurement and the visible projection.