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.

Recent advancements have integrated Language Models (LMs) into a drug discovery pipeline. However, existing models mostly work with SMILES and SELFIES chemical string representations, which lack spatial features vital for drug discovery. Additionally, attempts to translate chemical 3D structures into text format encounter issues such as excessive length and insufficient atom connectivity information. To address these issues, we introduce nach0-pc, a model combining domain-specific encoder and textual representation to handle spatial arrangement of atoms effectively. Our approach utilizes a molecular point cloud encoder for concise and order-invariant structure representation. We introduce a novel pre-training scheme for molecular point clouds to distillate the knowledge from spatial molecular structures datasets. After fine-tuning within both single-task and multi-task frameworks, nach0-pc demonstrates performance comparable with other diffusion models in terms of generated samples quality across several established spatial molecular generation tasks. Notably, our model is a multi-task approach, in contrast to diffusion models being limited to single tasks. Additionally, it is capable of processing point cloud-related data, which language models are not capable of handling due to memory limitations. These lead to our model having reduced training and inference time while maintaining on par performance.

Text-to-SQL enables users to interact with databases through natural language, simplifying the retrieval and synthesis of information. Despite the success of large language models (LLMs) in converting natural language questions into SQL queries, their broader adoption is limited by two main challenges: achieving robust generalization across diverse queries and ensuring interpretative confidence in their predictions. To tackle these issues, our research investigates the integration of selective classifiers into Text-to-SQL systems. We analyse the trade-off between coverage and risk using entropy based confidence estimation with selective classifiers and assess its impact on the overall performance of Text-to-SQL models. Additionally, we explore the models' initial calibration and improve it with calibration techniques for better model alignment between confidence and accuracy. Our experimental results show that encoder-decoder T5 is better calibrated than in-context-learning GPT 4 and decoder-only Llama 3, thus the designated external entropy-based selective classifier has better performance. The study also reveal that, in terms of error detection, selective classifier with a higher probability detects errors associated with irrelevant questions rather than incorrect query generations.

Alexandr Nesterov, Andrey Sakhovskiy, Ivan Sviridov, Airat Valiev, Vladimir Makharev, Petr Anokhin, Galina Zubkova, Elena Tutubalina. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Daniil Moskovskiy, Nikita Sushko, Sergey Pletenev, Elena Tutubalina, Alexander Panchenko. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

In recent years, there has been substantial progress in using pretrained Language Models (LMs) on a range of tasks aimed at improving the understanding of biomedical texts. Nonetheless, existing biomedical LLMs show limited comprehension of complex, domain-specific concept structures and the factual information encoded in biomedical Knowledge Graphs (KGs). In this work, we propose BALI (Biomedical Knowledge Graph and Language Model Ali gnment), a novel joint LM and KG pre-training method that augments an LM with external knowledge by the simultaneous learning of a dedicated KG encoder and aligning the representations of both the LM and the graph. For a given textual sequence, we link biomedical concept mentions to the Unified Medical Language System (UMLS) KG and utilize local KG subgraphs as cross-modal positive samples for these mentions. Our empirical findings indicate that implementing our method on several leading biomedical LMs, such as PubMedBERT and BioLinkBERT, improves their performance on a range of language understanding tasks and the quality of entity representations, even with minimal pre-training on a small alignment dataset sourced from PubMed scientific abstracts.

We present a new manually annotated dataset of PubMed abstracts for concept normalization in Russian. It contains over 23,641 entity mentions in 756 documents linked to 4,544 unique concepts from the UMLS ontology. Compared to existing corpora, we explore two novel annotation characteristics: the nestedness of named entities and the incompleteness of the Russian medical terminology in UMLS. 4,424 entity mentions are linked to 1,535 unique English concepts absent in the Russian part of the UMLS ontology. We present several baselines for normalization over nested named entities obtained with state-of-the-art models such as SapBERT. Our experimental results show that models pre-trained on graph structural data from UMLS achieve superior performance in a zero-shot setting on bilingual terminology.

Molecular conformation optimization is crucial to computer-aided drug discovery and materials design. Traditional energy minimization techniques rely on iterative optimization methods that use molecular forces calculated by a physical simulator (oracle) as anti-gradients. However, this is a computationally expensive approach that requires many interactions with a physical simulator. One way to accelerate this procedure is to replace the physical simulator with a neural network. Despite recent progress in neural networks for molecular conformation energy prediction, such models are prone to errors due to distribution shift, leading to inaccurate energy minimization. We find that the quality of energy minimization with neural networks can be improved by providing optimization trajectories as additional training data. Still, obtaining complete optimization trajectories demands a lot of additional computations. To reduce the required additional data, we present the Gradual Optimization Learning Framework (GOLF) for energy minimization with neural networks. The framework consists of an efficient data-collecting scheme and an external optimizer. The external optimizer utilizes gradients from the energy prediction model to generate optimization trajectories, and the data-collecting scheme selects additional training data to be processed by the physical simulator. Our results demonstrate that the neural network trained with GOLF performs \textit{on par} with the oracle on a benchmark of diverse drug-like molecules using significantly less additional data.

Methods of computational quantum chemistry provide accurate approximations of molecular properties crucial for computer-aided drug discovery and other areas of chemical science. However, high computational complexity limits the scalability of their applications.Neural network potentials (NNPs) are a promising alternative to quantum chemistry methods, but they require large and diverse datasets for training.This work presents a new dataset and benchmark called $\nabla^2$DFT that is based on the nablaDFT.It contains twice as much molecular structures, three times more conformations, new data types and tasks, and state-of-the-art models.The dataset includes energies, forces, 17 molecular properties, Hamiltonian and overlap matrices, and a wavefunction object.All calculations were performed at the DFT level ($\omega$B97X-D/def2-SVP) for each conformation. Moreover, $\nabla^2$DFT is the first dataset that contains relaxation trajectories for a substantial number of drug-like molecules. We also introduce a novel benchmark for evaluating NNPs in molecular property prediction, Hamiltonian prediction, and conformational optimization tasks. Finally, we propose an extendable framework for training NNPs and implement 10 models within it.

Mark Rofin, Vladislav Mikhailov, Mikhail Florinsky, Andrey Kravchenko, Tatiana Shavrina, Elena Tutubalina, Daniel Karabekyan, Ekaterina Artemova. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. 2023.

MOTIVATION: This article describes NEREL-BIO-an annotation scheme and corpus of PubMed abstracts in Russian and smaller number of abstracts in English. NEREL-BIO extends the general domain dataset NEREL by introducing domain-specific entity types. NEREL-BIO annotation scheme covers both general and biomedical domains making it suitable for domain transfer experiments. NEREL-BIO provides annotation for nested named entities as an extension of the scheme employed for NEREL. Nested named entities may cross entity boundaries to connect to shorter entities nested within longer entities, making them harder to detect. RESULTS: NEREL-BIO contains annotations for 700+ Russian and 100+ English abstracts. All English PubMed annotations have corresponding Russian counterparts. Thus, NEREL-BIO comprises the following specific features: annotation of nested named entities, it can be used as a benchmark for cross-domain (NEREL → NEREL-BIO) and cross-language (English → Russian) transfer. We experiment with both transformer-based sequence models and machine reading comprehension models and report their results. AVAILABILITY AND IMPLEMENTATION: The dataset and annotation guidelines are freely available at https://github.com/nerel-ds/NEREL-BIO.

State of the art neural methods for open information extraction (OpenIE) usually extract triplets (or tuples) iteratively in an autoregressive or predicate-based manner in order not to produce duplicates. In this work, we propose a different approach to the problem that can be equally or more successful. Namely, we present a novel single-pass method for OpenIE inspired by object detection algorithms from computer vision. We use an order-agnostic loss based on bipartite matching that forces unique predictions and a Transformer-based encoder-only architecture for sequence labeling. The proposed approach is faster and shows superior or similar performance in comparison with state of the art models on standard benchmarks in terms of both quality metrics and inference time. Our model sets the new state of the art performance of 67.7% F1 on CaRB evaluated as OIE2016 while being 3.35x faster at inference than previous state of the art. We also evaluate the multilingual version of our model in the zero-shot setting for two languages and introduce a strategy for generating synthetic multilingual data to fine-tune the model for each specific language. In this setting, we show performance improvement of 15% on multilingual Re-OIE2016, reaching 75% F1 for both Portuguese and Spanish languages. Code and models are available at https://github.com/sberbank-ai/DetIE.

Medical data annotation requires highly qualified expertise. Despite the efforts devoted to medical entity linking in different languages, available data is very sparse in terms of both data volume and languages. In this work, we establish benchmarks for cross-lingual medical entity linking using clinical reports, clinical guidelines, and medical research papers. We present a test set filtering procedure designed to analyze the “hard cases” of entity linking approaching zero-shot cross-lingual transfer learning, evaluate state-of-the-art models, and draw several interesting conclusions based on our evaluation results.

In this paper, we describe entity linking annotation over nested named entities in the recently released Russian NEREL dataset for information extraction. The NEREL collection is currently the largest Russian dataset annotated with entities and relations. It includes 933 news texts with annotation of 29 entity types and 49 relation types. The paper describes the main design principles behind NEREL’s entity linking annotation, provides its statistics, and reports evaluation results for several entity linking baselines. To date, 38,152 entity mentions in 933 documents are linked to Wikidata. The NEREL dataset is publicly available.

MOTIVATION: Clinical trials are the essential stage of every drug development program for the treatment to become available to patients. Despite the importance of well-structured clinical trial databases and their tremendous value for drug discovery and development such instances are very rare. Presently large-scale information on clinical trials is stored in clinical trial registers which are relatively structured, but the mappings to external databases of drugs and diseases are increasingly lacking. The precise production of such links would enable us to interrogate richer harmonized datasets for invaluable insights. RESULTS: We present a neural approach for medical concept normalization of diseases and drugs. Our two-stage approach is based on Bidirectional Encoder Representations from Transformers (BERT). In the training stage, we optimize the relative similarity of mentions and concept names from a terminology via triplet loss. In the inference stage, we obtain the closest concept name representation in a common embedding space to a given mention representation. We performed a set of experiments on a dataset of abstracts and a real-world dataset of trial records with interventions and conditions mapped to drug and disease terminologies. The latter includes mentions associated with one or more concepts (in-KB) or zero (out-of-KB, nil prediction). Experiments show that our approach significantly outperforms baseline and state-of-the-art architectures. Moreover, we demonstrate that our approach is effective in knowledge transfer from the scientific literature to clinical trial data. AVAILABILITY AND IMPLEMENTATION: We make code and data freely available at https://github.com/insilicomedicine/DILBERT.

MOTIVATION: Drugs and diseases play a central role in many areas of biomedical research and healthcare. Aggregating knowledge about these entities across a broader range of domains and languages is critical for information extraction (IE) applications. To facilitate text mining methods for analysis and comparison of patient's health conditions and adverse drug reactions reported on the Internet with traditional sources such as drug labels, we present a new corpus of Russian language health reviews. RESULTS: The Russian Drug Reaction Corpus (RuDReC) is a new partially annotated corpus of consumer reviews in Russian about pharmaceutical products for the detection of health-related named entities and the effectiveness of pharmaceutical products. The corpus itself consists of two parts, the raw one and the labeled one. The raw part includes 1.4 million health-related user-generated texts collected from various Internet sources, including social media. The labeled part contains 500 consumer reviews about drug therapy with drug- and disease-related information. Labels for sentences include health-related issues or their absence. The sentences with one are additionally labeled at the expression level for identification of fine-grained subtypes such as drug classes and drug forms, drug indications and drug reactions. Further, we present a baseline model for named entity recognition (NER) and multilabel sentence classification tasks on this corpus. The macro F1 score of 74.85% in the NER task was achieved by our RuDR-BERT model. For the sentence classification task, our model achieves the macro F1 score of 68.82% gaining 7.47% over the score of BERT model trained on Russian data. AVAILABILITY AND IMPLEMENTATION: We make the RuDReC corpus and pretrained weights of domain-specific BERT models freely available at https://github.com/cimm-kzn/RuDReC. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

OBJECTIVE: Research on pharmacovigilance from social media data has focused on mining adverse drug events (ADEs) using annotated datasets, with publications generally focusing on 1 of 3 tasks: ADE classification, named entity recognition for identifying the span of ADE mentions, and ADE mention normalization to standardized terminologies. While the common goal of such systems is to detect ADE signals that can be used to inform public policy, it has been impeded largely by limited end-to-end solutions for large-scale analysis of social media reports for different drugs. MATERIALS AND METHODS: We present a dataset for training and evaluation of ADE pipelines where the ADE distribution is closer to the average 'natural balance' with ADEs present in about 7% of the tweets. The deep learning architecture involves an ADE extraction pipeline with individual components for all 3 tasks. RESULTS: The system presented achieved state-of-the-art performance on comparable datasets and scored a classification performance of F1 = 0.63, span extraction performance of F1 = 0.44 and an end-to-end entity resolution performance of F1 = 0.34 on the presented dataset. DISCUSSION: The performance of the models continues to highlight multiple challenges when deploying pharmacovigilance systems that use social media data. We discuss the implications of such models in the downstream tasks of signal detection and suggest future enhancements. CONCLUSION: Mining ADEs from Twitter posts using a pipeline architecture requires the different components to be trained and tuned based on input data imbalance in order to ensure optimal performance on the end-to-end resolution task.

Understanding image advertisements is a challenging task, often requiring non-literal interpretation.We argue that standard image-based predictions are insufficient for symbolism prediction.Following the intuition that texts and images are complementary in advertising, we introduce a multimodal ensemble of a state of the art image-based classifier, a classifier based on an object detection architecture, and a fine-tuned language model applied to texts extracted from ads by OCR.The resulting system establishes a new state of the art in symbolism prediction.

Linking of biomedical entity mentions to various terminologies of chemicals, diseases, genes, adverse drug reactions is a challenging task, often requiring non-syntactic interpretation. A large number of biomedical corpora and state-of-the-art models have been introduced in the past five years. However, there are no general guidelines regarding the evaluation of models on these corpora in single- and cross-terminology settings. In this work, we perform a comparative evaluation of various benchmarks and study the efficiency of state-of-the-art neural architectures based on Bidirectional Encoder Representations from Transformers (BERT) for linking of three entity types across three domains: research abstracts, drug labels, and user-generated texts on drug therapy in English. We have made the source code and results available at https://github.com/insilicomedicine/Fair-Evaluation-BERT.

This article explores the problems of assigning documents to a limited number of topics and automating the process of topic structuring of Russian educational texts. For this purpose, we compiled an original corpus of school textbooks on Social Science. We utilized the Latent Dirichlet Allocation model for selection and comparative analysis of topics in the textbooks of different grades. This approach allows the reconstruction of the matrix of topics for each textbook in the orpus. The research demonstrated a grade ranked character of the topics in the text collection under study, in particular, there is a higher cohesion of topics in high school. The research also offers an innovative methodology of quantitative describing topics dynamics in the textbook collection. It allows visualization and comparison of strategies for presenting educational topics by different authors. The results received can be beneficial for both textbook writers as well as teachers and schoolchildren.

Recent research has shown the advantages of using autoencoders based on deep neural networks for collaborative filtering. In particular, the recently proposed Mult-VAE model, which used the multinomial likelihood variational autoencoders, has shown excellent results for top-N recommendations. In this work, we propose the Recommender VAE (RecVAE) model that originates from our research on regularization techniques for variational autoencoders. RecVAE introduces several novel ideas to improve Mult-VAE, including a novel composite prior distribution for the latent codes, a new approach to setting the beta hyperparameter for the beta-VAE framework, and a new approach to training based on alternating updates. In experimental evaluation, we show that RecVAE significantly outperforms previously proposed autoencoder-based models, including Mult-VAE and RaCT, across classical collaborative filtering datasets, and present a detailed ablation study to assess our new developments. Code and models are available at https://github.com/ilya-shenbin/RecVAE.

The rapidly growing availability of user reviews has become an important resource for companies to detect customer dissatisfaction from textual opinions. Much research in opinion mining focuses on extracting customers' opinions from products' reviews and predicting their sentiment orientation or ratings with the aim of helping other users to make a decision on whether to buy a product. However, there have been few recent studies conducted on business-related opinion tasks to extract more refined opinions about a product's quality problems or technical failures. The focus of this study is the extraction of problem phrases, mentioned in user reviews about products. We explore main opinion mining tasks to determine whether given text from reviews contains a mention of a problem. We formulate research questions and propose knowledge-based methods and probabilistic models to classify users' phrases and extract latent problem indicators, aspects and related sentiments from online reviews.