Elena Tutubalina

Natalia Loukachevitch, Suresh Manandhar, Elina Baral et al.

This paper 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. 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 (MRC) models and report their results. The dataset is freely available at https://github.com/nerel-ds/NEREL-BIO.

Michael Vasilkovsky, Anton Alekseev, Valentin Malykh et al.

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 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.

Andrey Sakhovskiy, Elena Tutubalina

In this paper, we focus on the classification of tweets as sources of potential signals for adverse drug effects (ADEs) or drug reactions (ADRs). Following the intuition that text and drug structure representations are complementary, we introduce a multimodal model with two components. These components are state-of-the-art BERT-based models for language understanding and molecular property prediction. Experiments were carried out on multilingual benchmarks of the Social Media Mining for Health Research and Applications (#SMM4H) initiative. Our models obtained state-of-the-art results of 0.61 F1 and 0.57 F1 on #SMM4H 2021 Shared Tasks 1a and 2 in English and Russian, respectively. On the classification of French tweets from SMM4H 2020 Task 1, our approach pushes the state of the art by an absolute gain of 8% F1. Our experiments show that the molecular information obtained from neural networks is more beneficial for ADE classification than traditional molecular descriptors. The source code for our models is freely available at https://github.com/Andoree/smm4h_2021_classification.

Tatiana Shamardina, Vladislav Mikhailov, Daniil Chernianskii et al.

We present the shared task on artificial text detection in Russian, which is organized as a part of the Dialogue Evaluation initiative, held in 2022. The shared task dataset includes texts from 14 text generators, i.e., one human writer and 13 text generative models fine-tuned for one or more of the following generation tasks: machine translation, paraphrase generation, text summarization, text simplification. We also consider back-translation and zero-shot generation approaches. The human-written texts are collected from publicly available resources across multiple domains. The shared task consists of two sub-tasks: (i) to determine if a given text is automatically generated or written by a human; (ii) to identify the author of a given text. The first task is framed as a binary classification problem. The second task is a multi-class classification problem. We provide count-based and BERT-based baselines, along with the human evaluation on the first sub-task. A total of 30 and 8 systems have been submitted to the binary and multi-class sub-tasks, correspondingly. Most teams outperform the baselines by a wide margin. We publicly release our codebase, human evaluation results, and other materials in our GitHub repository (https://github.com/dialogue-evaluation/RuATD).

Ekaterina Artemova, Maxim Zmeev, Natalia Loukachevitch et al.

The RuNNE Shared Task approaches the problem of nested named entity recognition. The annotation schema is designed in such a way, that an entity may partially overlap or even be nested into another entity. This way, the named entity "The Yermolova Theatre" of type "organization" houses another entity "Yermolova" of type "person". We adopt the Russian NEREL dataset for the RuNNE Shared Task. NEREL comprises news texts written in the Russian language and collected from the Wikinews portal. The annotation schema includes 29 entity types. The nestedness of named entities in NEREL reaches up to six levels. The RuNNE Shared Task explores two setups. (i) In the general setup all entities occur more or less with the same frequency. (ii) In the few-shot setup the majority of entity types occur often in the training set. However, some of the entity types are have lower frequency, being thus challenging to recognize. In the test set the frequency of all entity types is even. This paper reports on the results of the RuNNE Shared Task. Overall the shared task has received 156 submissions from nine teams. Half of the submissions outperform a straightforward BERT-based baseline in both setups. This paper overviews the shared task setup and discusses the submitted systems, discovering meaning insights for the problem of nested NER. The links to the evaluation platform and the data from the shared task are available in our github repository: https://github.com/dialogue-evaluation/RuNNE.

Micha Livne, Zulfat Miftahutdinov, Elena Tutubalina et al.

Large Language Models (LLMs) have substantially driven scientific progress in various domains, and many papers have demonstrated their ability to tackle complex problems with creative solutions. Our paper introduces a new foundation model, nach0, capable of solving various chemical and biological tasks: biomedical question answering, named entity recognition, molecular generation, molecular synthesis, attributes prediction, and others. nach0 is a multi-domain and multi-task encoder-decoder LLM pre-trained on unlabeled text from scientific literature, patents, and molecule strings to incorporate a range of chemical and linguistic knowledge. We employed instruction tuning, where specific task-related instructions are utilized to fine-tune nach0 for the final set of tasks. To train nach0 effectively, we leverage the NeMo framework, enabling efficient parallel optimization of both base and large model versions. Extensive experiments demonstrate that our model outperforms state-of-the-art baselines on single-domain and cross-domain tasks. Furthermore, it can generate high-quality outputs in molecular and textual formats, showcasing its effectiveness in multi-domain setups.

Qinghua Zhou, Ellina Aleshina, Andrey Lovyagin et al.

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.

Mark Rofin, Vladislav Mikhailov, Mikhail Florinskiy et al.

The development of state-of-the-art systems in different applied areas of machine learning (ML) is driven by benchmarks, which have shaped the paradigm of evaluating generalisation capabilities from multiple perspectives. Although the paradigm is shifting towards more fine-grained evaluation across diverse tasks, the delicate question of how to aggregate the performances has received particular interest in the community. In general, benchmarks follow the unspoken utilitarian principles, where the systems are ranked based on their mean average score over task-specific metrics. Such aggregation procedure has been viewed as a sub-optimal evaluation protocol, which may have created the illusion of progress. This paper proposes Vote'n'Rank, a framework for ranking systems in multi-task benchmarks under the principles of the social choice theory. We demonstrate that our approach can be efficiently utilised to draw new insights on benchmarking in several ML sub-fields and identify the best-performing systems in research and development case studies. The Vote'n'Rank's procedures are more robust than the mean average while being able to handle missing performance scores and determine conditions under which the system becomes the winner.

Usman Naseem, Robert Geislinger, Juan Ren et al.

We present SemEval-2026 Task 9, a shared task on online polarization detection, covering 22 languages and comprising over 110K annotated instances. Each data instance is multi-labeled with the presence of polarization, polarization type, and polarization manifestation. Participants were asked to predict labels in three sub-tasks: (1) detecting the presence of polarization, (2) identifying the type of polarization, and (3) recognizing the polarization manifestation. The three tasks attracted over 1,000 participants worldwide and more than 10k submission on Codabench. We received final submissions from 67 teams and 73 system description papers. We report the baseline results and analyze the performance of the best-performing systems, highlighting the most common approaches and the most effective methods across different subtasks and languages. The dataset of this task is publicly available.

Artem Tsypin, Leonid Ugadiarov, Kuzma Khrabrov et al.

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 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, it takes around $5 \times 10^5$ additional conformations to match the physical simulator's optimization quality. In this work, we present the Gradual Optimization Learning Framework (GOLF) for energy minimization with neural networks that significantly reduces the required additional data. 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 on par with the oracle on a benchmark of diverse drug-like molecules using $50$x less additional data.

Andrey Galichin, Alexey Dontsov, Polina Druzhinina et al.

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. Code available at https://github.com/AIRI-Institute/SAE-Reasoning

Alexey Dontsov, Dmitrii Korzh, Alexey Zhavoronkin et al.

Machine Unlearning (MU) is critical for removing private or hazardous information from deep learning models. While MU has advanced significantly in unimodal (text or vision) settings, multimodal unlearning (MMU) remains underexplored due to the lack of open benchmarks for evaluating cross-modal data removal. To address this gap, we introduce CLEAR, the first open-source benchmark designed specifically for MMU. CLEAR contains 200 fictitious individuals and 3,700 images linked with corresponding question-answer pairs, enabling a thorough evaluation across modalities. We conduct a comprehensive analysis of 11 MU methods (e.g., SCRUB, gradient ascent, DPO) across four evaluation sets, demonstrating that jointly unlearning both modalities outperforms single-modality approaches. The dataset is available at https://huggingface.co/datasets/therem/CLEAR

Vera Davydova, Huabin Yang, Elena Tutubalina

The COVID-19 pandemic has sparked numerous discussions on social media platforms, with users sharing their views on topics such as mask-wearing and vaccination. To facilitate the evaluation of neural models for stance detection and premise classification, we organized the Social Media Mining for Health (SMM4H) 2022 Shared Task 2. This competition utilized manually annotated posts on three COVID-19-related topics: school closures, stay-at-home orders, and wearing masks. In this paper, we extend the previous work and present newly collected data on vaccination from Twitter to assess the performance of models on a different topic. To enhance the accuracy and effectiveness of our evaluation, we employed various strategies to aggregate tweet texts with claims, including models with feature-level (early) fusion and dual-view architectures from SMM4H 2022 leaderboard. Our primary objective was to create a valuable dataset and perform an extensive experimental evaluation to support future research in argument mining in the health domain.

Mikhail Seleznyov, Mikhail Chaichuk, Gleb Ershov et al.

Large Language Models (LLMs) are highly sensitive to subtle, non-semantic variations in prompt phrasing and formatting. In this work, we present the first systematic evaluation of 5 methods for improving prompt robustness within a unified experimental framework. We benchmark these techniques on 8 models from Llama, Qwen and Gemma families across 52 tasks from Natural Instructions dataset. Our evaluation covers robustness methods from both fine-tuned and in-context learning paradigms, and tests their generalization against multiple types of distribution shifts. Finally, we extend our analysis to GPT-4.1 and DeepSeek V3 to assess frontier models' current robustness to format perturbations. Our findings offer actionable insights into the relative effectiveness of these robustness methods, enabling practitioners to make informed decisions when aiming for stable and reliable LLM performance in real-world applications. Code: https://github.com/AIRI-Institute/when-punctuation-matters.

Mikhail Seleznyov, Daniil Korbut, Viktor Moskvoretskii et al.

Uncertainty quantification (UQ) methods for large language models are predominantly designed by hand based on domain knowledge and heuristics, limiting their scalability and generality. We apply LLM-powered evolutionary search to automatically discover unsupervised UQ methods represented as Python programs. On the task of atomic claim verification, our evolved methods outperform strong manually-designed baselines, achieving up to 6.7% relative ROC-AUC improvement across 9 datasets while generalizing robustly out-of-distribution. Qualitative analysis reveals that different LLMs employ qualitatively distinct evolutionary strategies: Claude models consistently design high-feature-count linear estimators, while Gpt-oss-120B gravitates toward simpler and more interpretable positional weighting schemes. Surprisingly, only Sonnet 4.5 and Opus 4.5 reliably leverage increased method complexity to improve performance -- Opus 4.6 shows an unexpected regression relative to its predecessor. Overall, our results indicate that LLM-powered evolutionary search is a promising paradigm for automated, interpretable hallucination detector design.

Daniil Moskovskiy, Nikita Sushko, Sergey Pletenev et al.

Existing approaches to multilingual text detoxification are hampered by the scarcity of parallel multilingual datasets. In this work, we introduce a pipeline for the generation of multilingual parallel detoxification data. We also introduce SynthDetoxM, a manually collected and synthetically generated multilingual parallel text detoxification dataset comprising 16,000 high-quality detoxification sentence pairs across German, French, Spanish and Russian. The data was sourced from different toxicity evaluation datasets and then rewritten with nine modern open-source LLMs in few-shot setting. Our experiments demonstrate that models trained on the produced synthetic datasets have superior performance to those trained on the human-annotated MultiParaDetox dataset even in data limited setting. Models trained on SynthDetoxM outperform all evaluated LLMs in few-shot setting. We release our dataset and code to help further research in multilingual text detoxification.

Nikolas Evkarpidi, Elena Tutubalina

This paper presents a system developed for SemEval 2025 Task 8: Question Answering (QA) over tabular data. Our approach integrates several key components: text-to-SQL and text-to-code generation modules, a self-correction mechanism, and a retrieval-augmented generation (RAG). Additionally, it includes an end-to-end (E2E) module, all orchestrated by a large language model (LLM). Through ablation studies, we analyzed the effects of different parts of our pipeline and identified the challenges that are still present in this field. During the evaluation phase of the competition, our solution achieved an accuracy of 80%, resulting in a top-13 ranking among the 38 participating teams. Our pipeline demonstrates a significant improvement in accuracy for open-source models and achieves a performance comparable to proprietary LLMs in QA tasks over tables. The code is available at GitHub repository.

Oleg Somov, Mikhail Chaichuk, Mikhail Seleznyov et al.

Schema-guided reasoning pipelines ask LLMs to produce explicit intermediate structures -- rubrics, checklists, verification queries -- before committing to a final decision. But do these structures causally determine the output, or merely accompany it? We introduce a causal evaluation protocol that makes this directly measurable: by selecting tasks where a deterministic function maps intermediate structures to decisions, every controlled edit implies a unique correct output. Across eight models and three benchmarks, models appear self-consistent with their own intermediate structures but fail to update predictions after intervention in up to 60% of cases -- revealing that apparent faithfulness is fragile once the intermediate structure changes. When derivation of the final decision from the structure is delegated to an external tool, this fragility largely disappears; however, prompts which ask to prioritize the intermediate structure over the original input do not materially close the gap. Overall, intermediate structures in schema-guided pipelines function as influential context rather than stable causal mediators.

Borisiuk Anna, Andrey Savchenko, Alexander Panchenko et al.

Machine Unlearning (MU) enables Large Language Models (LLMs) to remove unsafe or outdated information. However, existing work assumes that all facts are equally forgettable and largely ignores whether the forgotten knowledge originates from pretraining or supervised fine-tuning (SFT). In this paper, we introduce DUAL (Dual Unlearning Evaluation across Training Stages), a benchmark of 28.6k Wikidata-derived triplets annotated with fact popularity using Wikipedia link counts and LLM-based salience scores. Our experiments show that pretrained and SFT models respond differently to unlearning. An SFT step on the forget data yields smoother forgetting, more stable tuning, and 10-50% higher retention, while direct unlearning on pretrained models remains unstable and prone to relearning or catastrophic forgetting.

Artem Alekseev, Mikhail Chaichuk, Miron Butko et al.

Large language models excel in question-answering (QA) yet still struggle with multi-hop reasoning and temporal questions. Query-based knowledge graph QA (KGQA) offers a modular alternative by generating executable queries instead of direct answers. We explore multi-stage query-based framework for WikiData QA, proposing multi-stage approach that enhances performance on challenging multi-hop and temporal benchmarks. Through generalization and rejection studies, we evaluate robustness across multi-hop and temporal QA datasets. Additionally, we introduce a novel entity linking and predicate matching method using CoT reasoning. Our results demonstrate the potential of query-based multi-stage KGQA framework for improving multi-hop and temporal QA with small language models. Code and data: https://github.com/ar2max/NLDB-KGQA-System

Aleksandr Nesterov, Andrey Sakhovskiy, Ivan Sviridov et al.

This study investigates the feasibility of automating clinical coding in Russian, a language with limited biomedical resources. We present a new dataset for ICD coding, which includes diagnosis fields from electronic health records (EHRs) annotated with over 10,000 entities and more than 1,500 unique ICD codes. This dataset serves as a benchmark for several state-of-the-art models, including BERT, LLaMA with LoRA, and RAG, with additional experiments examining transfer learning across domains (from PubMed abstracts to medical diagnosis) and terminologies (from UMLS concepts to ICD codes). We then apply the best-performing model to label an in-house EHR dataset containing patient histories from 2017 to 2021. Our experiments, conducted on a carefully curated test set, demonstrate that training with the automated predicted codes leads to a significant improvement in accuracy compared to manually annotated data from physicians. We believe our findings offer valuable insights into the potential for automating clinical coding in resource-limited languages like Russian, which could enhance clinical efficiency and data accuracy in these contexts. Our code and dataset are available at https://github.com/auto-icd-coding/ruccod.

Natalia Loukachevitch, Ekaterina Artemova, Tatiana Batura et al.

In this paper, we present NEREL, a Russian dataset for named entity recognition and relation extraction. NEREL is significantly larger than existing Russian datasets: to date it contains 56K annotated named entities and 39K annotated relations. Its important difference from previous datasets is annotation of nested named entities, as well as relations within nested entities and at the discourse level. NEREL can facilitate development of novel models that can extract relations between nested named entities, as well as relations on both sentence and document levels. NEREL also contains the annotation of events involving named entities and their roles in the events. The NEREL collection is available via https://github.com/nerel-ds/NEREL.

Elena Tutubalina, Ilseyar Alimova, Zulfat Miftahutdinov et al.

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 labelled one. The raw part includes 1.4 million health-related user-generated texts collected from various Internet sources, including social media. The labelled 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 labelled 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 multi-label 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. We make the RuDReC corpus and pretrained weights of domain-specific BERT models freely available at https://github.com/cimm-kzn/RuDReC

Ilya Shenbin, Anton Alekseev, Elena Tutubalina et al.

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 $β$ hyperparameter for the $β$-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.

Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara et al.

This is an overview of the thirteenth edition of the BioASQ challenge in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2025. BioASQ is a series of international challenges promoting advances in large-scale biomedical semantic indexing and question answering. This year, BioASQ consisted of new editions of the two established tasks, b and Synergy, and four new tasks: a) Task MultiClinSum on multilingual clinical summarization. b) Task BioNNE-L on nested named entity linking in Russian and English. c) Task ELCardioCC on clinical coding in cardiology. d) Task GutBrainIE on gut-brain interplay information extraction. In this edition of BioASQ, 83 competing teams participated with more than 1000 distinct submissions in total for the six different shared tasks of the challenge. Similar to previous editions, several participating systems achieved competitive performance, indicating the continuous advancement of the state-of-the-art in the field.

Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara et al.

This is an overview of the twelfth edition of the BioASQ challenge in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2024. BioASQ is a series of international challenges promoting advances in large-scale biomedical semantic indexing and question answering. This year, BioASQ consisted of new editions of the two established tasks b and Synergy, and two new tasks: a) MultiCardioNER on the adaptation of clinical entity detection to the cardiology domain in a multilingual setting, and b) BIONNE on nested NER in Russian and English. In this edition of BioASQ, 37 competing teams participated with more than 700 distinct submissions in total for the four different shared tasks of the challenge. Similarly to previous editions, most of the participating systems achieved competitive performance, suggesting the continuous advancement of the state-of-the-art in the field.

Oleg Somov, Elena Tutubalina

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.

Mikhail Chaichuk, Sushant Gautam, Steven Hicks et al.

The generation of realistic medical images from text descriptions has significant potential to address data scarcity challenges in healthcare AI while preserving patient privacy. This paper presents a comprehensive study of text-to-image synthesis in the medical domain, comparing two distinct approaches: (1) fine-tuning large pre-trained latent diffusion models and (2) training small, domain-specific models. We introduce a novel model named MSDM, an optimized architecture based on Stable Diffusion that integrates a clinical text encoder, variational autoencoder, and cross-attention mechanisms to better align medical text prompts with generated images. Our study compares two approaches: fine-tuning large pre-trained models (FLUX, Kandinsky) versus training compact domain-specific models (MSDM). Evaluation across colonoscopy (MedVQA-GI) and radiology (ROCOv2) datasets reveals that while large models achieve higher fidelity, our optimized MSDM delivers comparable quality with lower computational costs. Quantitative metrics and qualitative evaluations by medical experts reveal strengths and limitations of each approach.

Artem Vazhentsev, Maria Marina, Daniil Moskovskiy et al.

Trustworthiness is a core research challenge for agentic AI systems built on Large Language Models (LLMs). To enhance trust, natural language claims from diverse sources, including human-written text, web content, and model outputs, are commonly checked for factuality by retrieving external knowledge and using an LLM to verify the faithfulness of claims to the retrieved evidence. As a result, such methods are constrained by retrieval errors and external data availability, while leaving the models intrinsic fact-verification capabilities largely unused. We propose the task of fact-checking without retrieval, focusing on the verification of arbitrary natural language claims, independent of their source. To study this setting, we introduce a comprehensive evaluation framework focused on generalization, testing robustness to (i) long-tail knowledge, (ii) variation in claim sources, (iii) multilinguality, and (iv) long-form generation. Across 9 datasets, 18 methods and 3 models, our experiments indicate that logit-based approaches often underperform compared to those that leverage internal model representations. Building on this finding, we introduce INTRA, a method that exploits interactions between internal representations and achieves state-of-the-art performance with strong generalization. More broadly, our work establishes fact-checking without retrieval as a promising research direction that can complement retrieval-based frameworks, improve scalability, and enable the use of such systems as reward signals during training or as components integrated into the generation process.

Maksim Kuznetsov, Airat Valiev, Alex Aliper et al.

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.

Pavel Braslavski, Dmitrii Iarosh, Nikita Sushko et al.

We present a configurable pipeline for generating multilingual sets of entities with specified characteristics, such as domain, geographical location and popularity, using data from Wikipedia and Wikidata. These datasets are intended for evaluating the factuality of LLMs' long-form generation, thereby complementing evaluation based on short-form QA datasets. We present the RiDiC dataset as an example of this approach. RiDiC contains 3,000 entities from three domains -- rivers, natural disasters, and car models -- spanning different popularity tiers. Each entity is accompanied by its geographical location, English and Chinese names (if available) and relevant English and Chinese Wikipedia content, which is used to evaluate LLMs' responses. Generations about RiDiC entities were obtained from three LLMs in English and Chinese. These were then evaluated using a third-party factuality checker, which showed that entities from our dataset caused even frontier models to hallucinate. To facilitate the evaluation of LLMs' long-form factuality in multiple languages, the code, data, and generation/evaluation scripts have been released.

Anton Korznikov, Andrey Galichin, Alexey Dontsov et al.

Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features. Recent work has introduced multiple SAE variants and successfully scaled them to frontier models. Despite much excitement, a growing number of negative results in downstream tasks casts doubt on whether SAEs recover meaningful features. To directly investigate this, we perform two complementary evaluations. On a synthetic setup with known ground-truth features, we demonstrate that SAEs recover only $9\%$ of true features despite achieving $71\%$ explained variance, showing that they fail at their core task even when reconstruction is strong. To evaluate SAEs on real activations, we introduce three baselines that constrain SAE feature directions or their activation patterns to random values. Through extensive experiments across multiple SAE architectures, we show that our baselines match fully-trained SAEs in interpretability (0.87 vs 0.90), sparse probing (0.69 vs 0.72), and causal editing (0.73 vs 0.72). Together, these results suggest that SAEs in their current state do not reliably decompose models' internal mechanisms.

Viktoriia Zinkovich, Anton Antonov, Andrei Spiridonov et al.

Multimodal large language models (MLLMs) have shown impressive capabilities in vision-language tasks such as reasoning segmentation, where models generate segmentation masks based on textual queries. While prior work has primarily focused on perturbing image inputs, semantically equivalent textual paraphrases-crucial in real-world applications where users express the same intent in varied ways-remain underexplored. To address this gap, we introduce a novel adversarial paraphrasing task: generating grammatically correct paraphrases that preserve the original query meaning while degrading segmentation performance. To evaluate the quality of adversarial paraphrases, we develop a comprehensive automatic evaluation protocol validated with human studies. Furthermore, we introduce SPARTA-a black-box, sentence-level optimization method that operates in the low-dimensional semantic latent space of a text autoencoder, guided by reinforcement learning. SPARTA achieves significantly higher success rates, outperforming prior methods by up to 2x on both the ReasonSeg and LLMSeg-40k datasets. We use SPARTA and competitive baselines to assess the robustness of advanced reasoning segmentation models. We reveal that they remain vulnerable to adversarial paraphrasing-even under strict semantic and grammatical constraints. All code and data will be released publicly upon acceptance.

Tyler A. Chang, Catherine Arnett, Abdelrahman Eldesokey et al. · uw

To date, there exist almost no culturally-specific evaluation benchmarks for large language models (LLMs) that cover a large number of languages and cultures. In this paper, we present Global PIQA, a participatory commonsense reasoning benchmark for over 100 languages, constructed by hand by 335 researchers from 65 countries around the world. The 116 language varieties in Global PIQA cover five continents, 14 language families, and 23 writing systems. In the non-parallel split of Global PIQA, over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements. We find that state-of-the-art LLMs perform well on Global PIQA in aggregate, but they exhibit weaker performance in lower-resource languages (up to a 37% accuracy gap, despite random chance at 50%). Open models generally perform worse than proprietary models. Global PIQA highlights that in many languages and cultures, everyday knowledge remains an area for improvement, alongside more widely-discussed capabilities such as complex reasoning and expert knowledge. Beyond its uses for LLM evaluation, we hope that Global PIQA provides a glimpse into the wide diversity of cultures in which human language is embedded.

Daria Cherniuk, Nikita Sukhorukov, Nikita Sushko et al.

Retrieval-augmented generation has emerged as one of the most effective approaches for code completion, particularly when context from a surrounding repository is essential. However, incorporating context significantly extends sequence length, leading to slower inference - a critical limitation for interactive settings such as IDEs. In this work, we introduce LlavaCode, a framework that compresses code into compact, semantically rich representations interpretable by code LLM, enhancing generation quality while reducing the retrieved context to only a few compressed single-token vectors. Using a small projector module we can significantly increase the EM and ES metrics of coding model with negligible latency increase. Our experiments demonstrate that compressed context enables 20-38% reduction in Time-to-First-Token (TTFT) on line completion tasks compared to full-RAG pipelines.

Nikita Afonin, Nikita Andriyanov, Nikhil Bageshpura et al.

Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across three datasets, three frontier models produce broadly misaligned responses at rates between 2% and 17% given 64 narrow in-context examples, and up to 58% with 256 examples. We also examine mechanisms of EM by eliciting step-by-step reasoning (while leaving in-context examples unchanged). Manual analysis of the resulting chain-of-thought shows that 67.5% of misaligned traces explicitly rationalize harmful outputs by adopting a reckless or dangerous ''persona'', echoing prior results on finetuning-induced EM.

Anton Korznikov, Andrey Galichin, Alexey Dontsov et al.

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 2-27%. Alarmingly, steering benign features from a sparse autoencoder (SAE), a common source of interpretable directions, increases these rates by a further 2-4%. 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.

Anton Korznikov, Andrey Galichin, Alexey Dontsov et al.

Sparse autoencoders (SAEs) are a technique for sparse decomposition of neural network activations into human-interpretable features. However, current SAEs suffer from feature absorption, where specialized features capture instances of general features creating representation holes, and feature composition, where independent features merge into composite representations. In this work, we introduce Orthogonal SAE (OrtSAE), a novel approach aimed to mitigate these issues by enforcing orthogonality between the learned features. By implementing a new training procedure that penalizes high pairwise cosine similarity between SAE features, OrtSAE promotes the development of disentangled features while scaling linearly with the SAE size, avoiding significant computational overhead. We train OrtSAE across different models and layers and compare it with other methods. We find that OrtSAE discovers 9% more distinct features, reduces feature absorption (by 65%) and composition (by 15%), improves performance on spurious correlation removal (+6%), and achieves on-par performance for other downstream tasks compared to traditional SAEs.

Andrey Sakhovskiy, Elena Tutubalina

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 Alignment), 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.

Pavel Tikhonov, Ivan Oseledets, Elena Tutubalina

In-context learning (ICL) enables Large Language Models (LLMs) to adapt to new tasks using few examples, with task vectors - specific hidden state activations - hypothesized to encode task information. Existing studies are limited by small-scale benchmarks, restricting comprehensive analysis. We introduce QuiteAFew, a novel dataset of 3,096 diverse few-shot tasks, each with 30 input-output pairs derived from the Alpaca dataset. Experiments with Llama-3-8B on QuiteAFew reveal: (1) task vector performance peaks at an intermediate layer (e.g., 15th), (2) effectiveness varies significantly by task type, and (3) complex tasks rely on multiple, subtask-specific vectors rather than a single vector, suggesting distributed task knowledge representation.

Kuzma Khrabrov, Anton Ber, Artem Tsypin et al.

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 ($ω$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.

Anton Alekseev, Elena Tutubalina, Sejeong Kwon et al.

In this work, we explore the constructive side of online reviews: advice, tips, requests, and suggestions that users provide about goods, venues, services, and other items of interest. To reduce training costs and annotation efforts needed to build a classifier for a specific label set, we present and evaluate several entailment-based zero-shot approaches to suggestion classification in a label-fully-unseen fashion. In particular, we introduce the strategy of assigning target class labels to sentences in English language with user intentions, which significantly improves prediction quality. The proposed strategies are evaluated with a comprehensive experimental study that validated our results both quantitatively and qualitatively.

Ilseyar Alimova, Elena Tutubalina

Automatic monitoring of adverse drug events (ADEs) or reactions (ADRs) is currently receiving significant attention from the biomedical community. In recent years, user-generated data on social media has become a valuable resource for this task. Neural models have achieved impressive performance on automatic text classification for ADR detection. Yet, training and evaluation of these methods are carried out on user-generated texts about a targeted drug. In this paper, we assess the robustness of state-of-the-art neural architectures across different drug groups. We investigate several strategies to use pseudo-labeled data in addition to a manually annotated train set. Out-of-dataset experiments diagnose the bottleneck of supervised models in terms of breakdown performance, while additional pseudo-labeled data improves overall results regardless of the text selection strategy.

Daria Bakshandaeva, Denis Dimitrov, Vladimir Arkhipkin et al.

Supporting the current trend in the AI community, we present the AI Journey 2021 Challenge called Fusion Brain, the first competition which is targeted to make the universal architecture which could process different modalities (in this case, images, texts, and code) and solve multiple tasks for vision and language. The Fusion Brain Challenge combines the following specific tasks: Code2code Translation, Handwritten Text recognition, Zero-shot Object Detection, and Visual Question Answering. We have created datasets for each task to test the participants' submissions on it. Moreover, we have collected and made publicly available a new handwritten dataset in both English and Russian, which consists of 94,128 pairs of images and texts. We also propose a multimodal and multitask architecture - a baseline solution, in the center of which is a frozen foundation model and which has been trained in Fusion mode along with Single-task mode. The proposed Fusion approach proves to be competitive and more energy-efficient compared to the task-specific one.

Zulfat Miftahutdinov, Artur Kadurin, Roman Kudrin et al.

Concept normalization in free-form texts is a crucial step in every text-mining pipeline. Neural architectures based on Bidirectional Encoder Representations from Transformers (BERT) have achieved state-of-the-art results in the biomedical domain. In the context of drug discovery and development, clinical trials are necessary to establish the efficacy and safety of drugs. We investigate the effectiveness of transferring concept normalization from the general biomedical domain to the clinical trials domain in a zero-shot setting with an absence of labeled data. We propose a simple and effective two-stage neural approach based on fine-tuned BERT architectures. In the first stage, we train a metric learning model that optimizes relative similarity of mentions and concepts via triplet loss. The model is trained on available labeled corpora of scientific abstracts to obtain vector embeddings of concept names and entity mentions from texts. In the second stage, we find the closest concept name representation in an embedding space to a given clinical mention. We evaluated several models, including state-of-the-art architectures, on a dataset of abstracts and a real-world dataset of trial records with interventions and conditions mapped to drug and disease terminologies. Extensive experiments validate the effectiveness of our approach in knowledge transfer from the scientific literature to clinical trials.

Vitaly Ivanin, Ekaterina Artemova, Tatiana Batura et al.

We show-case an application of information extraction methods, such as named entity recognition (NER) and relation extraction (RE) to a novel corpus, consisting of documents, issued by a state agency. The main challenges of this corpus are: 1) the annotation scheme differs greatly from the one used for the general domain corpora, and 2) the documents are written in a language other than English. Unlike expectations, the state-of-the-art transformer-based models show modest performance for both tasks, either when approached sequentially, or in an end-to-end fashion. Our experiments have demonstrated that fine-tuning on a large unlabeled corpora does not automatically yield significant improvement and thus we may conclude that more sophisticated strategies of leveraging unlabelled texts are demanded. In this paper, we describe the whole developed pipeline, starting from text annotation, baseline development, and designing a shared task in hopes of improving the baseline. Eventually, we realize that the current NER and RE technologies are far from being mature and do not overcome so far challenges like ours.

Ekaterina Artemova, Tatiana Batura, Anna Golenkovskaya et al.

In this paper we present a corpus of Russian strategic planning documents, RuREBus. This project is grounded both from language technology and e-government perspectives. Not only new language sources and tools are being developed, but also their applications to e-goverment research. We demonstrate the pipeline for creating a text corpus from scratch. First, the annotation schema is designed. Next texts are marked up using human-in-the-loop strategy, so that preliminary annotations are derived from a machine learning model and are manually corrected. The amount of annotated texts is large enough to showcase what insights can be gained from RuREBus.

Anton Alekseev, Elena Tutubalina, Valentin Malykh et al.

Deep learning architectures based on self-attention have recently achieved and surpassed state of the art results in the task of unsupervised aspect extraction and topic modeling. While models such as neural attention-based aspect extraction (ABAE) have been successfully applied to user-generated texts, they are less coherent when applied to traditional data sources such as news articles and newsgroup documents. In this work, we introduce a simple approach based on sentence filtering in order to improve topical aspects learned from newsgroups-based content without modifying the basic mechanism of ABAE. We train a probabilistic classifier to distinguish between out-of-domain texts (outer dataset) and in-domain texts (target dataset). Then, during data preparation we filter out sentences that have a low probability of being in-domain and train the neural model on the remaining sentences. The positive effect of sentence filtering on topic coherence is demonstrated in comparison to aspect extraction models trained on unfiltered texts.

Juan M. Banda, Ramya Tekumalla, Guanyu Wang et al.

As the COVID-19 pandemic continues its march around the world, an unprecedented amount of open data is being generated for genetics and epidemiological research. The unparalleled rate at which many research groups around the world are releasing data and publications on the ongoing pandemic is allowing other scientists to learn from local experiences and data generated in the front lines of the COVID-19 pandemic. However, there is a need to integrate additional data sources that map and measure the role of social dynamics of such a unique world-wide event into biomedical, biological, and epidemiological analyses. For this purpose, we present a large-scale curated dataset of over 152 million tweets, growing daily, related to COVID-19 chatter generated from January 1st to April 4th at the time of writing. This open dataset will allow researchers to conduct a number of research projects relating to the emotional and mental responses to social distancing measures, the identification of sources of misinformation, and the stratified measurement of sentiment towards the pandemic in near real time.

Sergey Nikolenko, Elena Tutubalina, Zulfat Miftahutdinov et al.

We introduce an entity-centric search engineCommentsRadarthatpairs entity queries with articles and user opinions covering a widerange of topics from top commented sites. The engine aggregatesarticles and comments for these articles, extracts named entities,links them together and with knowledge base entries, performssentiment analysis, and aggregates the results, aiming to mine fortemporal trends and other insights. In this work, we present thegeneral engine, discuss the models used for all steps of this pipeline,and introduce several case studies that discover important insightsfrom online commenting data.