Ivan P. Yamshchikov

Taido Purason, Pavel Chizhov, Ivan P. Yamshchikov et al.

Tokenizer adaptation plays an important role in transferring pre-trained language models to new domains or languages. In this work, we address two complementary aspects of this process: vocabulary extension and pruning. The common approach to extension trains a new tokenizer on domain-specific text and appends the tokens that do not overlap with the existing vocabulary, which often results in many tokens that are unreachable or never used. We propose continued BPE training, which adapts a pre-trained tokenizer by continuing the BPE merge learning process on new data. Experiments across multiple languages and model families show that this approach improves tokenization efficiency and leads to better utilization of added vocabulary. We also introduce leaf-based vocabulary pruning, which removes redundant tokens while preserving model quality. Together, these methods provide practical tools for controlled vocabulary modification, which we release as an open-source package.

Ivan P. Yamshchikov, Alexey Tikhonov, Yorgos Pantis et al.

The extensive surviving corpus of the ancient scholar Plutarch of Chaeronea (ca. 45-120 CE) also contains several texts which, according to current scholarly opinion, did not originate with him and are therefore attributed to an anonymous author Pseudo-Plutarch. These include, in particular, the work Placita Philosophorum (Quotations and Opinions of the Ancient Philosophers), which is extremely important for the history of ancient philosophy. Little is known about the identity of that anonymous author and its relation to other authors from the same period. This paper presents a BERT language model for Ancient Greek. The model discovers previously unknown statistical properties relevant to these literary, philosophical, and historical problems and can shed new light on this authorship question. In particular, the Placita Philosophorum, together with one of the other Pseudo-Plutarch texts, shows similarities with the texts written by authors from an Alexandrian context (2nd/3rd century CE).

Ivan P. Yamshchikov, Alexey Tikhonov

This paper argues that a deeper understanding of narrative and the successful generation of longer subjectively interesting texts is a vital bottleneck that hinders the progress in modern Natural Language Processing (NLP) and may even be in the whole field of Artificial Intelligence. We demonstrate that there are no adequate datasets, evaluation methods, and even operational concepts that could be used to start working on narrative processing.

Alexey Tikhonov, Ivan P. Yamshchikov

In the rapidly evolving landscape of Large Language Models (LLMs), introduction of well-defined and standardized evaluation methodologies remains a crucial challenge. This paper traces the historical trajectory of LLM evaluations, from the foundational questions posed by Alan Turing to the modern era of AI research. We categorize the evolution of LLMs into distinct periods, each characterized by its unique benchmarks and evaluation criteria. As LLMs increasingly mimic human-like behaviors, traditional evaluation proxies, such as the Turing test, have become less reliable. We emphasize the pressing need for a unified evaluation system, given the broader societal implications of these models. Through an analysis of common evaluation methodologies, we advocate for a qualitative shift in assessment approaches, underscoring the importance of standardization and objective criteria. This work serves as a call for the AI community to collaboratively address the challenges of LLM evaluation, ensuring their reliability, fairness, and societal benefit.

Pavel Chizhov, Catherine Arnett, Elizaveta Korotkova et al.

Language models can largely benefit from efficient tokenization. However, they still mostly utilize the classical BPE algorithm, a simple and reliable method. This has been shown to cause such issues as under-trained tokens and sub-optimal compression that may affect the downstream performance. We introduce Picky BPE, a modified BPE algorithm that carries out vocabulary refinement during tokenizer training. Our method improves vocabulary efficiency, eliminates under-trained tokens, and does not compromise text compression. Our experiments show that our method does not reduce the downstream performance, and in several cases improves it.

Catherine Arnett, Eliot Jones, Ivan P. Yamshchikov et al.

Open-source large language models are becoming increasingly available and popular among researchers and practitioners. While significant progress has been made on open-weight models, open training data is a practice yet to be adopted by the leading open-weight models creators. At the same time, there researchers are working to make language models safer. We propose a data curation pipeline to reduce harmful outputs by models trained on public domain data. There are unique challenges to working with public domain data, as these sources differ from web text in both form and content. Many sources are historical documents and are the result of Optical Character Recognition (OCR). Consequently, current state-of-the-art approaches to toxicity filtering are often infeasible or inappropriate for open data models. In this paper, we introduce a new fully open-source pipeline for open-data toxicity filtering. Our contributions are threefold. We create a custom training dataset, ToxicCommons, which is composed of texts which have been classified across five different dimensions (racial/origin-based, gender/sex-based, religious, ability-based discrimination, and violence). We use this dataset to train a custom classifier, Celadon, that can be used to detect toxic content in open data more efficiently at a larger scale. Finally, we describe the balanced approach to content filtration that optimizes safety filtering with respect to the filtered data available for training.

Elizaveta Zhemchuzhina, Nikolai Filippov, Ivan P. Yamshchikov

This paper studies the limits of language models' statistical learning in the context of Zipf's law. First, we demonstrate that Zipf-law token distribution emerges irrespective of the chosen tokenization. Second, we show that Zipf distribution is characterized by two distinct groups of tokens that differ both in terms of their frequency and their semantics. Namely, the tokens that have a one-to-one correspondence with one semantic concept have different statistical properties than those with semantic ambiguity. Finally, we demonstrate how these properties interfere with statistical learning procedures motivated by distributional semantics.

Priyanka Singh, Vladislav D. Mosin, Ivan P. Yamshchikov

Working within specific NLP subdomains presents significant challenges, primarily due to a persistent deficit of data. Stringent privacy concerns and limited data accessibility often drive this shortage. Additionally, the medical domain demands high accuracy, where even marginal improvements in model performance can have profound impacts. In this study, we investigate the potential of vocabulary transfer to enhance model performance in biomedical NLP tasks. Specifically, we focus on vocabulary extension, a technique that involves expanding the target vocabulary to incorporate domain-specific biomedical terms. Our findings demonstrate that vocabulary extension, leads to measurable improvements in both downstream model performance and inference time.

Alexey Tikhonov, Lisa Bylinina, Ivan P. Yamshchikov

We show differences between a language-and-vision model CLIP and two text-only models - FastText and SBERT - when it comes to the encoding of individuation information. We study latent representations that CLIP provides for substrates, granular aggregates, and various numbers of objects. We demonstrate that CLIP embeddings capture quantitative differences in individuation better than models trained on text-only data. Moreover, the individuation hierarchy we deduce from the CLIP embeddings agrees with the hierarchies proposed in linguistics and cognitive science.

Anna Bykova, Nikolay Filippov, Ivan P. Yamshchikov

This paper presents a large anonymized dataset of homelessness alongside insights into the data-driven rehabilitation of homeless people. The dataset was gathered by a large nonprofit organization working on rehabilitating the homeless for twenty years. This is the first dataset that we know of that contains rich information on thousands of homeless individuals seeking rehabilitation. We show how data analysis can help to make the rehabilitation of homeless people more effective and successful. Thus, we hope this paper alerts the data science community to the problem of homelessness.

Evgenii Kniazev, Arseny Kravchenko, Igor Rekun et al.

We present app.build (https://github.com/appdotbuild/agent/), an open-source framework that improves LLM-based application generation through systematic validation and structured environments. Our approach combines multi-layered validation pipelines, stack-specific orchestration, and model-agnostic architecture, implemented across three reference stacks. Through evaluation on 30 generation tasks, we demonstrate that comprehensive validation achieves 73.3% viability rate with 30% reaching perfect quality scores, while open-weights models achieve 80.8% of closed-model performance when provided structured environments. The open-source framework has been adopted by the community, with over 3,000 applications generated to date. This work demonstrates that scaling reliable AI agents requires scaling environments, not just models -- providing empirical insights and complete reference implementations for production-oriented agent systems.

Pierre-Carl Langlais, Pavel Chizhov, Mattia Nee et al.

We introduce a new generation of small reasoning models for RAG, search, and source summarization. Pleias-RAG-350m and Pleias-RAG-1B are mid-trained on a large synthetic dataset emulating the retrieval of a wide variety of multilingual open sources from the Common Corpus. They provide native support for citation and grounding with literal quotes and reintegrate multiple features associated with RAG workflows, such as query routing, query reformulation, and source reranking. Pleias-RAG-350m and Pleias-RAG-1B outperform SLMs below 4 billion parameters on standardized RAG benchmarks (HotPotQA, 2wiki) and are competitive with popular larger models, including Qwen-2.5-7B, Llama-3.1-8B, and Gemma-3-4B. They are the only SLMs to date maintaining consistent RAG performance across leading European languages and ensuring systematic reference grounding for statements. Due to their size and ease of deployment on constrained infrastructure and higher factuality by design, the models unlock a range of new use cases for generative AI.

Gleb Schmidt, Svetlana Gorovaia, Ivan P. Yamshchikov

This paper evaluates the performance of Large Language Models (LLMs) in authorship attribution and authorship verification tasks for Latin texts of the Patristic Era. The study showcases that LLMs can be robust in zero-shot authorship verification even on short texts without sophisticated feature engineering. Yet, the models can also be easily "mislead" by semantics. The experiments also demonstrate that steering the model's authorship analysis and decision-making is challenging, unlike what is reported in the studies dealing with high-resource modern languages. Although LLMs prove to be able to beat, under certain circumstances, the traditional baselines, obtaining a nuanced and truly explainable decision requires at best a lot of experimentation.

Aleksandra Sorokovikova, Pavel Chizhov, Iuliia Eremenko et al.

Modern language models are trained on large amounts of data. These data inevitably include controversial and stereotypical content, which contains all sorts of biases related to gender, origin, age, etc. As a result, the models express biased points of view or produce different results based on the assigned personality or the personality of the user. In this paper, we investigate various proxy measures of bias in large language models (LLMs). We find that evaluating models with pre-prompted personae on a multi-subject benchmark (MMLU) leads to negligible and mostly random differences in scores. However, if we reformulate the task and ask a model to grade the user's answer, this shows more significant signs of bias. Finally, if we ask the model for salary negotiation advice, we see pronounced bias in the answers. With the recent trend for LLM assistant memory and personalization, these problems open up from a different angle: modern LLM users do not need to pre-prompt the description of their persona since the model already knows their socio-demographics.

Pierre-Carl Langlais, Carlos Rosas Hinostroza, Mattia Nee et al.

Large Language Models (LLMs) are pre-trained on large amounts of data from different sources and domains. These data most often contain trillions of tokens with large portions of copyrighted or proprietary content, which hinders the usage of such models under AI legislation. This raises the need for truly open pre-training data that is compliant with the data security regulations. In this paper, we introduce Common Corpus, the largest open dataset for language model pre-training. The data assembled in Common Corpus are either uncopyrighted or under permissible licenses and amount to about two trillion tokens. The dataset contains a wide variety of languages, ranging from the main European languages to low-resource ones rarely present in pre-training datasets; in addition, it includes a large portion of code data. The diversity of data sources in terms of covered domains and time periods opens up the paths for both research and entrepreneurial needs in diverse areas of knowledge. In this technical report, we present the detailed provenance of data assembling and the details of dataset filtering and curation. Being already used by such industry leaders as Anthropic and multiple LLM training projects, we believe that Common Corpus will become a critical infrastructure for open science research in LLMs.

Dmitry Vikhorev, Daria Galimzianova, Svetlana Gorovaia et al.

Humor generation is a challenging task in natural language processing due to limited resources and the quality of existing datasets. Available humor language resources often suffer from toxicity and duplication, limiting their effectiveness for training robust models. This paper proposes CleanComedy, a specialized, partially annotated toxicity-filtered corpus of English and Russian jokes collected from various sources. We study the effectiveness of our data filtering approach through a survey on humor and toxicity levels in various joke groups. In addition, we study advances in computer humor generation by comparing jokes written by humans with various groups of generative jokes, including our baseline models trained on the CleanComedy datasets.

Johannes Römisch, Svetlana Gorovaia, Mariia Halchynska et al.

This article explores the zero-shot performance of state-of-the-art large language models (LLMs) on one of the most challenging tasks in authorship analysis: sentence-level style change detection. Benchmarking four LLMs on the official PAN~2024 and 2025 "Multi-Author Writing Style Analysis" datasets, we present several observations. First, state-of-the-art generative models are sensitive to variations in writing style - even at the granular level of individual sentences. Second, their accuracy establishes a challenging baseline for the task, outperforming suggested baselines of the PAN competition. Finally, we explore the influence of semantics on model predictions and present evidence suggesting that the latest generation of LLMs may be more sensitive to content-independent and purely stylistic signals than previously reported.

Maxim K. Surkov, Ivan P. Yamshchikov

Evaluation plays a significant role in modern natural language processing. Most modern NLP benchmarks consist of arbitrary sets of tasks that neither guarantee any generalization potential for the model once applied outside the test set nor try to minimize the resource consumption needed for model evaluation. This paper presents a theoretical instrument and a practical algorithm to calculate similarity between benchmark tasks, we call this similarity measure "Vygotsky distance". The core idea of this similarity measure is that it is based on relative performance of the "students" on a given task, rather that on the properties of the task itself. If two tasks are close to each other in terms of Vygotsky distance the models tend to have similar relative performance on them. Thus knowing Vygotsky distance between tasks one can significantly reduce the number of evaluation tasks while maintaining a high validation quality. Experiments on various benchmarks, including GLUE, SuperGLUE, CLUE, and RussianSuperGLUE, demonstrate that a vast majority of NLP benchmarks could be at least 40% smaller in terms of the tasks included. Most importantly, Vygotsky distance could also be used for the validation of new tasks thus increasing the generalization potential of the future NLP models.

Gleb Schmidt, Johannes Römisch, Mariia Halchynska et al.

Style change detection - identifying the points in a document where writing style shifts - remains one of the most important and challenging problems in computational authorship analysis. At PAN 2025, the shared task challenges participants to detect style switches at the most fine-grained level: individual sentences. The task spans three datasets, each designed with controlled and increasing thematic variety within documents. We propose to address this problem by modeling the content of each problem instance - that is, a series of sentences - as a whole, using a Sequential Sentence Pair Classifier (SSPC). The architecture leverages a pre-trained language model (PLM) to obtain representations of individual sentences, which are then fed into a bidirectional LSTM (BiLSTM) to contextualize them within the document. The BiLSTM-produced vectors of adjacent sentences are concatenated and passed to a multi-layer perceptron for prediction per adjacency. Building on the work of previous PAN participants classical text segmentation, the approach is relatively conservative and lightweight. Nevertheless, it proves effective in leveraging contextual information and addressing what is arguably the most challenging aspect of this year's shared task: the notorious problem of "stylistically shallow", short sentences that are prevalent in the proposed benchmark data. Evaluated on the official PAN-2025 test datasets, the model achieves strong macro-F1 scores of 0.923, 0.828, and 0.724 on the EASY, MEDIUM, and HARD data, respectively, outperforming not only the official random baselines but also a much more challenging one: claude-3.7-sonnet's zero-shot performance.

Christeena Varghese, Sergey Koshelev, Ivan P. Yamshchikov

This study explores four methods of generating paraphrases in Malayalam, utilizing resources available for English paraphrasing and pre-trained Neural Machine Translation (NMT) models. We evaluate the resulting paraphrases using both automated metrics, such as BLEU, METEOR, and cosine similarity, as well as human annotation. Our findings suggest that automated evaluation measures may not be fully appropriate for Malayalam, as they do not consistently align with human judgment. This discrepancy underscores the need for more nuanced paraphrase evaluation approaches especially for highly agglutinative languages.

Ilya Smirnov, Ivan P. Yamshchikov

The word mover's distance (WMD) is a popular semantic similarity metric for two texts. This position paper studies several possible extensions of WMD. We experiment with the frequency of words in the corpus as a weighting factor and the geometry of the word vector space. We validate possible extensions of WMD on six document classification datasets. Some proposed extensions show better results in terms of the k-nearest neighbor classification error than WMD.

Maxim K. Surkov, Vladislav D. Mosin, Ivan P. Yamshchikov

Current state-of-the-art NLP systems use large neural networks that require lots of computational resources for training. Inspired by human knowledge acquisition, researchers have proposed curriculum learning, - sequencing of tasks (task-based curricula) or ordering and sampling of the datasets (data-based curricula) that facilitate training. This work investigates the benefits of data-based curriculum learning for large modern language models such as BERT and T5. We experiment with various curricula based on a range of complexity measures and different sampling strategies. Extensive experiments on different NLP tasks show that curricula based on various complexity measures rarely has any benefits while random sampling performs either as well or better than curricula.

Alexey Tikhonov, Ivan P. Yamshchikov

This paper presents a new natural language processing task - Actionable Entities Recognition (AER) - recognition of entities that protagonists could interact with for further plot development. Though similar to classical Named Entity Recognition (NER), it has profound differences. In particular, it is crucial for interactive fiction, where the agent needs to detect entities that might be useful in the future. We also discuss if AER might be further helpful for the systems dealing with narrative processing since actionable entities profoundly impact the causal relationship in a story. We validate the proposed task on two previously available datasets and present a new benchmark dataset for the AER task that includes 5550 descriptions with one or more actionable entities.

Shaul Solomon, Adam Cohn, Hernan Rosenblum et al.

Estimation of semantic similarity is crucial for a variety of natural language processing (NLP) tasks. In the absence of a general theory of semantic information, many papers rely on human annotators as the source of ground truth for semantic similarity estimation. This paper investigates the ambiguities inherent in crowd-sourced semantic labeling. It shows that annotators that treat semantic similarity as a binary category (two sentences are either similar or not similar and there is no middle ground) play the most important role in the labeling. The paper offers heuristics to filter out unreliable annotators and stimulates further discussions on human perception of semantic similarity.

Anastasia Malysheva, Alexey Tikhonov, Ivan P. Yamshchikov

Narrative generation and analysis are still on the fringe of modern natural language processing yet are crucial in a variety of applications. This paper proposes a feature extraction method for plot dynamics. We present a dataset that consists of the plot descriptions for thirteen thousand TV shows alongside meta-information on their genres and dynamic plots extracted from them. We validate the proposed tool for plot dynamics extraction and discuss possible applications of this method to the tasks of narrative analysis and generation.

Igor Samenko, Alexey Tikhonov, Ivan P. Yamshchikov

This paper shows that, modern word embeddings contain information that distinguishes synonyms and antonyms despite small cosine similarities between corresponding vectors. This information is encoded in the geometry of the embeddings and could be extracted with a straight-forward and intuitive manifold learning procedure or a contrasting map. Such a map is trained on a small labeled subset of the data and can produce new embeddings that explicitly highlight specific semantic attributes of the word. The new embeddings produced by the map are shown to improve the performance on downstream tasks.