Arbeiten

Modern AI systems often comprise multiple learnable components that can be naturally organized as graphs. A central challenge is the end-to-end training of such systems without restrictive architectural or training assumptions. Such tasks fit the theory and approaches of the collaborative Multi-Agent Reinforcement Learning (MARL) field. We introduce Reinforcement Networks, a general framework for MARL that organizes agents as vertices in a directed acyclic graph (DAG). This structure extends hierarchical RL to arbitrary DAGs, enabling flexible credit assignment and scalable coordination while avoiding strict topologies, fully centralized training, and other limitations of current approaches. We formalize training and inference methods for the Reinforcement Networks framework and connect it to the LevelEnv concept to support reproducible construction, training, and evaluation. We demonstrate the effectiveness of our approach on several collaborative MARL setups by developing several Reinforcement Networks models that achieve improved performance over standard MARL baselines. Beyond empirical gains, Reinforcement Networks unify hierarchical, modular, and graph-structured views of MARL, opening a principled path toward designing and training complex multi-agent systems. We conclude with theoretical and practical directions - richer graph morphologies, compositional curricula, and graph-aware exploration. That positions Reinforcement Networks as a foundation for a new line of research in scalable, structured MARL.

Text trees as a data structure occur in numerous machine learning tasks like hierarchical summarization and automatic mind map generation. One of the main methods of quality evaluation in these tasks is comparison with reference hierarchies created by experts. The method used so far to compare text hierarchies, as shown in this work, poorly accounts for their structure and text semantics relative to phrasing. To address this issue, we propose a new metric on the set of text trees — text tree edit distance (TTED), based on tree edit distance with semantic distance between texts measured using a large language model. To evaluate how the metric reflects different aspects of text tree difference, we introduce special quality coefficients that reflect the sensitivity of a metric to paraphrasing relative to structural and semantic differences of text trees. Using these coefficients, we conduct extensive testing of the proposed metric and its modifications compared to a baseline used in previous works to compare text hierarchies, which shows that TTED indeed captures significant differences between text trees more accurately than the previously used method. We also provide a practical implementation of TTED for further usage.

Against the backdrop of active LLM development, their tendency to hallucinate, as well as the growing volume of texts they generate, the validation of facts has become increasingly important and relevant.We propose ruSciFact 1 , a new benchmark for fact-checking scientific claims in Russian.ruSciFact is structured as a NLI task, the goal is to verify whether a fact is confirmed by a given abstract.To generate facts, we used an 3-step pipeline based on LLaMA-405B, validating the resulting sentences with the help of assessors-terminologists.The ruSciFact dataset consists of 1128 pairs in the format , which we are releasing as open source together with the benchmark code.Additionally, we are opensourcing the fact-generation pipeline 2 , which facilitates the expansion of the dataset to specific scientific domains.We evaluated several popular language models on ruSciFact, including text embedders and generative models.The results show that this benchmark allows to effectively assess the fact-checking capabilities of LLMs in Russian.

A wide class of natural language processing tasks is solved using markup.At the moment, the vast majority of models and datasets rely on a simple markup structure containing only fragments and labels.Moreover, simple classification metrics such as F1, Precision, Recall are used to evaluate the model's accuracy.The problem with such metrics is that they do not take into account all aspects of the markup structure and that they are applicable only under the assumption of the existence of an ideal markup.This paper proposes a more general and universal markup structure that allows solving complex problems and builds a methodology for multi-criteria evaluation of text markup models based on inconsistent expert markup.After that, the application of the constructed method is considered to assess the quality of the model obtained within the winning algorithm of the "READ//ABLE" competition, which focused on building an effective essay markup system.The results demonstrate that the new markup structure and evaluation approach provides a more comprehensive and accurate assessment of model performance, addressing the limitations of traditional metrics by accounting for complex markup scenarios and expert inconsistencies.

The Universal data markup structures empower the annotation of multiple text fragments (multi-spans or multi-fragments furthermore) to analyze large collections of content. Multi-spans have demonstrated helpful in tackling issues related to the programmed discovery of semantic blunders in school essays or human values in social media writings. Labeling multi-fragment information has made it conceivable to form an interdisciplinary classification of human values. This classifier consists of 105 labels grouped into 7 categories, and a corresponding dataset has been created. Subsequent ML experiments have been designed to demonstrate the effectiveness of the multi-spans structure in recovering annotations of human values. The accuracy of the multi-fragment detector is 0.943 for material values and 0.957 for legal awareness (a subject of civic engagement and citizenship).

Modern machine translation systems are trained on large volumes of parallel data obtained using heuristic methods of the Internet bypassing. The poor quality of the data leads to systematic translation errors, which can be quite noticeable from the human point of view. To fix such errors a human based models hypotheses re-ranking is introduced in this work. In this paper the use of human markup is shown not only to increase the overall quality of translation, but also to significantly reduce the number of systematic translation errors. In addition, the relative simplicity of human markup and its integration in the model training process opens up new opportunities in the field of domain adaptation of translation models for new domains like online retail.

Topic modelling is fundamentally a soft clustering problem (of known objects – documents, over unknown clusters – topics). That is, the task is incorrectly posed. In particular, the topic models are unstable and incomplete. All this leads to the fact that the process of finding a good topic model (repeated hyperparameter selection, model training, and topic quality assessment) can be particularly long and labor-intensive. We aim to simplify the process, to make it more deterministic and provable. To this end, we present a method for iterative training of a topic model. The essence of the method is that a series of related topic models are trained so that each subsequent model is at least as good as the previous one, i.e., that it retains all the good topics found earlier. The connection between the models is achieved by additive regularization. The result of this iterative training is the last topic model in the series, which we call the iteratively updated additively regularized topic model (ITAR). Experiments conducted on several collections of natural language texts show that the proposed ITAR model performs better than other popular topic models (LDA, ARTM, BERTopic), its topics are diverse, and its perplexity (ability to "explain" the underlying data) is moderate.

This paper presents the solution of the LomonosovMSU team for the SemEval-2024 Task 4 "Multilingual Detection of Persuasion Techniques in Memes" competition for the English language task.During the task solving process, generative and BERT-like (training classifiers on top of embedder models) approaches were tested for subtask №1, as well as an BERT-like approach on top of multimodal embedder models for subtasks №2a/№2b.The models were trained using datasets provided by the competition organizers, enriched with filtered datasets from previous SemEval competitions.The following results were achieved: 18th place for subtask №1, 9th place for subtask №2a, and 11th place for subtask №2b.The code for the solutions is available at github 1 .

Modern natural language processing models such as transformers operate multimodal data. In the present paper, multimodal data is explored using multimodal topic modeling on transactional data of bank corporate clients. A definition of the importance of modality for the model is proposed on the basis of which improvements are considered for two modeling scenarios: preserving the maximum amount of information by balancing modalities and automatic selection of modality weights to optimize auxiliary criteria based on topic representations of documents. A model is proposed for adding numerical data to topic models in the form of modalities: each topic is assigned a normal distribution with learning parameters. Significant improvements are demonstrated in comparison with standard topic models on the problem of modeling bank corporate clients. Based on the topic representations of the bank’s customers, a 90-day delay on the loan is predicted.

This thematic issue presents selected articles of the 20th All-Russian Conference with international participation “Mathematical Pattern Recognition Methods” (MPRM) held on December 7–10, 2021 in Moscow.