Ivan Smirnov

Gleb Kuzmin, Neemesh Yadav, Ivan Smirnov et al.

We propose selective debiasing -- an inference-time safety mechanism designed to enhance the overall model quality in terms of prediction performance and fairness, especially in scenarios where retraining the model is impractical. The method draws inspiration from selective classification, where at inference time, predictions with low quality, as indicated by their uncertainty scores, are discarded. In our approach, we identify the potentially biased model predictions and, instead of discarding them, we remove bias from these predictions using LEACE -- a post-processing debiasing method. To select problematic predictions, we propose a bias quantification approach based on KL divergence, which achieves better results than standard uncertainty quantification methods. Experiments on text classification datasets with encoder-based classification models demonstrate that selective debiasing helps to reduce the performance gap between post-processing methods and debiasing techniques from the at-training and pre-processing categories.

Nikita Remnev, Sergei Obiedkov, Ekaterina Rakhilina et al.

Grammatical error correction is one of the fundamental tasks in Natural Language Processing. For the Russian language, most of the spellcheckers available correct typos and other simple errors with high accuracy, but often fail when faced with non-native (L2) writing, since the latter contains errors that are not typical for native speakers. In this paper, we propose a pipeline involving a language model intended for correcting errors in L2 Russian writing. The language model proposed is trained on untagged texts of the Newspaper subcorpus of the Russian National Corpus, and the quality of the model is validated against the RULEC-GEC corpus.

Yulong He, Ivan Smirnov, Dmitry Fedrushkov et al.

Effective evaluation of large language models (LLMs) remains a critical bottleneck, as conventional direct scoring often yields inconsistent and opaque judgments. In this work, we adapt the Analytic Hierarchy Process (AHP) to LLM-based evaluation and, more importantly, propose a confidence-aware Fuzzy AHP (FAHP) extension that models epistemic uncertainty via triangular fuzzy numbers modulated by LLM-generated confidence scores. Systematically validated on JudgeBench, our structured approach decomposes assessments into explicit criteria and incorporates uncertainty-aware aggregation, producing more calibrated judgments. Extensive experiments demonstrate that both crisp and fuzzy AHP consistently outperform direct scoring across model scales and dataset splits, with FAHP showing superior stability in uncertain comparison scenarios. Building on these insights, we propose \textbf{DualJudge}, a hybrid framework inspired by Dual-Process Theory that adaptively fuses holistic direct scores with structured AHP outputs via consistency-aware weighting. DualJudge achieves state-of-the-art performance, underscoring the complementary strengths of intuitive and deliberative evaluation paradigms. These results establish uncertainty-aware structured reasoning as a principled pathway toward more reliable LLM assessment. Code is available at https://github.com/hreyulog/AHP_llm_judge.

Ivan Smirnov, Camelia Oprea, Markus Strohmaier

Wikipedia is one of the most successful collaborative projects in history. It is the largest encyclopedia ever created, with millions of users worldwide relying on it as the first source of information as well as for fact-checking and in-depth research. As Wikipedia relies solely on the efforts of its volunteer-editors, its success might be particularly affected by toxic speech. In this paper, we analyze all 57 million comments made on user talk pages of 8.5 million editors across the six most active language editions of Wikipedia to study the potential impact of toxicity on editors' behaviour. We find that toxic comments consistently reduce the activity of editors, leading to an estimated loss of 0.5-2 active days per user in the short term. This amounts to multiple human-years of lost productivity when considering the number of active contributors to Wikipedia. The effects of toxic comments are even greater in the long term, as they significantly increase the risk of editors leaving the project altogether. Using an agent-based model, we demonstrate that toxicity attacks on Wikipedia have the potential to impede the progress of the entire project. Our results underscore the importance of mitigating toxic speech on collaborative platforms such as Wikipedia to ensure their continued success.

Elena Chistova, Ivan Smirnov

Coreference resolution is the task of identifying and grouping mentions referring to the same real-world entity. Previous neural models have mainly focused on learning span representations and pairwise scores for coreference decisions. However, current methods do not explicitly capture the referential choice in the hierarchical discourse, an important factor in coreference resolution. In this study, we propose a new approach that incorporates rhetorical information into neural coreference resolution models. We collect rhetorical features from automated discourse parses and examine their impact. As a base model, we implement an end-to-end span-based coreference resolver using a partially fine-tuned multilingual entity-aware language model LUKE. We evaluate our method on the RuCoCo-23 Shared Task for coreference resolution in Russian. Our best model employing rhetorical distance between mentions has ranked 1st on the development set (74.6% F1) and 2nd on the test set (73.3% F1) of the Shared Task. We hope that our work will inspire further research on incorporating discourse information in neural coreference resolution models.

Ivan Smirnov, Segun T. Aroyehun, Paul Plener et al.

The widespread adoption of automatic sentiment and emotion classifiers makes it important to ensure that these tools perform reliably across different populations. Yet their reliability is typically assessed using benchmarks that rely on third-party annotators rather than the individuals experiencing the emotions themselves, potentially concealing systematic biases. In this paper, we use a unique, large-scale dataset of more than one million self-annotated posts and a pre-registered research design to investigate gender biases in emotion detection across 414 combinations of models and emotion-related classes. We find that across different types of automatic classifiers and various underlying emotions, error rates are consistently higher for texts authored by men compared to those authored by women. We quantify how this bias could affect results in downstream applications and show that current machine learning tools, including large language models, should be applied with caution when the gender composition of a sample is not known or variable. Our findings demonstrate that sentiment analysis is not yet a solved problem, especially in ensuring equitable model behaviour across demographic groups.

Ivan Smirnov, Shangding Gu

Reinforcement learning (RL) has seen significant advancements through the application of various neural network architectures. In this study, we systematically investigate the performance of several neural networks in RL tasks, including Long Short-Term Memory (LSTM), Multi-Layer Perceptron (MLP), Mamba/Mamba-2, Transformer-XL, Gated Transformer-XL, and Gated Recurrent Unit (GRU). Through comprehensive evaluation across continuous control, discrete decision-making, and memory-based environments, we identify architecture-specific strengths and limitations. Our results reveal that: (1) MLPs excel in fully observable continuous control tasks, providing an optimal balance of performance and efficiency; (2) recurrent architectures like LSTM and GRU offer robust performance in partially observable environments with moderate memory requirements; (3) Mamba models achieve a 4.5x higher throughput compared to LSTM and a 3.9x increase over GRU, all while maintaining comparable performance; and (4) only Transformer-XL, Gated Transformer-XL, and Mamba-2 successfully solve the most challenging memory-intensive tasks, with Mamba-2 requiring 8x less memory than Transformer-XL. These findings provide insights for researchers and practitioners, enabling more informed architecture selection based on specific task characteristics and computational constraints. Code is available at: https://github.com/SafeRL-Lab/RLBenchNet

Douglas Eacersall, Lynette Pretorius, Ivan Smirnov et al.

The rapid adoption of generative artificial intelligence (GenAI) in research presents both opportunities and ethical challenges that should be carefully navigated. Although GenAI tools can enhance research efficiency through automation of tasks such as literature review and data analysis, their use raises concerns about aspects such as data accuracy, privacy, bias, and research integrity. This paper develops the ETHICAL framework, which is a practical guide for responsible GenAI use in research. Employing a constructivist case study examining multiple GenAI tools in real research contexts, the framework consists of seven key principles: Examine policies and guidelines, Think about social impacts, Harness understanding of the technology, Indicate use, Critically engage with outputs, Access secure versions, and Look at user agreements. Applying these principles will enable researchers to uphold research integrity while leveraging GenAI benefits. The framework addresses a critical gap between awareness of ethical issues and practical action steps, providing researchers with concrete guidance for ethical GenAI integration. This work has implications for research practice, institutional policy development, and the broader academic community while adapting to an AI-enhanced research landscape. The ETHICAL framework can serve as a foundation for developing AI literacy in academic settings and promoting responsible innovation in research methodologies.

Georg Ahnert, Ivan Smirnov, Florian Lemmerich et al.

Measures of algorithmic fairness often do not account for human perceptions of fairness that can substantially vary between different sociodemographics and stakeholders. The FairCeptron framework is an approach for studying perceptions of fairness in algorithmic decision making such as in ranking or classification. It supports (i) studying human perceptions of fairness and (ii) comparing these human perceptions with measures of algorithmic fairness. The framework includes fairness scenario generation, fairness perception elicitation and fairness perception analysis. We demonstrate the FairCeptron framework by applying it to a hypothetical university admission context where we collect human perceptions of fairness in the presence of minorities. An implementation of the FairCeptron framework is openly available, and it can easily be adapted to study perceptions of algorithmic fairness in other application contexts. We hope our work paves the way towards elevating the role of studies of human fairness perceptions in the process of designing algorithmic decision making systems.

Charles Auguste, Sean Malory, Ivan Smirnov

In this report we present two new ways of enforcing monotone constraints in regression and classification trees. One yields better results than the current LightGBM, and has a similar computation time. The other one yields even better results, but is much slower than the current LightGBM. We also propose a heuristic that takes into account that greedily splitting a tree by choosing a monotone split with respect to its immediate gain is far from optimal. Then, we compare the results with the current implementation of the constraints in the LightGBM library, using the well known Adult public dataset. Throughout the report, we mostly focus on the implementation of our methods that we made for the LightGBM library, even though they are general and could be implemented in any regression or classification tree. The best method we propose (a smarter way to split the tree coupled to a penalization of monotone splits) consistently beats the current implementation of LightGBM. With small or average trees, the loss reduction can be as high as 1% in the early stages of training and decreases to around 0.1% at the loss peak for the Adult dataset. The results would be even better with larger trees. In our experiments, we didn't do a lot of tuning of the regularization parameters, and we wouldn't be surprised to see that increasing the performance of our methods on test sets.

Ivan Smirnov, Florian Lemmerich, Markus Strohmaier

Many important decisions in societies such as school admissions, hiring, or elections are based on the selection of top-ranking individuals from a larger pool of candidates. This process is often subject to biases, which typically manifest as an under-representation of certain groups among the selected or accepted individuals. The most common approach to this issue is debiasing, for example via the introduction of quotas that ensure proportional representation of groups with respect to a certain, often binary attribute. Cases include quotas for women on corporate boards or ethnic quotas in elections. This, however, has the potential to induce changes in representation with respect to other attributes. For the case of two correlated binary attributes we show that quota-based debiasing based on a single attribute can worsen the representation of already underrepresented groups and decrease overall fairness of selection. We use several data sets from a broad range of domains from recidivism risk assessments to scientific citations to assess this effect in real-world settings. Our results demonstrate the importance of including all relevant attributes in debiasing procedures and that more efforts need to be put into eliminating the root causes of inequalities as purely numerical solutions such as quota-based debiasing might lead to unintended consequences.

Ivan Smirnov

It has already been established that digital traces can be used to predict various human attributes. In most cases, however, predictive models rely on features that are specific to a particular source of digital trace data. In contrast, short texts written by users $-$ tweets, posts, or comments $-$ are ubiquitous across multiple platforms. In this paper, we explore the predictive power of short texts with respect to the academic performance of their authors. We use data from a representative panel of Russian students that includes information about their educational outcomes and activity on a popular networking site, VK. We build a model to predict academic performance from users' posts on VK and then apply it to a different context. In particular, we show that the model could reproduce rankings of schools and universities from the posts of their students on social media. We also find that the same model could predict academic performance from tweets as well as from VK posts. The generalizability of a model trained on a relatively small data set could be explained by the use of continuous word representations trained on a much larger corpus of social media posts. This also allows for greater interpretability of model predictions.