Tatiana Shavrina

BigScience Workshop, Teven Le Scao, Angela Fan et al. · allen-ai, berkeley

Large language models (LLMs) have been shown to be able to perform new tasks based on a few demonstrations or natural language instructions. While these capabilities have led to widespread adoption, most LLMs are developed by resource-rich organizations and are frequently kept from the public. As a step towards democratizing this powerful technology, we present BLOOM, a 176B-parameter open-access language model designed and built thanks to a collaboration of hundreds of researchers. BLOOM is a decoder-only Transformer language model that was trained on the ROOTS corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). We find that BLOOM achieves competitive performance on a wide variety of benchmarks, with stronger results after undergoing multitask prompted finetuning. To facilitate future research and applications using LLMs, we publicly release our models and code under the Responsible AI License.

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).

Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster et al.

LLM agents hold significant promise for advancing scientific research. To accelerate this progress, we introduce AIRS-Bench (the AI Research Science Benchmark), a suite of 20 tasks sourced from state-of-the-art machine learning papers. These tasks span diverse domains, including language modeling, mathematics, bioinformatics, and time series forecasting. AIRS-Bench tasks assess agentic capabilities over the full research lifecycle -- including idea generation, experiment analysis and iterative refinement -- without providing baseline code. The AIRS-Bench task format is versatile, enabling easy integration of new tasks and rigorous comparison across different agentic frameworks. We establish baselines using frontier models paired with both sequential and parallel scaffolds. Our results show that agents exceed human SOTA in four tasks but fail to match it in sixteen others. Even when agents surpass human benchmarks, they do not reach the theoretical performance ceiling for the underlying tasks. These findings indicate that AIRS-Bench is far from saturated and offers substantial room for improvement. We open-source the AIRS-Bench task definitions and evaluation code to catalyze further development in autonomous scientific research.

Oleh Shliazhko, Alena Fenogenova, Maria Tikhonova et al.

Recent studies report that autoregressive language models can successfully solve many NLP tasks via zero- and few-shot learning paradigms, which opens up new possibilities for using the pre-trained language models. This paper introduces two autoregressive GPT-like models with 1.3 billion and 13 billion parameters trained on 60 languages from 25 language families using Wikipedia and Colossal Clean Crawled Corpus. We reproduce the GPT-3 architecture using GPT-2 sources and the sparse attention mechanism; Deepspeed and Megatron frameworks allow us to parallelize the training and inference steps effectively. The resulting models show performance on par with the recently released XGLM models by Facebook, covering more languages and enhancing NLP possibilities for low resource languages of CIS countries and Russian small nations. We detail the motivation for the choices of the architecture design, thoroughly describe the data preparation pipeline, and train five small versions of the model to choose the most optimal multilingual tokenization strategy. We measure the model perplexity in all covered languages and evaluate it on the wide spectre of multilingual tasks, including classification, generative, sequence labeling and knowledge probing. The models were evaluated with the zero-shot and few-shot methods. Furthermore, we compared the classification tasks with the state-of-the-art multilingual model XGLM. source code and the mGPT XL model are publicly released.

Dina Pisarevskaya, Tatiana Shavrina

The General QA field has been developing the methodology referencing the Stanford Question answering dataset (SQuAD) as the significant benchmark. However, compiling factual questions is accompanied by time- and labour-consuming annotation, limiting the training data's potential size. We present the WikiOmnia dataset, a new publicly available set of QA-pairs and corresponding Russian Wikipedia article summary sections, composed with a fully automated generative pipeline. The dataset includes every available article from Wikipedia for the Russian language. The WikiOmnia pipeline is available open-source and is also tested for creating SQuAD-formatted QA on other domains, like news texts, fiction, and social media. The resulting dataset includes two parts: raw data on the whole Russian Wikipedia (7,930,873 QA pairs with paragraphs for ruGPT-3 XL and 7,991,040 QA pairs with paragraphs for ruT5-large) and cleaned data with strict automatic verification (over 160,000 QA pairs with paragraphs for ruGPT-3 XL and over 3,400,000 QA pairs with paragraphs for ruT5-large).

Oleg Serikov, Vitaly Protasov, Ekaterina Voloshina et al.

Linguistic analysis of language models is one of the ways to explain and describe their reasoning, weaknesses, and limitations. In the probing part of the model interpretability research, studies concern individual languages as well as individual linguistic structures. The question arises: are the detected regularities linguistically coherent, or on the contrary, do they dissonate at the typological scale? Moreover, the majority of studies address the inherent set of languages and linguistic structures, leaving the actual typological diversity knowledge out of scope. In this paper, we present and apply the GUI-assisted framework allowing us to easily probe a massive number of languages for all the morphosyntactic features present in the Universal Dependencies data. We show that reflecting the anglo-centric trend in NLP over the past years, most of the regularities revealed in the mBERT model are typical for the western-European languages. Our framework can be integrated with the existing probing toolboxes, model cards, and leaderboards, allowing practitioners to use and share their standard probing methods to interpret multilingual models. Thus we propose a toolkit to systematize the multilingual flaws in multilingual models, providing a reproducible experimental setup for 104 languages and 80 morphosyntactic features. https://github.com/AIRI-Institute/Probing_framework

Ekaterina Voloshina, Oleg Serikov, Tatiana Shavrina

The probing methodology allows one to obtain a partial representation of linguistic phenomena stored in the inner layers of the neural network, using external classifiers and statistical analysis. Pre-trained transformer-based language models are widely used both for natural language understanding (NLU) and natural language generation (NLG) tasks making them most commonly used for downstream applications. However, little analysis was carried out, whether the models were pre-trained enough or contained knowledge correlated with linguistic theory. We are presenting the chronological probing study of transformer English models such as MultiBERT and T5. We sequentially compare the information about the language learned by the models in the process of training on corpora. The results show that 1) linguistic information is acquired in the early stages of training 2) both language models demonstrate capabilities to capture various features from various levels of language, including morphology, syntax, and even discourse, while they also can inconsistently fail on tasks that are perceived as easy. We also introduce the open-source framework for chronological probing research, compatible with other transformer-based models. https://github.com/EkaterinaVoloshina/chronological_probing

Dmitry Zmitrovich, Alexander Abramov, Andrey Kalmykov et al.

Transformer language models (LMs) are fundamental to NLP research methodologies and applications in various languages. However, developing such models specifically for the Russian language has received little attention. This paper introduces a collection of 13 Russian Transformer LMs, which spans encoder (ruBERT, ruRoBERTa, ruELECTRA), decoder (ruGPT-3), and encoder-decoder (ruT5, FRED-T5) architectures. We provide a report on the model architecture design and pretraining, and the results of evaluating their generalization abilities on Russian language understanding and generation datasets and benchmarks. By pretraining and releasing these specialized Transformer LMs, we aim to broaden the scope of the NLP research directions and enable the development of industrial solutions for the Russian language.

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.

Bingchen Zhao, Jenny Zhang, Chenxi Whitehouse et al.

Scientific discoveries must be communicated clearly to realize their full potential. Without effective communication, even the most groundbreaking findings risk being overlooked or misunderstood. The primary way scientists communicate their work and receive feedback from the community is through peer review. However, the current system often provides inconsistent feedback between reviewers, ultimately hindering the improvement of a manuscript and limiting its potential impact. In this paper, we introduce a novel method APRES powered by Large Language Models (LLMs) to update a scientific papers text based on an evaluation rubric. Our automated method discovers a rubric that is highly predictive of future citation counts, and integrate it with APRES in an automated system that revises papers to enhance their quality and impact. Crucially, this objective should be met without altering the core scientific content. We demonstrate the success of APRES, which improves future citation prediction by 19.6% in mean averaged error over the next best baseline, and show that our paper revision process yields papers that are preferred over the originals by human expert evaluators 79% of the time. Our findings provide strong empirical support for using LLMs as a tool to help authors stress-test their manuscripts before submission. Ultimately, our work seeks to augment, not replace, the essential role of human expert reviewers, for it should be humans who discern which discoveries truly matter, guiding science toward advancing knowledge and enriching lives.

Karen Hambardzumyan, Nicolas Baldwin, Edan Toledo et al.

Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators imposes a ceiling on search performance. We introduce AIRA$_2$, which addresses these bottlenecks through three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug interactively. On MLE-bench-30, AIRA$_2$ achieves a mean Percentile Rank of 71.8% at 24 hours - surpassing the previous best of 69.9% - and steadily improves to 76.0% at 72 hours. Ablation studies reveal that each component is necessary and that the "overfitting" reported in prior work was driven by evaluation noise rather than true data memorization.

Ekaterina Taktasheva, Tatiana Shavrina, Alena Fenogenova et al.

Recent advances in zero-shot and few-shot learning have shown promise for a scope of research and practical purposes. However, this fast-growing area lacks standardized evaluation suites for non-English languages, hindering progress outside the Anglo-centric paradigm. To address this line of research, we propose TAPE (Text Attack and Perturbation Evaluation), a novel benchmark that includes six more complex NLU tasks for Russian, covering multi-hop reasoning, ethical concepts, logic and commonsense knowledge. The TAPE's design focuses on systematic zero-shot and few-shot NLU evaluation: (i) linguistic-oriented adversarial attacks and perturbations for analyzing robustness, and (ii) subpopulations for nuanced interpretation. The detailed analysis of testing the autoregressive baselines indicates that simple spelling-based perturbations affect the performance the most, while paraphrasing the input has a more negligible effect. At the same time, the results demonstrate a significant gap between the neural and human baselines for most tasks. We publicly release TAPE (tape-benchmark.com) to foster research on robust LMs that can generalize to new tasks when little to no supervision is available.

Deepak Nathani, Lovish Madaan, Nicholas Roberts et al.

We introduce Meta MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. This is the first Gym environment for machine learning (ML) tasks, enabling research on reinforcement learning (RL) algorithms for training such agents. MLGym-bench consists of 13 diverse and open-ended AI research tasks from diverse domains such as computer vision, natural language processing, reinforcement learning, and game theory. Solving these tasks requires real-world AI research skills such as generating new ideas and hypotheses, creating and processing data, implementing ML methods, training models, running experiments, analyzing the results, and iterating through this process to improve on a given task. We evaluate a number of frontier large language models (LLMs) on our benchmarks such as Claude-3.5-Sonnet, Llama-3.1 405B, GPT-4o, o1-preview, and Gemini-1.5 Pro. Our MLGym framework makes it easy to add new tasks, integrate and evaluate models or agents, generate synthetic data at scale, as well as develop new learning algorithms for training agents on AI research tasks. We find that current frontier models can improve on the given baselines, usually by finding better hyperparameters, but do not generate novel hypotheses, algorithms, architectures, or substantial improvements. We open-source our framework and benchmark to facilitate future research in advancing the AI research capabilities of LLM agents.

Jenny Zhang, Bingchen Zhao, Wannan Yang et al.

Self-improving AI systems aim to reduce reliance on human engineering by learning to improve their own learning and problem-solving processes. Existing approaches to self-improvement rely on fixed, handcrafted meta-level mechanisms, fundamentally limiting how fast such systems can improve. The Darwin Gödel Machine (DGM) demonstrates open-ended self-improvement in coding by repeatedly generating and evaluating self-modified variants. Because both evaluation and self-modification are coding tasks, gains in coding ability can translate into gains in self-improvement ability. However, this alignment does not generally hold beyond coding domains. We introduce \textbf{hyperagents}, self-referential agents that integrate a task agent (which solves the target task) and a meta agent (which modifies itself and the task agent) into a single editable program. Crucially, the meta-level modification procedure is itself editable, enabling metacognitive self-modification, improving not only the task-solving behavior, but also the mechanism that generates future improvements. We instantiate this framework by extending DGM to create DGM-Hyperagents (DGM-H), eliminating the assumption of domain-specific alignment between task performance and self-modification skill to potentially support self-accelerating progress on any computable task. Across diverse domains, the DGM-H improves performance over time and outperforms baselines without self-improvement or open-ended exploration, as well as prior self-improving systems. Furthermore, the DGM-H improves the process by which it generates new agents (e.g., persistent memory, performance tracking), and these meta-level improvements transfer across domains and accumulate across runs. DGM-Hyperagents offer a glimpse of open-ended AI systems that do not merely search for better solutions, but continually improve their search for how to improve.

Albert Kornilov, Tatiana Shavrina

Recent advances in language modeling have demonstrated significant improvements in zero-shot capabilities, including in-context learning, instruction following, and machine translation for extremely under-resourced languages (Tanzer et al., 2024). However, many languages with limited written resources rely primarily on formal descriptions of grammar and vocabulary. In this paper, we introduce a set of benchmarks to evaluate how well models can extract and classify information from the complex descriptions found in linguistic grammars. We present a Retrieval-Augmented Generation (RAG)-based approach that leverages these descriptions for downstream tasks such as machine translation. Our benchmarks encompass linguistic descriptions for 248 languages across 142 language families, focusing on typological features from WALS and Grambank. This set of benchmarks offers the first comprehensive evaluation of language models' in-context ability to accurately interpret and extract linguistic features, providing a critical resource for scaling NLP to low-resource languages. The code and data are publicly available at \url{https://github.com/al-the-eigenvalue/RAG-on-grammars}.

Alena Fenogenova, Maria Tikhonova, Vladislav Mikhailov et al.

In the last year, new neural architectures and multilingual pre-trained models have been released for Russian, which led to performance evaluation problems across a range of language understanding tasks. This paper presents Russian SuperGLUE 1.1, an updated benchmark styled after GLUE for Russian NLP models. The new version includes a number of technical, user experience and methodological improvements, including fixes of the benchmark vulnerabilities unresolved in the previous version: novel and improved tests for understanding the meaning of a word in context (RUSSE) along with reading comprehension and common sense reasoning (DaNetQA, RuCoS, MuSeRC). Together with the release of the updated datasets, we improve the benchmark toolkit based on \texttt{jiant} framework for consistent training and evaluation of NLP-models of various architectures which now supports the most recent models for Russian. Finally, we provide the integration of Russian SuperGLUE with a framework for industrial evaluation of the open-source models, MOROCCO (MOdel ResOurCe COmparison), in which the models are evaluated according to the weighted average metric over all tasks, the inference speed, and the occupied amount of RAM. Russian SuperGLUE is publicly available at https://russiansuperglue.com/.

Tatiana Shavrina, Alena Fenogenova, Anton Emelyanov et al.

In this paper, we introduce an advanced Russian general language understanding evaluation benchmark -- RussianGLUE. Recent advances in the field of universal language models and transformers require the development of a methodology for their broad diagnostics and testing for general intellectual skills - detection of natural language inference, commonsense reasoning, ability to perform simple logical operations regardless of text subject or lexicon. For the first time, a benchmark of nine tasks, collected and organized analogically to the SuperGLUE methodology, was developed from scratch for the Russian language. We provide baselines, human level evaluation, an open-source framework for evaluating models (https://github.com/RussianNLP/RussianSuperGLUE), and an overall leaderboard of transformer models for the Russian language. Besides, we present the first results of comparing multilingual models in the adapted diagnostic test set and offer the first steps to further expanding or assessing state-of-the-art models independently of language.

Edan Toledo, Karen Hambardzumyan, Martin Josifoski et al. · meta-ai, oxford

AI research agents are demonstrating great potential to accelerate scientific progress by automating the design, implementation, and training of machine learning models. We focus on methods for improving agents' performance on MLE-bench, a challenging benchmark where agents compete in Kaggle competitions to solve real-world machine learning problems. We formalize AI research agents as search policies that navigate a space of candidate solutions, iteratively modifying them using operators. By designing and systematically varying different operator sets and search policies (Greedy, MCTS, Evolutionary), we show that their interplay is critical for achieving high performance. Our best pairing of search strategy and operator set achieves a state-of-the-art result on MLE-bench lite, increasing the success rate of achieving a Kaggle medal from 39.6% to 47.7%. Our investigation underscores the importance of jointly considering the search strategy, operator design, and evaluation methodology in advancing automated machine learning.

Bingchen Zhao, Despoina Magka, Minqi Jiang et al. · meta-ai, oxford

Rapid advancements in large language models (LLMs) have the potential to assist in scientific progress. A critical capability toward this endeavor is the ability to reproduce existing work. To evaluate the ability of AI agents to reproduce results in an active research area, we introduce the Automated LLM Speedrunning Benchmark, leveraging the research community contributions on the NanoGPT speedrun, a competition to train a GPT-2 model in the shortest time. Each of the 19 speedrun tasks provides the agent with the previous records training script, optionally paired with one of three hint formats, ranging from pseudocode to paper-like descriptions of the new records improvements. Records execute quickly by design and speedrun improvements encompass diverse code-level changes, ranging from high-level algorithmic advancements to hardware-aware optimizations. These features make the benchmark both accessible and realistic for the frontier problem of improving LLM training. We find that recent reasoning LLMs combined with SoTA scaffolds struggle to reimplement already-known innovations in our benchmark, even when given detailed hints. Our benchmark thus provides a simple, non-saturated measure of an LLMs ability to automate scientific reproduction, a necessary (but not sufficient) skill for an autonomous research agent.

Alexis Audran-Reiss, Jordi Armengol Estapé, Karen Hambardzumyan et al.

AI research agents offer the promise to accelerate scientific progress by automating the design, implementation, and training of machine learning models. However, the field is still in its infancy, and the key factors driving the success or failure of agent trajectories are not fully understood. We examine the role that ideation diversity plays in agent performance. First, we analyse agent trajectories on MLE-bench, a well-known benchmark to evaluate AI research agents, across different models and agent scaffolds. Our analysis reveals that different models and agent scaffolds yield varying degrees of ideation diversity, and that higher-performing agents tend to have increased ideation diversity. Further, we run a controlled experiment where we modify the degree of ideation diversity, demonstrating that higher ideation diversity results in stronger performance. Finally, we strengthen our results by examining additional evaluation metrics beyond the standard medal-based scoring of MLE-bench, showing that our findings still hold across other agent performance metrics.

Shalini Maiti, Amar Budhiraja, Bhavul Gauri et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse domains, but their training remains resource- and time-intensive, requiring massive compute power and careful orchestration of training procedures. Model souping-the practice of averaging weights from multiple models of the same architecture-has emerged as a promising pre- and post-training technique that can enhance performance without expensive retraining. In this paper, we introduce Soup Of Category Experts (SoCE), a principled approach for model souping that utilizes benchmark composition to identify optimal model candidates and applies non-uniform weighted averaging to maximize performance. Contrary to previous uniform-averaging approaches, our method leverages the observation that benchmark categories often exhibit low inter-correlations in model performance. SoCE identifies "expert" models for each weakly-correlated category cluster and combines them using optimized weighted averaging rather than uniform weights. We demonstrate that the proposed method improves performance and robustness across multiple domains, including multilingual capabilities, tool calling, and math and achieves state-of-the-art results on the Berkeley Function Calling Leaderboard.

Valentin Malykh, Alexander Kukushkin, Ekaterina Artemova et al.

The new generation of pre-trained NLP models push the SOTA to the new limits, but at the cost of computational resources, to the point that their use in real production environments is often prohibitively expensive. We tackle this problem by evaluating not only the standard quality metrics on downstream tasks but also the memory footprint and inference time. We present MOROCCO, a framework to compare language models compatible with \texttt{jiant} environment which supports over 50 NLU tasks, including SuperGLUE benchmark and multiple probing suites. We demonstrate its applicability for two GLUE-like suites in different languages.

Vladislav Mikhailov, Tatiana Shavrina

Named Entity Recognition (NER) is a fundamental task in the fields of natural language processing and information extraction. NER has been widely used as a standalone tool or an essential component in a variety of applications such as question answering, dialogue assistants and knowledge graphs development. However, training reliable NER models requires a large amount of labelled data which is expensive to obtain, particularly in specialized domains. This paper describes a method to learn a domain-specific NER model for an arbitrary set of named entities when domain-specific supervision is not available. We assume that the supervision can be obtained with no human effort, and neural models can learn from each other. The code, data and models are publicly available.

Taisia Glushkova, Alexey Machnev, Alena Fenogenova et al.

DaNetQA, a new question-answering corpus, follows (Clark et. al, 2019) design: it comprises natural yes/no questions. Each question is paired with a paragraph from Wikipedia and an answer, derived from the paragraph. The task is to take both the question and a paragraph as input and come up with a yes/no answer, i.e. to produce a binary output. In this paper, we present a reproducible approach to DaNetQA creation and investigate transfer learning methods for task and language transferring. For task transferring we leverage three similar sentence modelling tasks: 1) a corpus of paraphrases, Paraphraser, 2) an NLI task, for which we use the Russian part of XNLI, 3) another question answering task, SberQUAD. For language transferring we use English to Russian translation together with multilingual language fine-tuning.

Maria Ponomareva, Kira Droganova, Ivan Smurov et al.

This paper provides a comprehensive overview of the gapping dataset for Russian that consists of 7.5k sentences with gapping (as well as 15k relevant negative sentences) and comprises data from various genres: news, fiction, social media and technical texts. The dataset was prepared for the Automatic Gapping Resolution Shared Task for Russian (AGRR-2019) - a competition aimed at stimulating the development of NLP tools and methods for processing of ellipsis. In this paper, we pay special attention to the gapping resolution methods that were introduced within the shared task as well as an alternative test set that illustrates that our corpus is a diverse and representative subset of Russian language gapping sufficient for effective utilization of machine learning techniques.