Arrow Research search

Author name cluster

Iryna Gurevych

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

18 papers
2 author rows

Possible papers

18

TMLR Journal 2026 Journal Article

Hypothesis-Driven Feature Manifold Analysis in LLMs via Supervised Multi-Dimensional Scaling

  • Federico Tiblias
  • Irina Bigoulaeva
  • Jingcheng Niu
  • Simone Balloccu
  • Iryna Gurevych

The linear representation hypothesis states that language models (LMs) encode concepts as directions in their latent space, forming organized, multidimensional manifolds. Prior work has largely focused on identifying specific geometries for individual features, limiting its ability to generalize. We introduce Supervised Multi-Dimensional Scaling (SMDS), a model-agnostic method for evaluating and comparing competing feature manifold hypotheses. We apply SMDS to temporal reasoning as a case study and find that different features instantiate distinct geometric structures, including circles, lines, and clusters. SMDS reveals several consistent characteristics of these structures: they reflect the semantic properties of the concepts they represent, remain stable across model families and sizes, actively support reasoning, and dynamically reshape in response to contextual changes. Together, our findings shed light on the functional role of feature manifolds, supporting a model of entity-based reasoning in which LMs encode and transform structured representations.

PDF Details

AAAI Conference 2026 Conference Paper

SPARE: Single-Pass Annotation with Reference-Guided Evaluation for Automatic Process Supervision and Reward Modelling

  • Md Imbesat Hassan Rizvi
  • Xiaodan Zhu
  • Iryna Gurevych

Process or step-wise supervision has played a crucial role in advancing complex multi-step reasoning capabilities of Large Language Models (LLMs). However, efficient, high-quality automated process annotation remains a significant challenge. To address this, we introduce Single-Pass Annotation with Reference-Guided Evaluation (SPARE), a novel structured framework that enables efficient per-step annotation by jointly aligning solution steps to reference solutions and determine its accuracy with explicit reasoning in single generation. We demonstrate SPARE's effectiveness across four diverse datasets spanning mathematical reasoning (GSM8K, MATH), multi-hop question answering (MuSiQue-Ans), and spatial reasoning (SpaRP), showing consistent improvements in two applications: (1) training Process Reward Models (PRMs) for ranking and aggregating multiple generations, and (2) fine-tuning models via offline reinforcement learning for greedy decoding. On PROCESSBENCH, SPARE demonstrates data-efficient out-of-distribution generalization, using only ~16% of training samples compared to human-labeled and other synthetically trained baselines. Additionally, it achieves competitive performance with MCTS-based methods while offering 2.3x speedup in terms of total token count. Manual analysis reveals complementary precision-recall characteristics with MCTS approaches, suggesting potential for ensemble methods. These results establish SPARE as a practical and scalable solution for automatic process supervision in LLM reasoning.

PDF Details DOI

ICLR Conference 2025 Conference Paper

Differentially Private Steering for Large Language Model Alignment

  • Anmol Goel
  • Yaxi Hu
  • Iryna Gurevych
  • Amartya Sanyal

Aligning Large Language Models (LLMs) with human values and away from undesirable behaviors (such as hallucination) has become increasingly important. Recently, steering LLMs towards a desired behavior via activation editing has emerged as an effective method to mitigate harmful generations at inference-time. Activation editing modifies LLM representations by preserving information from positive demonstrations (e.g., truthful) and minimising information from negative demonstrations (e.g., hallucinations). When these demonstrations come from a private dataset, the aligned LLM may leak private information contained in those private samples. In this work, we present the first study of aligning LLM behavior with private datasets. Our work proposes the \textit{\underline{P}rivate \underline{S}teering for LLM \underline{A}lignment (PSA)} algorithm to edit LLM activations with differential privacy (DP) guarantees. We conduct extensive experiments on seven different benchmarks with open-source LLMs of different sizes (0.5B to 7B) and model families (LlaMa and Qwen). Our results show that PSA achieves DP guarantees for LLM alignment with minimal loss in performance, including alignment metrics, open-ended text generation quality, and general-purpose reasoning. We also develop the first Membership Inference Attack (MIA) for evaluating and auditing the empirical privacy for the problem of LLM steering via activation editing. Our attack is tailored for activation editing and relies solely on the generated texts without their associated probabilities. Our experiments support the theoretical guarantees by showing improved guarantees for our \textit{PSA} algorithm compared to several existing non-private techniques.

Details

TMLR Journal 2025 Journal Article

Illusion or Algorithm? Investigating Memorization, Emergence, and Symbolic Processing in In-Context Learning

  • Jingcheng Niu
  • Subhabrata Dutta
  • Ahmed Elshabrawy
  • Harish Tayyar Madabushi
  • Iryna Gurevych

Large-scale Transformer language models (LMs) trained solely on next-token prediction with web-scale data can solve a wide range of tasks after seeing just a few examples. The mechanism behind this capability, known as in-context learning (ICL), remains both controversial and poorly understood. Some studies argue that it is merely the result of memorizing vast amounts of data, while others contend that it reflects a fundamental, symbolic algorithmic development in LMs. In this work, we introduce a suite of investigative tasks and a novel method to systematically investigate ICL by leveraging the full Pythia scaling suite, including interim checkpoints that capture progressively larger amount of training data. By carefully exploring ICL performance on downstream tasks and simultaneously conducting a mechanistic analysis of the residual stream's subspace, we demonstrate that ICL extends beyond mere "memorization" of the training corpus, yet does not amount to the implementation of an independent symbolic algorithm. Our results also clarify several aspects of ICL, including the influence of training dynamics, model capabilities, and elements of mechanistic interpretability. Overall, our work advances the understanding of ICL and its implications, offering model developers insights into potential improvements and providing AI security practitioners with a basis for more informed guidelines.

PDF Details

ICLR Conference 2025 Conference Paper

ObscuraCoder: Powering Efficient Code LM Pre-Training Via Obfuscation Grounding

  • Indraneil Paul
  • Haoyi Yang
  • Goran Glavas
  • Kristian Kersting
  • Iryna Gurevych

Language models (LMs) have become a staple of the code-writing toolbox. Their pre-training recipe has, however, remained stagnant over recent years, barring the occasional changes in data sourcing and filtering strategies. In particular, research exploring modifications to Code-LMs' pre-training objectives, geared towards improving data efficiency and better disentangling between syntax and semantics, has been noticeably sparse, especially compared with corresponding efforts in natural language LMs. In this work, we examine grounding on obfuscated code as a means of helping Code-LMs look beyond the surface-form syntax and enhance their pre-training sample efficiency. To this end, we compile ObscuraX, a dataset of approximately 55M source and obfuscated code pairs in seven languages. Subsequently, we pre-train ObscuraCoder models, ranging in size from 255M to 2.8B parameters, on a 272B-token corpus that includes ObscuraX and demonstrate that our obfuscation-based pre-training recipe leads to consistent improvements in Code-LMs' abilities compared to both vanilla autoregressive pre-training as well as existing de-obfuscation (DOBF) objectives. ObscuraCoder demonstrates sizeable gains across multiple tests of syntactic and semantic code understanding, along with improved capabilities in multilingual code completion, multilingual code commit summarization, and multi-purpose library-oriented code generation.

Details

ICLR Conference 2025 Conference Paper

Uncertainty-Aware Decoding with Minimum Bayes Risk

  • Nico Daheim
  • Clara Meister
  • Thomas Möllenhoff
  • Iryna Gurevych

Despite their outstanding performance in the majority of scenarios, contemporary language models still occasionally generate undesirable outputs, for example, hallucinated text. While such behaviors have previously been linked to uncertainty, there is a notable lack of methods that actively consider uncertainty during text generation. In this work, we show how Minimum Bayes Risk (MBR) decoding, which selects model generations according to an expected risk, can be generalized into a principled uncertainty-aware decoding method. In short, we account for model uncertainty during decoding by incorporating a posterior over model parameters into MBR’s computation of expected risk. We show that this modified expected risk is useful for both choosing outputs and deciding when to abstain from generation and can provide improvements without incurring overhead. We benchmark different methods for learning posteriors and show that performance improves with prediction diversity. We release our code publicly.

Details

ICLR Conference 2024 Conference Paper

Model Merging by Uncertainty-Based Gradient Matching

  • Nico Daheim
  • Thomas Möllenhoff
  • Edoardo M. Ponti
  • Iryna Gurevych
  • Mohammad Emtiyaz Khan

Models trained on different datasets can be merged by a weighted-averaging of their parameters, but why does it work and when can it fail? Here, we connect the inaccuracy of weighted-averaging to mismatches in the gradients and propose a new uncertainty-based scheme to improve the performance by reducing the mismatch. The connection also reveals implicit assumptions in other schemes such as averaging, task arithmetic, and Fisher-weighted averaging. Our new method gives consistent improvements for large language models and vision transformers, both in terms of performance and robustness to hyperparameters.

Details

ICML Conference 2024 Conference Paper

Variational Learning is Effective for Large Deep Networks

  • Yuesong Shen
  • Nico Daheim
  • Bai Cong
  • Peter Nickl
  • Gian Maria Marconi
  • Clement Bazan
  • Rio Yokota
  • Iryna Gurevych

We give extensive empirical evidence against the common belief that variational learning is ineffective for large neural networks. We show that an optimizer called Improved Variational Online Newton (IVON) consistently matches or outperforms Adam for training large networks such as GPT-2 and ResNets from scratch. IVON’s computational costs are nearly identical to Adam but its predictive uncertainty is better. We show several new use cases of IVON where we improve finetuning and model merging in Large Language Models, accurately predict generalization error, and faithfully estimate sensitivity to data. We find overwhelming evidence that variational learning is effective. Code is available at https: //github. com/team-approx-bayes/ivon.

Details

AILAW Journal 2023 Journal Article

Mining legal arguments in court decisions

  • Ivan Habernal
  • Daniel Faber
  • Nicola Recchia
  • Sebastian Bretthauer
  • Iryna Gurevych
  • Indra Spiecker genannt Döhmann
  • Christoph Burchard

Abstract Identifying, classifying, and analyzing arguments in legal discourse has been a prominent area of research since the inception of the argument mining field. However, there has been a major discrepancy between the way natural language processing (NLP) researchers model and annotate arguments in court decisions and the way legal experts understand and analyze legal argumentation. While computational approaches typically simplify arguments into generic premises and claims, arguments in legal research usually exhibit a rich typology that is important for gaining insights into the particular case and applications of law in general. We address this problem and make several substantial contributions to move the field forward. First, we design a new annotation scheme for legal arguments in proceedings of the European Court of Human Rights (ECHR) that is deeply rooted in the theory and practice of legal argumentation research. Second, we compile and annotate a large corpus of 373 court decisions (2.3M tokens and 15k annotated argument spans). Finally, we train an argument mining model that outperforms state-of-the-art models in the legal NLP domain and provide a thorough expert-based evaluation. All datasets and source codes are available under open lincenses at https://github.com/trusthlt/mining-legal-arguments.

Details DOI

NeurIPS Conference 2021 Conference Paper

BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models

  • Nandan Thakur
  • Nils Reimers
  • Andreas Rücklé
  • Abhishek Srivastava
  • Iryna Gurevych

Existing neural information retrieval (IR) models have often been studied in homogeneous and narrow settings, which has considerably limited insights into their out-of-distribution (OOD) generalization capabilities. To address this, and to facilitate researchers to broadly evaluate the effectiveness of their models, we introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains and evaluate 10 state-of-the-art retrieval systems including lexical, sparse, dense, late-interaction, and re-ranking architectures on the BEIR benchmark. Our results show BM25 is a robust baseline and re-ranking and late-interaction based models on average achieve the best zero-shot performances, however, at high computational costs. In contrast, dense and sparse-retrieval models are computationally more efficient but often underperform other approaches, highlighting the considerable room for improvement in their generalization capabilities. We hope this framework allows us to better evaluate and understand existing retrieval systems and contributes to accelerating progress towards more robust and generalizable systems in the future. BEIR is publicly available at https: //github. com/UKPLab/beir.

PDF Details

NeurIPS Conference 2021 Conference Paper

SciGen: a Dataset for Reasoning-Aware Text Generation from Scientific Tables

  • Nafise Moosavi
  • Andreas Rücklé
  • Dan Roth
  • Iryna Gurevych

We introduce SciGen, a new challenge dataset consisting of tables from scientific articles and their corresponding descriptions, for the task of reasoning-aware data-to-text generation. Describing scientific tables goes beyond the surface realization of the table content and requires reasoning over table values. The unique properties of SciGen are that (1) tables mostly contain numerical values, and (2) the corresponding descriptions require arithmetic reasoning. SciGen is the first dataset that assesses the arithmetic reasoning capabilities of generation models on complex input structures, such as tables from scientific articles, and thus it opens new avenues for future research in reasoning-aware text generation and evaluation. The core part of SciGen, including the test data, is annotated by one of the authors of the corresponding articles. Such expert annotations do not scale to large training data sizes. To tackle this, we propose a pipeline for automatically extracting high-quality table-description pairs from the LaTeX sources of scientific articles. We study the effectiveness of state-of-the-art data-to-text generation models on SciGen and evaluate the results using common metrics and human evaluation. Our results and analyses show that adding high-quality unsupervised training data improves the correctness and reduces the hallucination in generated descriptions, however, the ability of state-of-the-art models is still severely limited on this task.

PDF Details

AAAI Conference 2020 Conference Paper

Fine-Grained Argument Unit Recognition and Classification

  • Dietrich Trautmann
  • Johannes Daxenberger
  • Christian Stab
  • Hinrich Schütze
  • Iryna Gurevych

Prior work has commonly defined argument retrieval from heterogeneous document collections as a sentence-level classification task. Consequently, argument retrieval suffers both from low recall and from sentence segmentation errors making it difficult for humans and machines to consume the arguments. In this work, we argue that the task should be performed on a more fine-grained level of sequence labeling. For this, we define the task as Argument Unit Recognition and Classification (AURC). We present a dataset of arguments from heterogeneous sources annotated as spans of tokens within a sentence, as well as with a corresponding stance. We show that and how such difficult argument annotations can be effectively collected through crowdsourcing with high interannotator agreement. The new benchmark, AURC-8, contains up to 15% more arguments per topic as compared to annotations on the sentence level. We identify a number of methods targeted at AURC sequence labeling, achieving close to human performance on known domains. Further analysis also reveals that, contrary to previous approaches, our methods are more robust against sentence segmentation errors. We publicly release our code and the AURC-8 dataset. 1

PDF Details

ICLR Conference 2020 Conference Paper

Latent Normalizing Flows for Many-to-Many Cross-Domain Mappings

  • Shweta Mahajan
  • Iryna Gurevych
  • Stefan Roth 0001

Learned joint representations of images and text form the backbone of several important cross-domain tasks such as image captioning. Prior work mostly maps both domains into a common latent representation in a purely supervised fashion. This is rather restrictive, however, as the two domains follow distinct generative processes. Therefore, we propose a novel semi-supervised framework, which models shared information between domains and domain-specific information separately. The information shared between the domains is aligned with an invertible neural network. Our model integrates normalizing flow-based priors for the domain-specific information, which allows us to learn diverse many-to-many mappings between the two domains. We demonstrate the effectiveness of our model on diverse tasks, including image captioning and text-to-image synthesis.

Details

AAAI Conference 2020 Conference Paper

Low Resource Sequence Tagging with Weak Labels

  • Edwin Simpson
  • Jonas Pfeiffer
  • Iryna Gurevych

Current methods for sequence tagging depend on large quantities of domain-specific training data, limiting their use in new, user-defined tasks with few or no annotations. While crowdsourcing can be a cheap source of labels, it often introduces errors that degrade the performance of models trained on such crowdsourced data. Another solution is to use transfer learning to tackle low resource sequence labelling, but current approaches rely heavily on similar high resource datasets in different languages. In this paper, we propose a domain adaptation method using Bayesian sequence combination to exploit pre-trained models and unreliable crowdsourced data that does not require high resource data in a different language. Our method boosts performance by learning the relationship between each labeller and the target task and trains a sequence labeller on the target domain with little or no goldstandard data. We apply our approach to labelling diagnostic classes in medical and educational case studies, showing that the model achieves strong performance though zero-shot transfer learning and is more effective than alternative ensemble methods. Using NER and information extraction tasks, we show how our approach can train a model directly from crowdsourced labels, outperforming pipeline approaches that first aggregate the crowdsourced data, then train on the aggregated labels.

PDF Details

AAAI Conference 2020 Conference Paper

Two Birds with One Stone: Investigating Invertible Neural Networks for Inverse Problems in Morphology

  • Gözde Gül Şahin
  • Iryna Gurevych

Most problems in natural language processing can be approximated as inverse problems such as analysis and generation at variety of levels from morphological (e. g. , cat+Plural↔cats) to semantic (e. g. , (call + 1 2)↔“Calculate one plus two. ”). Although the tasks in both directions are closely related, general approach in the field has been to design separate models specific for each task. However, having one shared model for both tasks, would help the researchers exploit the common knowledge among these problems with reduced time and memory requirements. We investigate a specific class of neural networks, called Invertible Neural Networks (INNs) (Ardizzone et al. 2019) that enable simultaneous optimization in both directions, hence allow addressing of inverse problems via a single model. In this study, we investigate INNs on morphological problems casted as inverse problems. We apply INNs to various morphological tasks with varying ambiguity and show that they provide competitive performance in both directions. We show that they are able to recover the morphological input parameters, i. e. , predicting the lemma (e. g. , cat) or the morphological tags (e. g. , Plural) when run in the reverse direction, without any significant performance drop in the forward direction, i. e. , predicting the surface form (e. g. , cats).

PDF Details

AAAI Conference 2019 Conference Paper

Challenges in the Automatic Analysis of Students’ Diagnostic Reasoning

  • Claudia Schulz
  • Christian M. Meyer
  • Iryna Gurevych

Diagnostic reasoning is a key component of many professions. To improve students’ diagnostic reasoning skills, educational psychologists analyse and give feedback on epistemic activities used by these students while diagnosing, in particular, hypothesis generation, evidence generation, evidence evaluation, and drawing conclusions. However, this manual analysis is highly time-consuming. We aim to enable the large-scale adoption of diagnostic reasoning analysis and feedback by automating the epistemic activity identification. We create the first corpus for this task, comprising diagnostic reasoning selfexplanations of students from two domains annotated with epistemic activities. Based on insights from the corpus creation and the task’s characteristics, we discuss three challenges for the automatic identification of epistemic activities using AI methods: the correct identification of epistemic activity spans, the reliable distinction of similar epistemic activities, and the detection of overlapping epistemic activities. We propose a separate performance metric for each challenge and thus provide an evaluation framework for future research. Indeed, our evaluation of various state-of-the-art recurrent neural network architectures reveals that current techniques fail to address some of these challenges.

PDF Details

AAAI Conference 2019 Conference Paper

COALA: A Neural Coverage-Based Approach for Long Answer Selection with Small Data

  • Andreas Rücklé
  • Nafise Sadat Moosavi
  • Iryna Gurevych

Current neural network based community question answering (cQA) systems fall short of (1) properly handling long answers which are common in cQA; (2) performing under small data conditions, where a large amount of training data is unavailable—i. e. , for some domains in English and even more so for a huge number of datasets in other languages; and (3) benefiting from syntactic information in the model—e. g. , to differentiate between identical lexemes with different syntactic roles. In this paper, we propose COALA, an answer selection approach that (a) selects appropriate long answers due to an effective comparison of all question-answer aspects, (b) has the ability to generalize from a small number of training examples, and (c) makes use of the information about syntactic roles of words. We show that our approach outperforms existing answer selection models by a large margin on six cQA datasets from different domains. Furthermore, we report the best results on the passage retrieval benchmark WikiPassageQA.

PDF Details

IJCAI Conference 2019 Conference Paper

Reward Learning for Efficient Reinforcement Learning in Extractive Document Summarisation

  • Yang Gao
  • Christian M. Meyer
  • Mohsen Mesgar
  • Iryna Gurevych

Document summarisation can be formulated as a sequential decision-making problem, which can be solved by Reinforcement Learning (RL) algorithms. The predominant RL paradigm for summarisation learns a cross-input policy, which requires considerable time, data and parameter tuning due to the huge search spaces and the delayed rewards. Learning input-specific RL policies is a more efficient alternative, but so far depends on handcrafted rewards, which are difficult to design and yield poor performance. We propose RELIS, a novel RL paradigm that learns a reward function with Learning-to-Rank (L2R) algorithms at training time and uses this reward function to train an input-specific RL policy at test time. We prove that RELIS guarantees to generate near-optimal summaries with appropriate L2R and RL algorithms. Empirically, we evaluate our approach on extractive multi-document summarisation. We show that RELIS reduces the training time by two orders of magnitude compared to the state-of-the-art models while performing on par with them.

PDF Details