Kiril Gashteovski

Vijay Viswanathan, Kiril Gashteovski, Carolin Lawrence et al. · cmu

Unlike traditional unsupervised clustering, semi-supervised clustering allows users to provide meaningful structure to the data, which helps the clustering algorithm to match the user's intent. Existing approaches to semi-supervised clustering require a significant amount of feedback from an expert to improve the clusters. In this paper, we ask whether a large language model can amplify an expert's guidance to enable query-efficient, few-shot semi-supervised text clustering. We show that LLMs are surprisingly effective at improving clustering. We explore three stages where LLMs can be incorporated into clustering: before clustering (improving input features), during clustering (by providing constraints to the clusterer), and after clustering (using LLMs post-correction). We find incorporating LLMs in the first two stages can routinely provide significant improvements in cluster quality, and that LLMs enable a user to make trade-offs between cost and accuracy to produce desired clusters. We release our code and LLM prompts for the public to use.

Gorjan Radevski, Kiril Gashteovski, Chia-Chien Hung et al.

Open Information Extraction (OIE) methods extract facts from natural language text in the form of ("subject"; "relation"; "object") triples. These facts are, however, merely surface forms, the ambiguity of which impedes their downstream usage; e.g., the surface phrase "Michael Jordan" may refer to either the former basketball player or the university professor. Knowledge Graphs (KGs), on the other hand, contain facts in a canonical (i.e., unambiguous) form, but their coverage is limited by a static schema (i.e., a fixed set of entities and predicates). To bridge this gap, we need the best of both worlds: (i) high coverage of free-text OIEs, and (ii) semantic precision (i.e., monosemy) of KGs. In order to achieve this goal, we propose a new benchmark with novel evaluation protocols that can, for example, measure fact linking performance on a granular triple slot level, while also measuring if a system has the ability to recognize that a surface form has no match in the existing KG. Our extensive evaluation of several baselines show that detection of out-of-KG entities and predicates is more difficult than accurate linking to existing ones, thus calling for more research efforts on this difficult task. We publicly release all resources (data, benchmark and code) on https://github.com/nec-research/fact-linking.

Haris Widjaja, Kiril Gashteovski, Wiem Ben Rim et al.

Knowledge Graphs (KGs) store information in the form of (head, predicate, tail)-triples. To augment KGs with new knowledge, researchers proposed models for KG Completion (KGC) tasks such as link prediction; i.e., answering (h; p; ?) or (?; p; t) queries. Such models are usually evaluated with averaged metrics on a held-out test set. While useful for tracking progress, averaged single-score metrics cannot reveal what exactly a model has learned -- or failed to learn. To address this issue, we propose KGxBoard: an interactive framework for performing fine-grained evaluation on meaningful subsets of the data, each of which tests individual and interpretable capabilities of a KGC model. In our experiments, we highlight the findings that we discovered with the use of KGxBoard, which would have been impossible to detect with standard averaged single-score metrics.

Bhushan Kotnis, Kiril Gashteovski, Julia Gastinger et al.

With Human-Centric Research (HCR) we can steer research activities so that the research outcome is beneficial for human stakeholders, such as end users. But what exactly makes research human-centric? We address this question by providing a working definition and define how a research pipeline can be split into different stages in which human-centric components can be added. Additionally, we discuss existing NLP with HCR components and define a series of guiding questions, which can serve as starting points for researchers interested in exploring human-centric research approaches. We hope that this work would inspire researchers to refine the proposed definition and to pose other questions that might be meaningful for achieving HCR.

Zhivar Sourati, Darshan Deshpande, Filip Ilievski et al.

Downstream applications often require text classification models to be accurate and robust. While the accuracy of the state-of-the-art Language Models (LMs) approximates human performance, they often exhibit a drop in performance on noisy data found in the real world. This lack of robustness can be concerning, as even small perturbations in the text, irrelevant to the target task, can cause classifiers to incorrectly change their predictions. A potential solution can be the family of Prototype-Based Networks (PBNs) that classifies examples based on their similarity to prototypical examples of a class (prototypes) and has been shown to be robust to noise for computer vision tasks. In this paper, we study whether the robustness properties of PBNs transfer to text classification tasks under both targeted and static adversarial attack settings. Our results show that PBNs, as a mere architectural variation of vanilla LMs, offer more robustness compared to vanilla LMs under both targeted and static settings. We showcase how PBNs' interpretability can help us to understand PBNs' robustness properties. Finally, our ablation studies reveal the sensitivity of PBNs' robustness to how strictly clustering is done in the training phase, as tighter clustering results in less robust PBNs.

Sascha Saralajew, Ammar Shaker, Zhao Xu et al.

With the rise of AI systems in real-world applications comes the need for reliable and trustworthy AI. An essential aspect of this are explainable AI systems. However, there is no agreed standard on how explainable AI systems should be assessed. Inspired by the Turing test, we introduce a human-centric assessment framework where a leading domain expert accepts or rejects the solutions of an AI system and another domain expert. By comparing the acceptance rates of provided solutions, we can assess how the AI system performs compared to the domain expert, and whether the AI system's explanations (if provided) are human-understandable. This setup -- comparable to the Turing test -- can serve as a framework for a wide range of human-centric AI system assessments. We demonstrate this by presenting two instantiations: (1) an assessment that measures the classification accuracy of a system with the option to incorporate label uncertainties; (2) an assessment where the usefulness of provided explanations is determined in a human-centric manner.

Luca Gioacchini, Giuseppe Siracusano, Davide Sanvito et al.

The advances made by Large Language Models (LLMs) have led to the pursuit of LLM agents that can solve intricate, multi-step reasoning tasks. As with any research pursuit, benchmarking and evaluation are key corner stones to efficient and reliable progress. However, existing benchmarks are often narrow and simply compute overall task success. To face these issues, we propose AgentQuest -- a framework where (i) both benchmarks and metrics are modular and easily extensible through well documented and easy-to-use APIs; (ii) we offer two new evaluation metrics that can reliably track LLM agent progress while solving a task. We exemplify the utility of the metrics on two use cases wherein we identify common failure points and refine the agent architecture to obtain a significant performance increase. Together with the research community, we hope to extend AgentQuest further and therefore we make it available under https://github.com/nec-research/agentquest.

Seungone Kim, Dongkeun Yoon, Kiril Gashteovski et al.

With the advancement of AI capabilities, AI reviewers are beginning to be deployed in scientific peer review, yet their capability and credibility remain in question: many scientists simply view them as probabilistic systems without the expertise to evaluate research, while other researchers are more optimistic about their readiness without concrete evidence. Understanding what AI reviewers do well, where they fall short, and what challenges remain is essential. However, existing evaluations of AI reviewers have focused on whether their verdicts match human verdicts (e.g., score alignment, acceptance prediction), which is insufficient to characterize their capabilities and limits. In this paper, we close this gap through a large-scale expert annotation study, in which 45 domain scientists in Physical, Biological, and Health Sciences spent 469 hours rating 2,960 individual criticisms (each targeting one specific aspect of a paper) from human-written and AI-generated reviews of 82 Nature-family papers on correctness, significance, and sufficiency of evidence. On a composite of all three dimensions, a reviewing agent powered by GPT-5.2 scores above each paper's top-rated human reviewer (60.0% vs. 48.2%, p = 0.009), while all three AI reviewers (including Gemini 3.0 Pro and Claude Opus 4.5) exceed the lowest-rated human across every dimension. AI reviewers' accurate criticisms are also more often rated significant and well-evidenced, and surface a distinct 26% of issues no human raises. However, AI reviewers overlap far more than humans do (21% vs. 3% for cross-reviewer pairs), and exhibit 16 recurring weaknesses humans do not share, such as limited subfield knowledge, lack of long context management over multiple files, and overly critical stance on minor issues. Overall, our results position current AI reviewers as complements to, not substitutes for, human reviewers.

Emmy Liu, Amanda Bertsch, Lintang Sutawika et al. · cmu

Improvements in language model capabilities are often attributed to increasing model size or training data, but in some cases smaller models trained on curated data or with different architectural decisions can outperform larger ones trained on more tokens. What accounts for this? To quantify the impact of these design choices, we meta-analyze 92 open-source pretrained models across a wide array of scales, including state-of-the-art open-weights models as well as less performant models and those with less conventional design decisions. We find that by incorporating features besides model size and number of training tokens, we can achieve a relative 3-28% increase in ability to predict downstream performance compared with using scale alone. Analysis of model design decisions reveal insights into data composition, such as the trade-off between language and code tasks at 15-25\% code, as well as the better performance of some architectural decisions such as choosing rotary over learned embeddings. Broadly, our framework lays a foundation for more systematic investigation of how model development choices shape final capabilities.

Gorjan Radevski, Kiril Gashteovski, Giwon Hong et al.

Deploying LLMs in real-world applications requires controllable output that satisfies multiple desiderata at the same time. While existing work extensively addresses LLM steering for a single behavior, \textit{compositional steering} -- i.e., steering LLMs simultaneously towards multiple behaviors -- remains an underexplored problem. In this work, we propose \emph{compositional steering tokens} for multi-behavior steering. We first embed individual behaviors, expressed as natural language instructions, into dedicated tokens via self-distillation. Contrary to most prior work, which operates in the activation space, our behavior steers live in the space of input tokens, enabling more effective zero-shot composition. We then train a dedicated \textit{composition token} on pairs of behaviors and show that it successfully captures the notion of composition: it generalizes well to \textit{unseen} compositions, including those with unseen behaviors as well as those with an unseen \textit{number} of behaviors. Our experiments across different LLM architectures show that steering tokens lead to superior multi-behavior control compared to competing approaches (instructions, activation steering, and LoRA merging). Moreover, we show that steering tokens complement natural language instructions, with their combination resulting in further gains.

Kiril Gashteovski, Mingying Yu, Bhushan Kotnis et al.

Intrinsic evaluations of OIE systems are carried out either manually -- with human evaluators judging the correctness of extractions -- or automatically, on standardized benchmarks. The latter, while much more cost-effective, is less reliable, primarily because of the incompleteness of the existing OIE benchmarks: the ground truth extractions do not include all acceptable variants of the same fact, leading to unreliable assessment of the models' performance. Moreover, the existing OIE benchmarks are available for English only. In this work, we introduce BenchIE: a benchmark and evaluation framework for comprehensive evaluation of OIE systems for English, Chinese, and German. In contrast to existing OIE benchmarks, BenchIE is fact-based, i.e., it takes into account informational equivalence of extractions: our gold standard consists of fact synsets, clusters in which we exhaustively list all acceptable surface forms of the same fact. Moreover, having in mind common downstream applications for OIE, we make BenchIE multi-faceted; i.e., we create benchmark variants that focus on different facets of OIE evaluation, e.g., compactness or minimality of extractions. We benchmark several state-of-the-art OIE systems using BenchIE and demonstrate that these systems are significantly less effective than indicated by existing OIE benchmarks. We make BenchIE (data and evaluation code) publicly available on https://github.com/gkiril/benchie.

Seungone Kim, Juyoung Suk, Xiang Yue et al. · cmu

Given the increasing use of synthetic data in language model (LM) post-training, an LM's ability to generate high-quality data has become nearly as crucial as its ability to solve problems directly. While prior works have focused on developing effective data generation methods, they lack systematic comparison of different LMs as data generators in a unified setting. To address this gap, we propose AgoraBench, a benchmark that provides standardized settings and metrics to evaluate LMs' data generation abilities. Through synthesizing 1.26 million training instances using 6 LMs and training 99 student models, we uncover key insights about LMs' data generation capabilities. First, we observe that LMs exhibit distinct strengths. For instance, GPT-4o excels at generating new problems, while Claude-3.5-Sonnet performs better at enhancing existing ones. Furthermore, our analysis reveals that an LM's data generation ability doesn't necessarily correlate with its problem-solving ability. Instead, multiple intrinsic features of data quality-including response quality, perplexity, and instruction difficulty-collectively serve as better indicators. Finally, we demonstrate that strategic choices in output format and cost-conscious model selection significantly impact data generation effectiveness.

Seungone Kim, Ian Wu, Jinu Lee et al. · cmu

As language model (LM) outputs get more and more natural, it is becoming more difficult than ever to evaluate their quality. Simultaneously, increasing LMs' "thinking" time through scaling test-time compute has proven an effective technique to solve challenging problems in domains such as math and code. This raises a natural question: can an LM's evaluation capability also be improved by spending more test-time compute? To answer this, we investigate employing reasoning models-LMs that natively generate long chain-of-thought reasoning-as evaluators. Specifically, we examine methods to leverage more test-time compute by (1) using reasoning models, and (2) prompting these models to evaluate not only the response as a whole (i.e., outcome evaluation) but also assess each step in the response separately (i.e., process evaluation). In experiments, we observe that the evaluator's performance improves monotonically when generating more reasoning tokens, similar to the trends observed in LM-based generation. Furthermore, we use these more accurate evaluators to rerank multiple generations, and demonstrate that spending more compute at evaluation time can be as effective as using more compute at generation time in improving an LM's problem-solving capability.

Daniel Rose, Chia-Chien Hung, Marco Lepri et al.

Differential Diagnosis (DDx) is a fundamental yet complex aspect of clinical decision-making, in which physicians iteratively refine a ranked list of possible diseases based on symptoms, antecedents, and medical knowledge. While recent advances in large language models (LLMs) have shown promise in supporting DDx, existing approaches face key limitations, including single-dataset evaluations, isolated optimization of components, unrealistic assumptions about complete patient profiles, and single-attempt diagnosis. We introduce a Modular Explainable DDx Agent (MEDDxAgent) framework designed for interactive DDx, where diagnostic reasoning evolves through iterative learning, rather than assuming a complete patient profile is accessible. MEDDxAgent integrates three modular components: (1) an orchestrator (DDxDriver), (2) a history taking simulator, and (3) two specialized agents for knowledge retrieval and diagnosis strategy. To ensure robust evaluation, we introduce a comprehensive DDx benchmark covering respiratory, skin, and rare diseases. We analyze single-turn diagnostic approaches and demonstrate the importance of iterative refinement when patient profiles are not available at the outset. Our broad evaluation demonstrates that MEDDxAgent achieves over 10% accuracy improvements in interactive DDx across both large and small LLMs, while offering critical explainability into its diagnostic reasoning process.

Filip Ilievski, Barbara Hammer, Frank van Harmelen et al.

Recent advances in AI -- including generative approaches -- have resulted in technology that can support humans in scientific discovery and forming decisions, but may also disrupt democracies and target individuals. The responsible use of AI and its participation in human-AI teams increasingly shows the need for AI alignment, that is, to make AI systems act according to our preferences. A crucial yet often overlooked aspect of these interactions is the different ways in which humans and machines generalise. In cognitive science, human generalisation commonly involves abstraction and concept learning. In contrast, AI generalisation encompasses out-of-domain generalisation in machine learning, rule-based reasoning in symbolic AI, and abstraction in neurosymbolic AI. In this perspective paper, we combine insights from AI and cognitive science to identify key commonalities and differences across three dimensions: notions of, methods for, and evaluation of generalisation. We map the different conceptualisations of generalisation in AI and cognitive science along these three dimensions and consider their role for alignment in human-AI teaming. This results in interdisciplinary challenges across AI and cognitive science that must be tackled to provide a foundation for effective and cognitively supported alignment in human-AI teaming scenarios.

Chenyue Zhou, Gürkan Solmaz, Flavio Cirillo et al.

Humanitarian Mine Action (HMA) addresses the challenge of detecting and removing landmines from conflict regions. Much of the life-saving operational knowledge produced by HMA agencies is buried in unstructured reports, limiting the transferability of information between agencies. To address this issue, we propose TextMine: the first dataset, evaluation framework and ontology-guided large language model (LLM) pipeline for knowledge extraction in the HMA domain. TextMine structures HMA reports into (subject, relation, object)-triples, thus creating domain-specific knowledge. To ensure real-world relevance, we created the dataset in collaboration with Cambodian Mine Action Center (CMAC). We further introduce a bias-aware evaluation framework that combines human-annotated triples with an LLM-as-Judge protocol to mitigate position bias in reference-free scoring. Our experiments show that ontology-aligned prompts improve extraction accuracy by up to 44.2%, reduce hallucinations by 22.5%, and enhance format adherence by 20.9% compared to baseline models. We publicly release the dataset and code.

Gorjan Radevski, Kiril Gashteovski, Shahbaz Syed et al.

Question Answering (QA) accounts for a significant portion of LLM usage "in the wild". However, LLMs sometimes produce false or misleading responses, also known as "hallucinations". Therefore, grounding the generated answers in contextually provided information -- i.e., providing evidence for the generated text -- is paramount for LLMs' trustworthiness. Providing this information is the task of context attribution. In this paper, we systematically study LLM-based approaches for this task, namely we investigate (i) zero-shot inference, (ii) LLM ensembling, and (iii) fine-tuning of small LMs on synthetic data generated by larger LLMs. Our key contribution is SynQA: a novel generative strategy for synthesizing context attribution data. Given selected context sentences, an LLM generates QA pairs that are supported by these sentences. This leverages LLMs' natural strengths in text generation while ensuring clear attribution paths in the synthetic training data. We show that the attribution data synthesized via SynQA is highly effective for fine-tuning small LMs for context attribution in different QA tasks and domains. Finally, with a user study, we validate the usefulness of small LMs (fine-tuned on synthetic data from SynQA) in context attribution for QA.

Masafumi Enomoto, Kunihiro Takeoka, Kosuke Akimoto et al.

Open-Domain Multi-Document Summarization (ODMDS) is the task of generating summaries from large document collections in response to user queries. This task is crucial for efficiently addressing diverse information needs from users. Traditional retrieve-then-summarize approaches fall short for open-ended queries in ODMDS tasks. These queries often require broader context than initially retrieved passages provide, making it challenging to retrieve all relevant information in a single search. While iterative retrieval methods has been explored for multi-hop question answering (MQA), it's impractical for ODMDS due to high latency from repeated LLM inference. Accordingly, we propose LightPAL, a lightweight passage retrieval method for ODMDS. LightPAL leverages an LLM to pre-construct a graph representing passage relationships, then employs random walk during retrieval, avoiding iterative LLM inference. Experiments demonstrate that LightPAL outperforms naive sparse and pre-trained dense retrievers in both retrieval and summarization metrics, while achieving higher efficiency compared to iterative MQA approaches.

David Dukić, Kiril Gashteovski, Goran Glavaš et al.

Event detection is a crucial information extraction task in many domains, such as Wikipedia or news. The task typically relies on trigger detection (TD) -- identifying token spans in the text that evoke specific events. While the notion of triggers should ideally be universal across domains, domain transfer for TD from high- to low-resource domains results in significant performance drops. We address the problem of negative transfer in TD by coupling triggers between domains using subject-object relations obtained from a rule-based open information extraction (OIE) system. We demonstrate that OIE relations injected through multi-task training can act as mediators between triggers in different domains, enhancing zero- and few-shot TD domain transfer and reducing performance drops, in particular when transferring from a high-resource source domain (Wikipedia) to a low(er)-resource target domain (news). Additionally, we combine this improved transfer with masked language modeling on the target domain, observing further TD transfer gains. Finally, we demonstrate that the gains are robust to the choice of the OIE system.

Bhushan Kotnis, Kiril Gashteovski, Daniel Oñoro Rubio et al.

Open Information Extraction (OpenIE) is the task of extracting (subject, predicate, object) triples from natural language sentences. Current OpenIE systems extract all triple slots independently. In contrast, we explore the hypothesis that it may be beneficial to extract triple slots iteratively: first extract easy slots, followed by the difficult ones by conditioning on the easy slots, and therefore achieve a better overall extraction. Based on this hypothesis, we propose a neural OpenIE system, milIE, that operates in an iterative fashion. Due to the iterative nature, the system is also modular -- it is possible to seamlessly integrate rule based extraction systems with a neural end-to-end system, thereby allowing rule based systems to supply extraction slots which milIE can leverage for extracting the remaining slots. We confirm our hypothesis empirically: milIE outperforms SOTA systems on multiple languages ranging from Chinese to Arabic. Additionally, we are the first to provide an OpenIE test dataset for Arabic and Galician.

Niklas Friedrich, Kiril Gashteovski, Mingying Yu et al.

Open Information Extraction (OIE) is the task of extracting facts from sentences in the form of relations and their corresponding arguments in schema-free manner. Intrinsic performance of OIE systems is difficult to measure due to the incompleteness of existing OIE benchmarks: the ground truth extractions do not group all acceptable surface realizations of the same fact that can be extracted from a sentence. To measure performance of OIE systems more realistically, it is necessary to manually annotate complete facts (i.e., clusters of all acceptable surface realizations of the same fact) from input sentences. We propose AnnIE: an interactive annotation platform that facilitates such challenging annotation tasks and supports creation of complete fact-oriented OIE evaluation benchmarks. AnnIE is modular and flexible in order to support different use case scenarios (i.e., benchmarks covering different types of facts). We use AnnIE to build two complete OIE benchmarks: one with verb-mediated facts and another with facts encompassing named entities. Finally, we evaluate several OIE systems on our complete benchmarks created with AnnIE. Our results suggest that existing incomplete benchmarks are overly lenient, and that OIE systems are not as robust as previously reported. We publicly release AnnIE under non-restrictive license.

Kiril Gashteovski, Sebastian Wanner, Sven Hertling et al.

Open information extraction (OIE) systems extract relations and their arguments from natural language text in an unsupervised manner. The resulting extractions are a valuable resource for downstream tasks such as knowledge base construction, open question answering, or event schema induction. In this paper, we release, describe, and analyze an OIE corpus called OPIEC, which was extracted from the text of English Wikipedia. OPIEC complements the available OIE resources: It is the largest OIE corpus publicly available to date (over 340M triples) and contains valuable metadata such as provenance information, confidence scores, linguistic annotations, and semantic annotations including spatial and temporal information. We analyze the OPIEC corpus by comparing its content with knowledge bases such as DBpedia or YAGO, which are also based on Wikipedia. We found that most of the facts between entities present in OPIEC cannot be found in DBpedia and/or YAGO, that OIE facts often differ in the level of specificity compared to knowledge base facts, and that OIE open relations are generally highly polysemous. We believe that the OPIEC corpus is a valuable resource for future research on automated knowledge base construction.