arXiv subjects

Goran Glavaš

Publications and source records attributed to Goran Glavaš.

At least 19 recordsLinked to original sources

arXiv2026

OctoLong: Mid-Training On Cross-Repository Code Contexts Enhances Long-Context Modeling

Context lengths of language models (LMs) have dramatically increased, driven by the demands for in-context learning, self-improvement, and long-horizon agentic workflows. However, existing long-context corpora are dominated by books, academic articles, and code repositories, which are finite and often lack long-distance dependencies. In this work, we introduce OctoLong, a context-engineering pipeline that instruments an AST parser, a language server backend, and a package manager to recursively retrieve code references, enabling the curation of dependency-rich code contexts up to millions of tokens long. We then train OctoLong-Instruct, a suite of capable long-context open LMs, derived from base models ranging in size from 600M to 14B parameters, via context-extension mid-training on a approx 50B-token mixture containing approx 6.2B tokens of OctoLong code contexts, followed by approx 10B tokens of instruction tuning. Our ablations and experimental evaluations against 18 strong long-context open-weight LMs show that supplanting just 12% of conventional context-extension corpora with OctoLong data yields substantial gains in long-range retrieval, long-term state tracking, repository-level code understanding, and downstream agentic tasks, while also enhancing API usage in regular coding scenarios.

cs.AI↗

arXiv2026

Themis: Training Robust Multilingual Code Reward Models for Flexible Multi-Criteria Scoring

Reward models (RMs) have become an indispensable fixture of the language model (LM) post-training playbook, enabling policy alignment and test-time scaling. Research on the application of RMs in code generation, however, has been comparatively sparse, with existing work largely focusing on execution feedback. This choice constrains post-training to optimizing functional correctness over self-contained executable code. In this work, we examine the training and evaluation of multilingual, multi-criteria code RMs. To this end, we first compile Themis-CodeRewardBench, a benchmark to evaluate code RMs across five preference dimensions (i.e., criteria) and eight programming languages, on which we profile 50+ code, math, and general-purpose RMs. Observing the limited proficiency of current RMs beyond scoring for functional correctness, we develop Themis-CodePreference, the largest open-source collection of code preferences to date (more than 350k preference pairs), and use it to train Themis-RM, a suite of multilingual code reward models for flexible multi-criteria scoring, ranging in size from 600M to 32B parameters. Our experiments and ablations demonstrate positive scaling trends, strong cross-lingual transfer when training on diverse preferences, and the importance of multi-criteria training for reliable code reward modeling.

cs.SE↗

arXiv2026

Compositional Steering of Large Language Models with Steering Tokens

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 steering of verifiable constraints (e.g., length, format, structure, language) 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.

cs.CL↗

arXiv2026

Don't Stop the Multi-Party! On Generating Synthetic Written Multi-Party Conversations with Constraints

Written Multi-Party Conversations (WMPCs) are widely studied across disciplines, with social media as a primary data source due to their accessibility. However, these datasets raise privacy concerns and often reflect platform-specific properties. For example, interactions between speakers may be limited due to rigid platform structures (e.g., threads, tree-like discussions), which yield overly simplistic interaction patterns (e.g., one-to-one "reply-to" links). This work explores the feasibility of generating synthetic WMPCs with instruction-tuned Large Language Models (LLMs) by providing deterministic constraints such as dialogue structure and participants' stance. We investigate two complementary strategies of leveraging LLMs in this context: (i.) LLMs as WMPC generators, where we task the LLM to generate a whole WMPC at once and (ii.) LLMs as WMPC parties, where the LLM generates one turn of the conversation at a time (made of speaker, addressee and message), provided the conversation history. We next introduce an analytical framework to evaluate compliance with the constraints, content quality, and interaction complexity for both strategies. Finally, we assess the level of obtained WMPCs via human and LLM-as-a-judge evaluations. We find stark differences among LLMs, with only some being able to generate high-quality WMPCs. We also find that turn-by-turn generation yields better conformance to constraints and higher linguistic variability than generating WMPCs in one pass. Nonetheless, our structural and qualitative evaluation indicates that both generation strategies can yield high-quality WMPCs.

cs.CL↗

arXiv2026

How Much Do LLMs Hallucinate across Languages? On Realistic Multilingual Estimation of LLM Hallucination

In the age of misinformation, hallucination - the tendency of Large Language Models (LLMs) to generate non-factual or unfaithful responses - represents the main risk for their global utility. Despite LLMs becoming increasingly multilingual, the vast majority of research on detecting and quantifying LLM hallucination are (a) English-centric and (b) focus on machine translation (MT) and summarization, tasks that are less common in realistic settings than open information seeking. In contrast, we aim to quantify the extent of LLM hallucination across languages in knowledge-intensive long-form question answering (LFQA). To this end, we train a multilingual hallucination detection model and conduct a large-scale study across 30 languages and 6 open-source LLM families. We start from an English hallucination detection dataset and rely on MT to translate-train a detection model. We also manually annotate gold data for five high-resource languages; we then demonstrate, for these languages, that the estimates of hallucination rates are similar between silver (LLM-generated) and gold test sets, validating the use of silver data for estimating hallucination rates for other languages. For the final rates estimation, we build open-domain QA dataset for 30 languages with LLM-generated prompts and Wikipedia articles as references. Our analysis shows that LLMs, in absolute terms, hallucinate more tokens in high-resource languages due to longer responses, but that the actual hallucination rates (i.e., normalized for length) seems uncorrelated with the sizes of languages' digital footprints. We also find that smaller LLMs hallucinate more, and significantly, LLMs with broader language support display higher hallucination rates.

cs.CL↗

arXiv2026

The Curious Case of Factual (Mis)Alignment between LLMs' Short- and Long-Form Answers

Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when writing about Einstein's life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs' answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs' trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.

cs.CL↗

arXiv2026

One Script Instead of Hundreds? On Pretraining Romanized Encoder Language Models

Exposing latent lexical overlap, script romanization has emerged as an effective strategy for improving cross-lingual transfer (XLT) in multilingual language models (mLMs). Most prior work, however, focused on setups that favor romanization the most: (1) transfer from high-resource Latin-script to low-resource non-Latin-script languages and/or (2) between genealogically closely related languages with different scripts. It thus remains unclear whether romanization is a good representation choice for pretraining general-purpose mLMs, or, more precisely, if information loss associated with romanization harms performance for high-resource languages. We address this gap by pretraining encoder LMs from scratch on both romanized and original texts for six typologically diverse high-resource languages, investigating two potential sources of degradation: (i) loss of script-specific information and (ii) negative cross-lingual interference from increased vocabulary overlap. Using two romanizers with different fidelity profiles, we observe negligible performance loss for languages with segmental scripts, whereas languages with morphosyllabic scripts (Chinese and Japanese) suffer degradation that higher-fidelity romanization mitigates but cannot fully recover. Importantly, comparing monolingual LMs with their mLM counterpart, we find no evidence that increased subword overlap induces negative interference. We further show that romanization improves encoding efficiency (i.e., fertility) for segmental scripts at a negligible performance cost.

cs.CL↗

arXiv2025

TransAlign: Machine Translation Encoders are Strong Word Aligners, Too

In the absence of sizable training data for most world languages and NLP tasks, translation-based strategies such as translate-test -- evaluating on noisy source language data translated from the target language -- and translate-train -- training on noisy target language data translated from the source language -- have been established as competitive approaches for cross-lingual transfer (XLT). For token classification tasks, these strategies require label projection: mapping the labels from each token in the original sentence to its counterpart(s) in the translation. To this end, it is common to leverage multilingual word aligners (WAs) derived from encoder language models such as mBERT or LaBSE. Despite obvious associations between machine translation (MT) and WA, research on extracting alignments with MT models is largely limited to exploiting cross-attention in encoder-decoder architectures, yielding poor WA results. In this work, in contrast, we propose TransAlign, a novel word aligner that utilizes the encoder of a massively multilingual MT model. We show that TransAlign not only achieves strong WA performance but substantially outperforms popular WA and state-of-the-art non-WA-based label projection methods in MT-based XLT for token classification.

cs.CL↗

arXiv2025

Supervised In-Context Fine-Tuning for Generative Sequence Labeling

Sequence labeling (SL) tasks, where labels are assigned to tokens, are abundant in NLP (e.g., named entity recognition and aspect-based sentiment analysis). Owing to the intuition that they require bidirectional context, SL tasks are commonly tackled with encoder-only models. Recent work also shows that removing the causal mask in fine-tuning enables decoder-based LLMs to become effective token classifiers. Less work, however, focused on (supervised) generative SL, a more natural setting for causal LLMs. Due to their rapid scaling, causal LLMs applied to SL are expected to outperform encoders, whose own development has stagnated. In this work, we propose supervised in-context fine-tuning (SIFT) for generative SL. SIFT casts SL tasks as constrained response generation, natural to LLMs, combining in-context learning (ICL) from demonstrations with supervised fine-tuning. SIFT considerably outperforms both ICL and decoder-as-encoder fine-tuning baselines on a range of standard SL tasks. We further find that although long context hinders the performance of generative SL in both ICL and SIFT, this deficiency can be mitigated by removing the instruction, as instructions are shown to be largely unnecessary for achieving strong SL performance with SIFT. Our findings highlight strengths and limitations of SL with LLMs, underscoring the importance of a response-based generative task formulation for effective SL performance.

cs.CL↗

arXiv2025

ReCoVeR the Target Language: Language Steering without Sacrificing Task Performance

As they become increasingly multilingual, Large Language Models (LLMs) exhibit more language confusion, i.e., they tend to generate answers in a language different from the language of the prompt or the answer language explicitly requested by the user. In this work, we propose ReCoVeR (REducing language COnfusion in VEctor Representations), a novel lightweight approach for reducing language confusion based on language-specific steering vectors. We first isolate language vectors with the help of multi-parallel corpus and then effectively leverage those vectors for effective LLM steering via fixed (i.e., unsupervised) as well as trainable steering functions. Our extensive evaluation, encompassing three benchmarks and 18 languages, shows that ReCoVeR effectively mitigates language confusion in both monolingual and cross-lingual setups while at the same time -- and in contrast to prior language steering methods -- retaining task performance. Our data code is available at https://github.com/hSterz/recover.

cs.CL↗

arXiv2025

Fleurs-SLU: A Massively Multilingual Benchmark for Spoken Language Understanding

Spoken language understanding (SLU) is indispensable for half of all living languages that lack a formal writing system. Unlike for high-resource languages, for these languages, we cannot offload semantic understanding of speech to the cascade of automatic speech recognition (ASR) and text-based large language models (LLMs). Even if low-resource languages possess a writing system, ASR for these languages remains unreliable due to limited bimodal speech and text training data. Nonetheless, the evaluation of multilingual SLU is limited to shallow tasks such as intent classification or language identification. This is why we present Fleurs-SLU, a multilingual SLU benchmark that encompasses (i) 692 hours of speech for topical utterance classification in 102 languages and (ii) multiple-choice question answering via listening comprehension spanning 944 hours of speech across 92 languages. We extensively evaluate end-to-end speech classification models, cascaded systems that combine speech-to-text transcription with subsequent LLM-based classification, and multimodal speech-LLMs on Fleurs-SLU. Our results show that cascaded systems are more robust in multilingual SLU, though well-pretrained speech encoders can perform competitively in topical speech classification. Closed-source speech-LLMs match or surpass the performance of cascaded systems. We observe a strong correlation between robust multilingual ASR, effective speech-to-text translation, and strong multilingual SLU, indicating mutual benefits between acoustic and semantic speech representations.

cs.CL↗

arXiv2025

The Devil Is in the Word Alignment Details: On Translation-Based Cross-Lingual Transfer for Token Classification Tasks

Translation-based strategies for cross-lingual transfer XLT such as translate-train -- training on noisy target language data translated from the source language -- and translate-test -- evaluating on noisy source language data translated from the target language -- are competitive XLT baselines. In XLT for token classification tasks, however, these strategies include label projection, the challenging step of mapping the labels from each token in the original sentence to its counterpart(s) in the translation. Although word aligners (WAs) are commonly used for label projection, the low-level design decisions for applying them to translation-based XLT have not been systematically investigated. Moreover, recent marker-based methods, which project labeled spans by inserting tags around them before (or after) translation, claim to outperform WAs in label projection for XLT. In this work, we revisit WAs for label projection, systematically investigating the effects of low-level design decisions on token-level XLT: (i) the algorithm for projecting labels between (multi-)token spans, (ii) filtering strategies to reduce the number of noisily mapped labels, and (iii) the pre-tokenization of the translated sentences. We find that all of these substantially impact translation-based XLT performance and show that, with optimized choices, XLT with WA offers performance at least comparable to that of marker-based methods. We then introduce a new projection strategy that ensembles translate-train and translate-test predictions and demonstrate that it substantially outperforms the marker-based projection. Crucially, we show that our proposed ensembling also reduces sensitivity to low-level WA design choices, resulting in more robust XLT for token classification tasks.

cs.CL↗

arXiv2025

Problem Solving Through Human-AI Preference-Based Cooperation

While there is a widespread belief that artificial general intelligence (AGI) -- or even superhuman AI -- is imminent, complex problems in expert domains are far from being solved. We argue that such problems require human-AI cooperation and that the current state of the art in generative AI is unable to play the role of a reliable partner due to a multitude of shortcomings, including difficulty to keep track of a complex solution artifact (e.g., a software program), limited support for versatile human preference expression and lack of adapting to human preference in an interactive setting. To address these challenges, we propose HAICo2, a novel human-AI co-construction framework. We take first steps towards a formalization of HAICo2 and discuss the difficult open research problems that it faces.

cs.AI↗

arXiv2025

On Synthesizing Data for Context Attribution in Question Answering

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.

cs.IR↗

arXiv2025

On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures

Measurement systems (e.g., currencies) differ across cultures, but the conversions between them are well defined so that humans can state facts using any measurement system of their choice. Being available to users from diverse cultural backgrounds, large language models (LLMs) should also be able to provide accurate information irrespective of the measurement system at hand. Using newly compiled datasets we test if this is the case for seven open-source LLMs, addressing three key research questions: (RQ1) What is the default system used by LLMs for each type of measurement? (RQ2) Do LLMs' answers and their accuracy vary across different measurement systems? (RQ3) Can LLMs mitigate potential challenges w.r.t. underrepresented systems via reasoning? Our findings show that LLMs default to the measurement system predominantly used in the data. Additionally, we observe considerable instability and variance in performance across different measurement systems. While this instability can in part be mitigated by employing reasoning methods such as chain-of-thought (CoT), this implies longer responses and thereby significantly increases test-time compute (and inference costs), marginalizing users from cultural backgrounds that use underrepresented measurement systems.

cs.CL↗

arXiv2025

Modular Sentence Encoders: Separating Language Specialization from Cross-Lingual Alignment

Multilingual sentence encoders (MSEs) are commonly obtained by training multilingual language models to map sentences from different languages into a shared semantic space. As such, they are subject to curse of multilinguality, a loss of monolingual representational accuracy due to parameter sharing. Another limitation of MSEs is the trade-off between different task performance: cross-lingual alignment training distorts the optimal monolingual structure of semantic spaces of individual languages, harming the utility of sentence embeddings in monolingual tasks; cross-lingual tasks, such as cross-lingual semantic similarity and zero-shot transfer for sentence classification, may also require conflicting cross-lingual alignment strategies. In this work, we address both issues by means of modular training of sentence encoders. We first train language-specific monolingual modules to mitigate negative interference between languages (i.e., the curse). We then align all non-English sentence embeddings to the English by training cross-lingual alignment adapters, preventing interference with monolingual specialization from the first step. We train the cross-lingual adapters with two different types of data to resolve the conflicting requirements of different cross-lingual tasks. Monolingual and cross-lingual results on semantic text similarity and relatedness, bitext mining and sentence classification show that our modular solution achieves better and more balanced performance across all the tasks compared to full-parameter training of monolithic multilingual sentence encoders, especially benefiting low-resource languages.

cs.CL↗

arXiv2025

Large Language Models are Miscalibrated In-Context Learners

When adapting ICL with or without fine-tuning, we are curious about whether the instruction-tuned language model is able to achieve well-calibrated results without suffering from the problem of overconfidence (i.e., miscalibration) considering its strong instruction following ability, especially in such limited data setups. In this work, we deliver an in-depth analysis of the behavior across different choices of learning methods from the perspective of both performance and calibration. Through extensive controlled experiments, we observe that the miscalibration problem exists across all learning methods in low-resource setups. To achieve simultaneous gain for both in-task performance and calibration, we then study the potential of self-ensembling applied at different modeling stages (e.g., variations of in-context examples or variations in prompts or different ensembling strategies) to make the predictions more calibrated and have comparable or even better performance. We find that self-ensembling with max probability produces robust and calibrated predictions. Our work reveals the potential calibration problem of using ICL despite the improvements in task performance and sheds light on which learning paradigm to choose. We also provide practical guidelines for choosing learning paradigms depending on whether the data has been seen by the model before and a worthwhile solution via self-ensembling on how to enhance both task performance and calibration of LMs, which we hope could encourage further study.

cs.CL↗

arXiv2025

On the Potential of Large Language Models to Solve Semantics-Aware Process Mining Tasks

Large language models (LLMs) have shown to be valuable tools for tackling process mining tasks. Existing studies report on their capability to support various data-driven process analyses and even, to some extent, that they are able to reason about how processes work. This reasoning ability suggests that there is potential for LLMs to tackle semantics-aware process mining tasks, which are tasks that rely on an understanding of the meaning of activities and their relationships. Examples of these include process discovery, where the meaning of activities can indicate their dependency, whereas in anomaly detection the meaning can be used to recognize process behavior that is abnormal. In this paper, we systematically explore the capabilities of LLMs for such tasks. Unlike prior work, which largely evaluates LLMs in their default state, we investigate their utility through both in-context learning and supervised fine-tuning. Concretely, we define five process mining tasks requiring semantic understanding and provide extensive benchmarking datasets for evaluation. Our experiments reveal that while LLMs struggle with challenging process mining tasks when used out of the box or with minimal in-context examples, they achieve strong performance when fine-tuned for these tasks across a broad range of process types and industries.

cs.DB↗