arXiv subjects

Hafsteinn Einarsson

Publications and source records attributed to Hafsteinn Einarsson.

11 recordsLinked to original sources

arXiv2026

SAGA: Score-Weighted Adaptive Generation Alignment for Low-Resource Nordic Language Models

Preference optimisation has proven effective for improving large language models but typically relies on costly human preference annotations. Extending these methods to morphologically rich, low-resource languages remains challenging because such annotations are scarce. We present SAGA (Score-weighted Adaptive Generation Alignment), a parser-guided preference optimisation framework that replaces human labels with dependency-parser supervision. SAGA converts parser judgements into preference pairs for delta-DPO, combines parser quality with lexical diversity in a composite reward, filters low-information pairs using a reward-gap criterion, and monitors reward hacking to maintain reliable supervision. Across Danish, Icelandic, and Norwegian Bokmål using GPT-SW3-1.3B, SAGA consistently improves grammatical quality without requiring human preference labels. Danish parse success increases from 69.0% to 93.8%, Icelandic achieves a +4.5 percentage-point improvement on an independent Stanza evaluation (three-run mean +3.3 percentage points) while native speakers prefer SAGA outputs in 80% of pairwise comparisons, and Norwegian Bokmål improves by +28 percentage points. These results demonstrate that parser-derived supervision is a practical alternative to human preference annotation for grammatical alignment in low-resource languages where high-quality dependency parsers are available.

cs.LG↗

arXiv2026

Curated retrieval versus open web search in public AI information services: a coverage-trust trade-off

Public institutions increasingly use large language models (LLMs) to answer citizens' questions, often pairing a curated knowledge base with live web search, yet whether the sources behind these answers can be trusted has received little empirical scrutiny. We report a pre-launch expert evaluation of Evrópuvefur, an independent, government-funded service run by the University of Iceland that answers questions about the European Union, conducted as Iceland prepared for its referendum of 29 August 2026 on whether to resume EU accession talks. Five domain experts produced 551 evaluations of 449 AI-generated answers, scoring each against a seven-criterion quality rubric and, separately, flagging individual cited sources. We compared two retrieval paths: a curated local corpus (RAG) and open web search. In more than a third of the reviewed web-search answers (35%, 65 of 187), at least one cited source was flagged, almost always as untrustworthy or irrelevant; curated sources were flagged far less often and only for being out of date. Web search answered more questions, but at the cost of source quality; the curated corpus was trustworthy yet limited in coverage, and the model declined to respond when it fell short. The citation mix also passed over strong sources: across all 287 web-search answers, the system never cited RÚV, the public broadcaster and the country's most widely used news source. A companion prompt ablation shows how weak prompt-level steering is: a trusted-domain list in the system prompt raised the share of citations to listed domains only from 12% to 21%. Fluency and topical fit did not predict source trustworthiness. We argue that source trustworthiness is a measurable yet largely invisible dimension of information quality in public AI services, and we discuss transparency-oriented responses and their trade-offs.

cs.CY↗

arXiv2025

MazeEval: A Benchmark for Testing Sequential Decision-Making in Language Models

As Large Language Models (LLMs) increasingly power autonomous agents in robotics and embodied AI, understanding their spatial reasoning capabilities becomes crucial for ensuring reliable real-world deployment. Despite advances in language understanding, current research lacks evaluation of how LLMs perform spatial navigation without visual cues, a fundamental requirement for agents operating with limited sensory information. This paper addresses this gap by introducing MazeEval, a benchmark designed to isolate and evaluate pure spatial reasoning in LLMs through coordinate-based maze navigation tasks. Our methodology employs a function-calling interface where models navigate mazes of varying complexity ($5\times 5$ to $15\times 15$ grids) using only coordinate feedback and distance-to-wall information, excluding visual input to test fundamental spatial cognition. We evaluate eight state-of-the-art LLMs across identical mazes in both English and Icelandic to assess cross-linguistic transfer of spatial abilities. Our findings reveal striking disparities: while OpenAI's O3 achieves perfect navigation for mazes up to size $30\times 30$, other models exhibit catastrophic failure beyond $9\times 9$ mazes, with 100% of failures attributed to excessive looping behavior where models revisit a cell at least 10 times. We document a significant performance degradation in Icelandic, with models solving mazes 3-4 sizes smaller than in English, suggesting spatial reasoning in LLMs emerges from linguistic patterns rather than language-agnostic mechanisms. These results have important implications for global deployment of LLM-powered autonomous systems, showing spatial intelligence remains fundamentally constrained by training data availability and highlighting the need for architectural innovations to achieve reliable navigation across linguistic contexts.

cs.AI↗

arXiv2025

Hotter and Colder: A New Approach to Annotating Sentiment, Emotions, and Bias in Icelandic Blog Comments

This paper presents Hotter and Colder, a dataset designed to analyze various types of online behavior in Icelandic blog comments. Building on previous work, we used GPT-4o mini to annotate approximately 800,000 comments for 25 tasks, including sentiment analysis, emotion detection, hate speech, and group generalizations. Each comment was automatically labeled on a 5-point Likert scale. In a second annotation stage, comments with high or low probabilities of containing each examined behavior were subjected to manual revision. By leveraging crowdworkers to refine these automatically labeled comments, we ensure the quality and accuracy of our dataset resulting in 12,232 uniquely annotated comments and 19,301 annotations. Hotter and Colder provides an essential resource for advancing research in content moderation and automatically detectiong harmful online behaviors in Icelandic.

cs.CL↗

arXiv2025

FoQA: A Faroese Question-Answering Dataset

We present FoQA, a Faroese extractive question-answering (QA) dataset with 2,000 samples, created using a semi-automated approach combining Large Language Models (LLMs) and human validation. The dataset was generated from Faroese Wikipedia articles using GPT-4-turbo for initial QA generation, followed by question rephrasing to increase complexity and native speaker validation to ensure quality. We provide baseline performance metrics for FoQA across multiple models, including LLMs and BERT, demonstrating its effectiveness in evaluating Faroese QA performance. The dataset is released in three versions: a validated set of 2,000 samples, a complete set of all 10,001 generated samples, and a set of 2,395 rejected samples for error analysis.

cs.CL↗

arXiv2022

Cross-Lingual QA as a Stepping Stone for Monolingual Open QA in Icelandic

It can be challenging to build effective open question answering (open QA) systems for languages other than English, mainly due to a lack of labeled data for training. We present a data efficient method to bootstrap such a system for languages other than English. Our approach requires only limited QA resources in the given language, along with machine-translated data, and at least a bilingual language model. To evaluate our approach, we build such a system for the Icelandic language and evaluate performance over trivia style datasets. The corpora used for training are English in origin but machine translated into Icelandic. We train a bilingual Icelandic/English language model to embed English context and Icelandic questions following methodology introduced with DensePhrases (Lee et al., 2021). The resulting system is an open domain cross-lingual QA system between Icelandic and English. Finally, the system is adapted for Icelandic only open QA, demonstrating how it is possible to efficiently create an open QA system with limited access to curated datasets in the language of interest.

cs.CL↗

arXiv2022

Building an Icelandic Entity Linking Corpus

In this paper, we present the first Entity Linking corpus for Icelandic. We describe our approach of using a multilingual entity linking model (mGENRE) in combination with Wikipedia API Search (WAPIS) to label our data and compare it to an approach using WAPIS only. We find that our combined method reaches 53.9% coverage on our corpus, compared to 30.9% using only WAPIS. We analyze our results and explain the value of using a multilingual system when working with Icelandic. Additionally, we analyze the data that remain unlabeled, identify patterns and discuss why they may be more difficult to annotate.

cs.CL↗

arXiv2022

A Warm Start and a Clean Crawled Corpus -- A Recipe for Good Language Models

We train several language models for Icelandic, including IceBERT, that achieve state-of-the-art performance in a variety of downstream tasks, including part-of-speech tagging, named entity recognition, grammatical error detection and constituency parsing. To train the models we introduce a new corpus of Icelandic text, the Icelandic Common Crawl Corpus (IC3), a collection of high quality texts found online by targeting the Icelandic top-level-domain (TLD). Several other public data sources are also collected for a total of 16GB of Icelandic text. To enhance the evaluation of model performance and to raise the bar in baselines for Icelandic, we translate and adapt the WinoGrande dataset for co-reference resolution. Through these efforts we demonstrate that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus. We further show that initializing models using existing multilingual models can lead to state-of-the-art results for some downstream tasks.

cs.CL↗

arXiv2018

The linear hidden subset problem for the (1+1) EA with scheduled and adaptive mutation rates

We study unbiased $(1+1)$ evolutionary algorithms on linear functions with an unknown number $n$ of bits with non-zero weight. Static algorithms achieve an optimal runtime of $O(n (\ln n)^{2+ε})$, however, it remained unclear whether more dynamic parameter policies could yield better runtime guarantees. We consider two setups: one where the mutation rate follows a fixed schedule, and one where it may be adapted depending on the history of the run. For the first setup, we give a schedule that achieves a runtime of $(1\pm o(1))βn \ln n$, where $β\approx 3.552$, which is an asymptotic improvement over the runtime of the static setup. Moreover, we show that no schedule admits a better runtime guarantee and that the optimal schedule is essentially unique. For the second setup, we show that the runtime can be further improved to $(1\pm o(1)) e n \ln n$, which matches the performance of algorithms that know $n$ in advance. Finally, we study the related model of initial segment uncertainty with static position-dependent mutation rates, and derive asymptotically optimal lower bounds. This answers a question by Doerr, Doerr, and Kötzing.

cs.NE↗

arXiv2015

Bootstrap percolation with inhibition

Bootstrap percolation is a prominent framework for studying the spreading of activity on a graph. We begin with an initial set of active vertices. The process then proceeds in rounds, and further vertices become active as soon as they have a certain number of active neighbors. A recurring feature in bootstrap percolation theory is an `all-or-nothing' phenomenon: either the size of the starting set is so small that the process stops very soon, or it percolates (almost) completely. Motivated by several important phenomena observed in various types of real-world networks we propose in this work a variant of bootstrap percolation that exhibits a vastly different behavior. Our graphs have two types of vertices: some of them obstruct the diffusion, while the others facilitate it. We study the effect of this setting by analyzing the process on Erdős-Rényi random graphs. Our main findings are two-fold. First we show that the presence of vertices hindering the diffusion does not result in a stable behavior: tiny changes in the size of the starting set can dramatically influence the size of the final active set. In particular, the process is non-monotone: a larger starting set can result in a smaller final set. In the second part of the paper we show that this phenomenom arises from the round-based approach: if we move to a continuous time model in which every edge draws its transmission time randomly, then we gain stability, and the process stops with an active set that contains a non-trivial constant fraction of all vertices. Moreover, we show that in the continuous time model percolation occurs significantly faster compared to the classical round-based model. Our findings are in line with empirical observations and demonstrate the importance of introducing various types of vertex behaviors in the mathematical model.

math.PR↗

arXiv2014

Connectivity Thresholds for Bounded Size Rules

In an Achlioptas process, starting with a graph that has n vertices and no edge, in each round $d \geq 1$ edges are drawn uniformly at random, and using some rule exactly one of them is chosen and added to the evolving graph. For the class of Achlioptas processes we investigate how much impact the rule has on one of the most basic properties of a graph: connectivity. Our main results are twofold. First, we study the prominent class of bounded size rules, which select the edge to add according to the component sizes of its vertices, treating all sizes larger than some constant equally. For such rules we provide a fine analysis that exposes the limiting distribution of the number of rounds until the graph gets connected, and we give a detailed picture of the dynamics of the formation of the single component from smaller components. Second, our results allow us to study the connectivity transition of all Achlioptas processes, in the sense that we identify a process that accelerates it as much as possible.

math.PR↗