arXiv subjects

Patrick Haller

Publications and source records attributed to Patrick Haller.

At least 19 recordsLinked to original sources

arXiv2026

Repetition over Diversity: High-Signal Data Filtering for Sample-Efficient German Language Modeling

Recent research has shown that filtering massive English web corpora into high-quality subsets significantly improves training efficiency. However, for high-resource non-English languages like German, French, or Japanese, aggressive filtering creates a strategic dilemma: should practitioners prioritize diversity by training once on large amounts of lightly filtered web data, or prioritize quality by strictly filtering for a high-quality core and repeating it over multiple epochs? We investigate this trade-off for German by constructing hierarchical quality filters applied to 500M web documents, comparing multi-epoch training on the filtered subsets against single-pass training on a diverse corpus. Our experiments across multiple model scales and token budgets show that repeating high-quality data consistently outperforms single-pass training on larger, less filtered sets. Notably, the performance gap persists even after 7 epochs. Our findings suggest that for non-English LLMs, semantic concentration through quality filtering offers a more viable path to efficient language modeling than simply maximizing unique data volume. We release our German language models (called Boldt), as well as our cleaned evaluation benchmarks to the research community. Our experiments indicate that they achieve state-of-the-art results despite training on 10-360x fewer tokens than comparable models.

cs.CL↗

arXiv2026

What Matters When Building Universal Multilingual Named Entity Recognition Models?

Recent progress in universal multilingual named entity recognition (NER) has been driven by multilingual transformer models, task-specific architectures, custom loss functions, and large-scale training datasets. However, despite substantial prior work, we find that many design decisions for such models are made without systematic justification, with individual components evaluated only in combination rather than in isolation. We argue that this impedes progress in the field by making it difficult to identify which choices improve multilingual generalization. In this work, we conduct an extensive empirical evaluation on transformer backbones, architectures, training objectives, data composition, and threshold selection for zero-shot, multilingual NER models. Building on these findings, we present Otter, a small encoder model that achieves consistent improvements over strong multilingual NER baselines, outperforming similarly sized models by 5.7 percentage points in F1. Further, it remains within 1.4 points of Qwen3-32B and 5.4 points below Gemma3-27B, while being roughly 90x smaller and one to two orders of magnitude faster at inference.

cs.CL↗

arXiv2025

FiNERweb: Datasets and Artifacts for Scalable Multilingual Named Entity Recognition

Recent multilingual named entity recognition (NER) work has shown that large language models (LLMs) can provide effective synthetic supervision, yet such datasets have mostly appeared as by-products of broader experiments rather than as systematic, reusable resources. We introduce FiNERweb, a dataset-creation pipeline that scales the teacher-student paradigm to 91 languages and 25 scripts. Building on FineWeb-Edu, our approach trains regression models to identify NER-relevant passages and annotates them with multilingual LLMs, resulting in about 225k passages with 235k distinct entity labels. Our experiments show that the regression model achieves more than 84 F1, and that models trained on FiNERweb obtain comparable or improved performance in zero shot transfer settings on English, Thai, and Swahili, despite being trained on 19x less data than strong baselines. In addition, we assess annotation quality using LLM-as-a-judge and observe consistently high scores for both faithfulness (3.99 out of 5) and completeness (4.05 out of 5), indicating reliable and informative annotations. Further, we release the dataset with both English labels and translated label sets in the respective target languages because we observe that the performance of current state-of-the-art models drops by 0.02 to 0.09 F1 when evaluated using target language labels instead of English ones. We release FiNERweb together with all accompanying artifacts to the research community in order to facilitate more effective student-teacher training for multilingual named entity recognition.

cs.CL↗

arXiv2025

Sample-Efficient Language Modeling with Linear Attention and Lightweight Enhancements

We study architectural and optimization techniques for sample-efficient language modeling under the constraints of the BabyLM 2025 shared task. Our model, BLaLM, replaces self-attention with a linear-time mLSTM token mixer and explores lightweight enhancements, including short convolutions, sliding window attention with dynamic modulation, and Hedgehog feature maps. To support training in low-resource settings, we curate a high-quality corpus emphasizing readability and pedagogical structure. Experiments across both STRICT and STRICT-SMALL tracks show that (1) linear attention combined with sliding window attention consistently improves zero-shot performance, and (2) the Muon optimizer stabilizes convergence and reduces perplexity over AdamW. These results highlight effective strategies for efficient language modeling without relying on scale.

cs.CL↗

arXiv2025

PISA-Bench: The PISA Index as a Multilingual and Multimodal Metric for the Evaluation of Vision-Language Models

Vision-language models (VLMs) have demonstrated remarkable progress in multimodal reasoning. However, existing benchmarks remain limited in terms of high-quality, human-verified examples. Many current datasets rely on synthetically generated content by large language models (LLMs). Furthermore, most datasets are limited to English, as manual quality assurance of translated samples is time-consuming and costly. To fill this gap, we introduce PISA-Bench, a multilingual benchmark derived from English examples of the expert-created PISA tests, a unified framework for the assessment of student competencies in over eighty countries. Each example consists of human-extracted instructions, questions, answer options, and images, enriched with question type categories, and has been translated from English into five additional languages (Spanish, German, Chinese, French, and Italian), resulting in a fully parallel corpus covering six languages. We evaluate state-of-the-art vision-language models on PISA-Bench and find that especially small models (<20B parameters) fail to achieve high test scores. We further find substantial performance degradation on non-English splits as well as high error-rates when models are tasked with spatial and geometric reasoning. By releasing the dataset and evaluation framework, we provide a resource for advancing research on multilingual multimodal reasoning.

cs.CV↗

arXiv2025

LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance

For Large Language Models (LLMs) to be reliable, they must learn robust knowledge that can be generally applied in diverse settings -- often unlike those seen during training. Yet, extensive research has shown that LLM performance can be brittle, with models exhibiting excessive sensitivity to trivial input variations. In this work, we explore whether this brittleness is a direct result of unstable internal knowledge representations. To explore this question, we build on previous work showing that LLM representations encode statement truthfulness -- i.e., true, factual statements can be easily separated from false, inaccurate ones. Specifically, we test the robustness of learned knowledge by evaluating representation separability on samples that have undergone superficial transformations to drive them out-of-distribution (OOD), such as typos or reformulations. By applying semantically-preserving perturbations, we study how separability degrades as statements become more OOD, across four LLM families, five evaluation datasets, and three knowledge probing methods. Our results reveal that internal representations of statement truthfulness collapse as the samples' presentations become less similar to those seen during pre-training. While LLMs can often distinguish between true and false statements when they closely resemble the pre-training data, this ability is highly dependent on the statement's exact surface form. These findings offer a possible explanation for brittle benchmark performance: LLMs may learn shallow, non-robust knowledge representations that allow for only limited generalizability. Our work presents a fundamental challenge for the utility of truthfulness probes, and more broadly, calls for further research on improving the robustness of learned knowledge representations.

cs.CL↗

arXiv2025

Leveraging In-Context Learning for Political Bias Testing of LLMs

A growing body of work has been querying LLMs with political questions to evaluate their potential biases. However, this probing method has limited stability, making comparisons between models unreliable. In this paper, we argue that LLMs need more context. We propose a new probing task, Questionnaire Modeling (QM), that uses human survey data as in-context examples. We show that QM improves the stability of question-based bias evaluation, and demonstrate that it may be used to compare instruction-tuned models to their base versions. Experiments with LLMs of various sizes indicate that instruction tuning can indeed change the direction of bias. Furthermore, we observe a trend that larger models are able to leverage in-context examples more effectively, and generally exhibit smaller bias scores in QM. Data and code are publicly available.

cs.CL↗

arXiv2025

From Data to Knowledge: Evaluating How Efficiently Language Models Learn Facts

Sample efficiency is a crucial property of language models with practical implications for training efficiency. In real-world text, information follows a long-tailed distribution. Yet, we expect models to learn and recall frequent and infrequent facts. Sample-efficient models are better equipped to handle this challenge of learning and retaining rare information without requiring excessive exposure. This study analyzes multiple models of varying architectures and sizes, all trained on the same pre-training data. By annotating relational facts with their frequencies in the training corpus, we examine how model performance varies with fact frequency. Our findings show that most models perform similarly on high-frequency facts but differ notably on low-frequency facts. This analysis provides new insights into the relationship between model architecture, size, and factual learning efficiency.

cs.CL↗

arXiv2025

What Matters in Linearizing Language Models? A Comparative Study of Architecture, Scale, and Task Adaptation

Linearization has emerged as a strategy for developing efficient language models (LMs). Starting from an existing Transformer-based LM, linearization replaces the attention component with computationally efficient subquadratic \textit{token mixers}. However, as an increasing number of mixers are proposed, it remains unclear which inductive biases are best suited to inherit the original Transformer's capabilities. Furthermore, it is unknown how linearization is affected by parameter and token budget scaling. To address these questions, we propose a unified setup to compare seven representative architectures, including xLSTM, GLA, and Gated DeltaNet. Our findings reveal that performance hierarchies remain stable from 140M to 1.7B parameters, with error-correcting update rules demonstrating superior scaling exponents. We show that performance gaps are established early and persist through asymptotic maturity at 10B tokens, suggesting that state resolution is a more fundamental bottleneck than the distillation budget. Finally, while most models adapt to instruction tuning, only gated delta-rule formulations maintain the precision necessary for long-context retrieval, whereas additive models suffer from irreversible state saturation. These results suggest that for successful linearization, architectural inductive biases remain the primary constraint that cannot be overcome by simply scaling training compute.

cs.CL↗

arXiv2025

MastermindEval: A Simple But Scalable Reasoning Benchmark

Recent advancements in large language models (LLMs) have led to remarkable performance across a wide range of language understanding and mathematical tasks. As a result, increasing attention has been given to assessing the true reasoning capabilities of LLMs, driving research into commonsense, numerical, logical, and qualitative reasoning. However, with the rapid progress of reasoning-focused models such as OpenAI's o1 and DeepSeek's R1, there has been a growing demand for reasoning benchmarks that can keep pace with ongoing model developments. In this paper, we introduce MastermindEval, a simple, scalable, and interpretable deductive reasoning benchmark inspired by the board game Mastermind. Our benchmark supports two evaluation paradigms: (1) agentic evaluation, in which the model autonomously plays the game, and (2) deductive reasoning evaluation, in which the model is given a pre-played game state with only one possible valid code to infer. In our experimental results we (1) find that even easy Mastermind instances are difficult for current models and (2) demonstrate that the benchmark is scalable to possibly more advanced models in the future Furthermore, we investigate possible reasons why models cannot deduce the final solution and find that current models are limited in deducing the concealed code as the number of statement to combine information from is increasing.

cs.CL↗

arXiv2024

BabyHGRN: Exploring RNNs for Sample-Efficient Training of Language Models

This paper explores the potential of recurrent neural networks (RNNs) and other subquadratic architectures as competitive alternatives to transformer-based models in low-resource language modeling scenarios. We utilize HGRN2 (Qin et al., 2024), a recently proposed RNN-based architecture, and comparatively evaluate its effectiveness against transformer-based baselines and other subquadratic architectures (LSTM, xLSTM, Mamba). Our experimental results show that BABYHGRN, our HGRN2 language model, outperforms transformer-based models in both the 10M and 100M word tracks of the challenge, as measured by their performance on the BLiMP, EWoK, GLUE and BEAR benchmarks. Further, we show the positive impact of knowledge distillation. Our findings challenge the prevailing focus on transformer architectures and indicate the viability of RNN-based models, particularly in resource-constrained environments.

cs.CL↗

arXiv2024

Familiarity: Better Evaluation of Zero-Shot Named Entity Recognition by Quantifying Label Shifts in Synthetic Training Data

Zero-shot named entity recognition (NER) is the task of detecting named entities of specific types (such as 'Person' or 'Medicine') without any training examples. Current research increasingly relies on large synthetic datasets, automatically generated to cover tens of thousands of distinct entity types, to train zero-shot NER models. However, in this paper, we find that these synthetic datasets often contain entity types that are semantically highly similar to (or even the same as) those in standard evaluation benchmarks. Because of this overlap, we argue that reported F1 scores for zero-shot NER overestimate the true capabilities of these approaches. Further, we argue that current evaluation setups provide an incomplete picture of zero-shot abilities since they do not quantify the label shift (i.e., the similarity of labels) between training and evaluation datasets. To address these issues, we propose Familiarity, a novel metric that captures both the semantic similarity between entity types in training and evaluation, as well as their frequency in the training data, to provide an estimate of label shift. It allows researchers to contextualize reported zero-shot NER scores when using custom synthetic training datasets. Further, it enables researchers to generate evaluation setups of various transfer difficulties for fine-grained analysis of zero-shot NER.

cs.CL↗

arXiv2024

EMTeC: A Corpus of Eye Movements on Machine-Generated Texts

The Eye Movements on Machine-Generated Texts Corpus (EMTeC) is a naturalistic eye-movements-while-reading corpus of 107 native English speakers reading machine-generated texts. The texts are generated by three large language models using five different decoding strategies, and they fall into six different text type categories. EMTeC entails the eye movement data at all stages of pre-processing, i.e., the raw coordinate data sampled at 2000 Hz, the fixation sequences, and the reading measures. It further provides both the original and a corrected version of the fixation sequences, accounting for vertical calibration drift. Moreover, the corpus includes the language models' internals that underlie the generation of the stimulus texts: the transition scores, the attention scores, and the hidden states. The stimuli are annotated for a range of linguistic features both at text and at word level. We anticipate EMTeC to be utilized for a variety of use cases such as, but not restricted to, the investigation of reading behavior on machine-generated text and the impact of different decoding strategies; reading behavior on different text types; the development of new pre-processing, data filtering, and drift correction algorithms; the cognitive interpretability and enhancement of language models; and the assessment of the predictive power of surprisal and entropy for human reading times. The data at all stages of pre-processing, the model internals, and the code to reproduce the stimulus generation, data pre-processing and analyses can be accessed via https://github.com/DiLi-Lab/EMTeC/.

cs.CL↗

arXiv2024

Language models emulate certain cognitive profiles: An investigation of how predictability measures interact with individual differences

To date, most investigations on surprisal and entropy effects in reading have been conducted on the group level, disregarding individual differences. In this work, we revisit the predictive power of surprisal and entropy measures estimated from a range of language models (LMs) on data of human reading times as a measure of processing effort by incorporating information of language users' cognitive capacities. To do so, we assess the predictive power of surprisal and entropy estimated from generative LMs on reading data obtained from individuals who also completed a wide range of psychometric tests. Specifically, we investigate if modulating surprisal and entropy relative to cognitive scores increases prediction accuracy of reading times, and we examine whether LMs exhibit systematic biases in the prediction of reading times for cognitively high- or low-performing groups, revealing what type of psycholinguistic subject a given LM emulates. Our study finds that in most cases, incorporating cognitive capacities increases predictive power of surprisal and entropy on reading times, and that generally, high performance in the psychometric tests is associated with lower sensitivity to predictability effects. Finally, our results suggest that the analyzed LMs emulate readers with lower verbal intelligence, suggesting that for a given target group (i.e., individuals with high verbal intelligence), these LMs provide less accurate predictability estimates.

cs.CL↗

arXiv2024

PECC: Problem Extraction and Coding Challenges

Recent advancements in large language models (LLMs) have showcased their exceptional abilities across various tasks, such as code generation, problem-solving and reasoning. Existing benchmarks evaluate tasks in isolation, yet the extent to which LLMs can understand prose-style tasks, identify the underlying problems, and then generate appropriate code solutions is still unexplored. Addressing this gap, we introduce PECC, a novel benchmark derived from Advent Of Code (AoC) challenges and Project Euler, including 2396 problems. Unlike conventional benchmarks, PECC requires LLMs to interpret narrative-embedded problems, extract requirements, and generate executable code. A key feature of our dataset is the complexity added by natural language prompting in chat-based evaluations, mirroring real-world instruction ambiguities. Results show varying model performance between narrative and neutral problems, with specific challenges in the Euler math-based subset with GPT-3.5-Turbo passing 50% of the AoC challenges and only 8% on the Euler problems. By probing the limits of LLMs' capabilities, our benchmark provides a framework to monitor and assess the subsequent progress of LLMs as a universal problem solver.

cs.AI↗

arXiv2024

PoTeC: A German Naturalistic Eye-tracking-while-reading Corpus

The Potsdam Textbook Corpus (PoTeC) is a naturalistic eye-tracking-while-reading corpus containing data from 75 participants reading 12 scientific texts. PoTeC is the first naturalistic eye-tracking-while-reading corpus that contains eye-movements from domain-experts as well as novices in a within-participant manipulation: It is based on a 2x2x2 fully-crossed factorial design which includes the participants' level of study and the participants' discipline of study as between-subject factors and the text domain as a within-subject factor. The participants' reading comprehension was assessed by a series of text comprehension questions and their domain knowledge was tested by text-independent background questions for each of the texts. The materials are annotated for a variety of linguistic features at different levels. We envision PoTeC to be used for a wide range of studies including but not limited to analyses of expert and non-expert reading strategies. The corpus and all the accompanying data at all stages of the preprocessing pipeline and all code used to preprocess the data are made available via GitHub: https://github.com/DiLi-Lab/PoTeC.

cs.CL↗

arXiv2024

Digital Comprehensibility Assessment of Simplified Texts among Persons with Intellectual Disabilities

Text simplification refers to the process of increasing the comprehensibility of texts. Automatic text simplification models are most commonly evaluated by experts or crowdworkers instead of the primary target groups of simplified texts, such as persons with intellectual disabilities. We conducted an evaluation study of text comprehensibility including participants with and without intellectual disabilities reading unsimplified, automatically and manually simplified German texts on a tablet computer. We explored four different approaches to measuring comprehensibility: multiple-choice comprehension questions, perceived difficulty ratings, response time, and reading speed. The results revealed significant variations in these measurements, depending on the reader group and whether the text had undergone automatic or manual simplification. For the target group of persons with intellectual disabilities, comprehension questions emerged as the most reliable measure, while analyzing reading speed provided valuable insights into participants' reading behavior.

cs.CL↗

arXiv2023

ScanDL: A Diffusion Model for Generating Synthetic Scanpaths on Texts

Eye movements in reading play a crucial role in psycholinguistic research studying the cognitive mechanisms underlying human language processing. More recently, the tight coupling between eye movements and cognition has also been leveraged for language-related machine learning tasks such as the interpretability, enhancement, and pre-training of language models, as well as the inference of reader- and text-specific properties. However, scarcity of eye movement data and its unavailability at application time poses a major challenge for this line of research. Initially, this problem was tackled by resorting to cognitive models for synthesizing eye movement data. However, for the sole purpose of generating human-like scanpaths, purely data-driven machine-learning-based methods have proven to be more suitable. Following recent advances in adapting diffusion processes to discrete data, we propose ScanDL, a novel discrete sequence-to-sequence diffusion model that generates synthetic scanpaths on texts. By leveraging pre-trained word representations and jointly embedding both the stimulus text and the fixation sequence, our model captures multi-modal interactions between the two inputs. We evaluate ScanDL within- and across-dataset and demonstrate that it significantly outperforms state-of-the-art scanpath generation methods. Finally, we provide an extensive psycholinguistic analysis that underlines the model's ability to exhibit human-like reading behavior. Our implementation is made available at https://github.com/DiLi-Lab/ScanDL.

cs.CL↗