arXiv subjects

Lukas Galke

Publications and source records attributed to Lukas Galke.

At least 19 recordsLinked to original sources

arXiv2026

Efficient Continual Learning for Small Language Models with a Discrete Key-Value Bottleneck

Continual learning remains a challenge across various natural language processing (NLP) tasks, as models updated with new training data often risk catastrophic forgetting of previously acquired knowledge. We introduce a discrete key-value bottleneck (DKVB) for encoder-only language models, enabling efficient continual learning through localized updates. Inspired by a discrete key-value bottleneck in vision, we consider new and NLP-specific challenges. We compare different bottleneck architectures for NLP and introduce a new, task-independent initialization technique for the discrete keys. We evaluate our DKVB for NLP in four continual learning scenarios and show that it alleviates catastrophic forgetting. Our experiments demonstrate that the proposed approach achieves competitive performance compared to popular continual learning methods while incurring lower computational costs. Furthermore, we show that DKVB remains effective even in challenging single-head continual learning scenarios where no task ID is provided.

cs.CL↗

arXiv2026

Your Extreme Multi-label Classifier is Secretly a Hierarchical Text Classifier for Free

Assigning a set of labels to a given text is a classification problem with many real-world applications, such as recommender systems. Two separate research streams address this issue. Hierarchical Text Classification (HTC) focuses on datasets with label pools of hundreds of entries, accompanied by a semantic label hierarchy. In contrast, eXtreme Multi-Label Text Classification (XML) considers very large sets of labels with up to millions of entries but without an explicit hierarchy. In XML methods, it is common to construct an artificial hierarchy in order to deal with the large label space before or during the training process. Here, we investigate how state-of-the-art HTC models perform when trained and tested on XML datasets and vice versa using three benchmark datasets from each of the two streams. Our results demonstrate that XML models, with their internally constructed hierarchy, are very effective HTC models. HTC models, on the other hand, are not equipped to handle the sheer label set size of XML datasets and achieve poor transfer results. We further argue that for a fair comparison in HTC and XML, more than one metric like F1 should be used but complemented with P@k and R-Precision.

cs.CL↗

arXiv2025

Not Everything That Counts Can Be Counted: A Case for Safe Qualitative AI

Artificial intelligence (AI) and large language models (LLM) are reshaping science, with most recent advances culminating in fully-automated scientific discovery pipelines. But qualitative research has been left behind. Researchers in qualitative methods are hesitant about AI adoption. Yet when they are willing to use AI at all, they have little choice but to rely on general-purpose tools like ChatGPT to assist with interview interpretation, data annotation, and topic modeling - while simultaneously acknowledging these system's well-known limitations of being biased, opaque, irreproducible, and privacy-compromising. This creates a critical gap: while AI has substantially advanced quantitative methods, the qualitative dimensions essential for meaning-making and comprehensive scientific understanding remain poorly integrated. We argue for developing dedicated qualitative AI systems built from the ground up for interpretive research. Such systems must be transparent, reproducible, and privacy-friendly. We review recent literature to show how existing automated discovery pipelines could be enhanced by robust qualitative capabilities, and identify key opportunities where safe qualitative AI could advance multidisciplinary and mixed-methods research.

cs.AI↗

arXiv2025

Guarded Query Routing for Large Language Models

Query routing, the task to route user queries to different large language model (LLM) endpoints, can be considered as a text classification problem. However, out-of-distribution queries must be handled properly, as those could be about unrelated domains, queries in other languages, or even contain unsafe text. Here, we thus study a guarded query routing problem, for which we first introduce the Guarded Query Routing Benchmark (GQR-Bench, released as Python package gqr), covers three exemplary target domains (law, finance, and healthcare), and seven datasets to test robustness against out-of-distribution queries. We then use GQR-Bench to contrast the effectiveness and efficiency of LLM-based routing mechanisms (GPT-4o-mini, Llama-3.2-3B, and Llama-3.1-8B), standard LLM-based guardrail approaches (LlamaGuard and NVIDIA NeMo Guardrails), continuous bag-of-words classifiers (WideMLP, fastText), and traditional machine learning models (SVM, XGBoost). Our results show that WideMLP, enhanced with out-of-domain detection capabilities, yields the best trade-off between accuracy (88%) and speed (<4ms). The embedding-based fastText excels at speed (<1ms) with acceptable accuracy (80%), whereas LLMs yield the highest accuracy (91%) but are comparatively slow (62ms for local Llama-3.1:8B and 669ms for remote GPT-4o-mini calls). Our findings challenge the automatic reliance on LLMs for (guarded) query routing and provide concrete recommendations for practical applications. Source code is available: https://github.com/williambrach/gqr.

cs.AI↗

arXiv2025

Gumbel-MPNN: Graph Rewiring with Gumbel-Softmax

Graph homophily has been considered an essential property for message-passing neural networks (MPNN) in node classification. Recent findings suggest that performance is more closely tied to the consistency of neighborhood class distributions. We demonstrate that the MPNN performance depends on the number of components of the overall neighborhood distribution within a class. By breaking down the classes into their neighborhood distribution components, we increase measures of neighborhood distribution informativeness but do not observe an improvement in MPNN performance. We propose a Gumbel-Softmax-based rewiring method that reduces deviations in neighborhood distributions. Our results show that our new method enhances neighborhood informativeness, handles long-range dependencies, mitigates oversquashing, and increases the classification performance of the MPNN. The code is available at https://github.com/Bobowner/Gumbel-Softmax-MPNN.

cs.LG↗

arXiv2025

Super-additive Cooperation in Language Model Agents

With the prospect of autonomous artificial intelligence (AI) agents, studying their tendency for cooperative behavior becomes an increasingly relevant topic. This study is inspired by the super-additive cooperation theory, where the combined effects of repeated interactions and inter-group rivalry have been argued to be the cause for cooperative tendencies found in humans. We devised a virtual tournament where language model agents, grouped into teams, face each other in a Prisoner's Dilemma game. By simulating both internal team dynamics and external competition, we discovered that this blend substantially boosts both overall and initial, one-shot cooperation levels (the tendency to cooperate in one-off interactions). This research provides a novel framework for large language models to strategize and act in complex social scenarios and offers evidence for how intergroup competition can, counter-intuitively, result in more cooperative behavior. These insights are crucial for designing future multi-agent AI systems that can effectively work together and better align with human values. Source code is available at https://github.com/pippot/Superadditive-cooperation-LLMs.

cs.AI↗

arXiv2025

Dynaword: From One-shot to Continuously Developed Datasets

Large-scale datasets are foundational for research and development in natural language processing. However, current approaches face three key challenges: (1) reliance on ambiguously licensed sources restricting use, sharing, and derivative works; (2) static dataset releases that prevent community contributions and diminish longevity; and (3) quality assurance processes restricted to publishing teams rather than leveraging community expertise. To address these limitations, we introduce two contributions: the Dynaword approach and Danish Dynaword. The Dynaword approach is a framework for creating large-scale, open datasets that can be continuously updated through community collaboration. Danish Dynaword is a concrete implementation that validates this approach and demonstrates its potential. Danish Dynaword contains over four times as many tokens as comparable releases, is exclusively openly licensed, and has received multiple contributions across industry and research. The repository includes light-weight tests to ensure data formatting, quality, and documentation, establishing a sustainable framework for ongoing community contributions and dataset evolution.

cs.CL↗

arXiv2025

Four Shades of Life Sciences: A Dataset for Disinformation Detection in the Life Sciences

Disseminators of disinformation often seek to attract attention or evoke emotions - typically to gain influence or generate revenue - resulting in distinctive rhetorical patterns that can be exploited by machine learning models. In this study, we explore linguistic and rhetorical features as proxies for distinguishing disinformative texts from other health and life-science text genres, applying both large language models and classical machine learning classifiers. Given the limitations of existing datasets, which mainly focus on fact checking misinformation, we introduce Four Shades of Life Sciences (FSoLS): a novel, labeled corpus of 2,603 texts on 14 life-science topics, retrieved from 17 diverse sources and classified into four categories of life science publications. The source code for replicating, and updating the dataset is available on GitHub: https://github.com/EvaSeidlmayer/FourShadesofLifeSciences

cs.CL↗

arXiv2025

Continual Quantization-Aware Pre-Training: When to transition from 16-bit to 1.58-bit pre-training for BitNet language models?

Large language models (LLMs) require immense resources for training and inference. Quantization, a technique that reduces the precision of model parameters, offers a promising solution for improving LLM efficiency and sustainability. While post-training quantization methods typically achieve 4-8 bits per parameter, recent research suggests that training LLMs with 1.58 bits per weight parameter from scratch can maintain model accuracy while greatly reducing memory requirements and energy consumption at inference time. Here, we investigate a training strategy for quantization-aware pre-training, where the models are first trained with 16-bit precision and then transition into 1.58-bit quantization-aware training. Our results on 11 downstream tasks show that this 16-to-1.58-bit training strategy is preferable over full 1.58-bit training and leaves models closer to those which have undergone 16-bit training. We further investigate the effects of retaining the optimizer state at the transition point and gradually phasing in quantization strength -- finding that both techniques alleviate the magnitude of loss spikes, but also that these effects can be compensated through further training.

cs.LG↗

arXiv2025

A Transformer-based Autoregressive Decoder Architecture for Hierarchical Text Classification

Recent approaches in hierarchical text classification (HTC) rely on the capabilities of a pre-trained transformer model and exploit the label semantics and a graph encoder for the label hierarchy. In this paper, we introduce an effective hierarchical text classifier RADAr (Transformer-based Autoregressive Decoder Architecture) that is based only on an off-the-shelf RoBERTa transformer to process the input and a custom autoregressive decoder with two decoder layers for generating the classification output. Thus, unlike existing approaches for HTC, the encoder of RADAr has no explicit encoding of the label hierarchy and the decoder solely relies on the label sequences of the samples observed during training. We demonstrate on three benchmark datasets that RADAr achieves results competitive to the state of the art with less training and inference time. Our model consistently performs better when organizing the label sequences from children to parents versus the inverse, as done in existing HTC approaches. Our experiments show that neither the label semantics nor an explicit graph encoder for the hierarchy is needed. This has strong practical implications for HTC as the architecture has fewer requirements and provides a speed-up by a factor of 2 at inference time. Moreover, training a separate decoder from scratch in conjunction with fine-tuning the encoder allows future researchers and practitioners to exchange the encoder part as new models arise. The source code is available at https://github.com/yousef-younes/RADAr.

cs.LG↗

arXiv2025

Are We Really Making Much Progress in Text Classification? A Comparative Review

We analyze various methods for single-label and multi-label text classification across well-known datasets, categorizing them into bag-of-words, sequence-based, graph-based, and hierarchical approaches. Despite the surge in methods like graph-based models, encoder-only pre-trained language models, notably BERT, remain state-of-the-art. However, recent findings suggest simpler models like logistic regression and trigram-based SVMs outperform newer techniques. While decoder-only generative language models show promise in learning with limited data, they lag behind encoder-only models in performance. We emphasize the superiority of discriminative language models like BERT over generative models for supervised tasks. Additionally, we highlight the literature's lack of robustness in method comparisons, particularly concerning basic hyperparameter optimizations like learning rate in fine-tuning encoder-only language models. Data availability: The source code is available at https://github.com/drndr/multilabel-text-clf All datasets used for our experiments are publicly available except the NYT dataset.

cs.CL↗

arXiv2024

Isotropy Matters: Soft-ZCA Whitening of Embeddings for Semantic Code Search

Low isotropy in an embedding space impairs performance on tasks involving semantic inference. Our study investigates the impact of isotropy on semantic code search performance and explores post-processing techniques to mitigate this issue. We analyze various code language models, examine isotropy in their embedding spaces, and its influence on search effectiveness. We propose a modified ZCA whitening technique to control isotropy levels in embeddings. Our results demonstrate that Soft-ZCA whitening improves the performance of pre-trained code language models and can complement contrastive fine-tuning.

cs.CL↗

arXiv2024

Learning and communication pressures in neural networks: Lessons from emergent communication

Finding and facilitating commonalities between the linguistic behaviors of large language models and humans could lead to major breakthroughs in our understanding of the acquisition, processing, and evolution of language. However, most findings on human-LLM similarity can be attributed to training on human data. The field of emergent machine-to-machine communication provides an ideal testbed for discovering which pressures are neural agents naturally exposed to when learning to communicate in isolation, without any human language to start with. Here, we review three cases where mismatches between the emergent linguistic behavior of neural agents and humans were resolved thanks to introducing theoretically-motivated inductive biases. By contrasting humans, large language models, and emergent communication agents, we then identify key pressures at play for language learning and emergence: communicative success, production effort, learnability, and other psycho-/sociolinguistic factors. We discuss their implications and relevance to the field of language evolution and acquisition. By mapping out the necessary inductive biases that make agents' emergent languages more human-like, we not only shed light on the underlying principles of human cognition and communication, but also inform and improve the very use of these models as valuable scientific tools for studying language learning, processing, use, and representation more broadly.

cs.CL↗

arXiv2024

When are 1.58 bits enough? A Bottom-up Exploration of BitNet Quantization

Contemporary machine learning models, such as language models, are powerful, but come with immense resource requirements both at training and inference time. It has been shown that decoder-only language models can be trained to a competitive state with ternary weights (1.58 bits per weight), facilitating efficient inference. Here, we start our exploration with non-transformer model architectures, investigating 1.58-bit training for multi-layer perceptrons and graph neural networks. Then, we explore 1.58-bit training in other transformer-based language models, namely encoder-only and encoder-decoder models. Our results show that in all of these settings, 1.58-bit training is on par with or sometimes even better than the standard 32/16-bit models.

cs.LG↗

arXiv2024

Tokenization and Morphology in Multilingual Language Models: A Comparative Analysis of mT5 and ByT5

Morphology is a crucial factor for multilingual language modeling as it poses direct challenges for tokenization. Here, we seek to understand how tokenization influences the morphological knowledge encoded in multilingual language models. Specifically, we capture the impact of tokenization by contrasting two multilingual language models: mT5 and ByT5. The two models share the same architecture, training objective, and training data and only differ in their tokenization strategies: subword tokenization vs.\@ character-level tokenization. Probing the morphological knowledge encoded in these models on four tasks and 17 languages, our analyses show that the models learn the morphological systems of some languages better than others and that morphological information is encoded in the middle and late layers. Finally, we show that languages with more irregularities benefit more from having a higher share of the pre-training data.

cs.CL↗

arXiv2024

What makes a language easy to deep-learn? Deep neural networks and humans similarly benefit from compositional structure

Deep neural networks drive the success of natural language processing. A fundamental property of language is its compositional structure, allowing humans to systematically produce forms for new meanings. For humans, languages with more compositional and transparent structures are typically easier to learn than those with opaque and irregular structures. However, this learnability advantage has not yet been shown for deep neural networks, limiting their use as models for human language learning. Here, we directly test how neural networks compare to humans in learning and generalizing different languages that vary in their degree of compositional structure. We evaluate the memorization and generalization capabilities of a large language model and recurrent neural networks, and show that both deep neural networks exhibit a learnability advantage for more structured linguistic input: neural networks exposed to more compositional languages show more systematic generalization, greater agreement between different agents, and greater similarity to human learners.

cs.CL↗

arXiv2024

POWN: Prototypical Open-World Node Classification

We consider the problem of \textit{true} open-world semi-supervised node classification, in which nodes in a graph either belong to known or new classes, with the latter not present during training. Existing methods detect and reject new classes but fail to distinguish between different new classes. We adapt existing methods and show they do not solve the problem sufficiently. We introduce a novel end-to-end approach for classification into known classes and new classes based on class prototypes, which we call Prototypical Open-World Learning for Node Classification (POWN). Our method combines graph semi-supervised learning, self-supervised learning, and pseudo-labeling to learn prototype representations of new classes in a zero-shot way. In contrast to existing solutions from the vision domain, POWN does not require data augmentation techniques for node classification. Experiments on benchmark datasets demonstrate the effectiveness of POWN, where it outperforms baselines by up to $20\%$ accuracy on the small and up to $30\%$ on the large datasets. Source code is available at https://github.com/Bobowner/POWN.

cs.LG↗

arXiv2023

GenCodeSearchNet: A Benchmark Test Suite for Evaluating Generalization in Programming Language Understanding

Language models can serve as a valuable tool for software developers to increase productivity. Large generative models can be used for code generation and code completion, while smaller encoder-only models are capable of performing code search tasks using natural language queries.These capabilities are heavily influenced by the quality and diversity of the available training data. Source code datasets used for training usually focus on the most popular languages and testing is mostly conducted on the same distributions, often overlooking low-resource programming languages. Motivated by the NLP generalization taxonomy proposed by Hupkes et.\,al., we propose a new benchmark dataset called GenCodeSearchNet (GeCS) which builds upon existing natural language code search datasets to systemically evaluate the programming language understanding generalization capabilities of language models. As part of the full dataset, we introduce a new, manually curated subset StatCodeSearch that focuses on R, a popular but so far underrepresented programming language that is often used by researchers outside the field of computer science. For evaluation and comparison, we collect several baseline results using fine-tuned BERT-style models and GPT-style large language models in a zero-shot setting.

cs.CL↗