arXiv subjects

Martin Potthast

Publications and source records attributed to Martin Potthast.

At least 19 recordsLinked to original sources

arXiv2026

When Attention Goes Blind: Numerical Failure in ALiBi Positional Encodings

We identify a previously overlooked failure mode of ALiBi positional encoding: its linear bias scaling underflows floating-point precision, which zeroes out a large fraction of attention weights and renders the affected attention heads partially blind. We analyze this failure mode, characterize its impact, and examine four mitigation strategies. We further demonstrate its occurrence in state-of-the-art pretrained models based on ALiBi. Comprehensive pretraining experiments with 148M-parameter decoder models help us to disentangle its effects from out-of-context degradation. We find that ALiBi's failure mode can substantially impair token retrieval while having only a minor effect on standard decoder benchmarks. We propose four training-time mitigation strategies and evaluate them individually and in combinations, finding that log-scaled distances yield the most consistent improvements in passkey retrieval. Despite this problem, default ALiBi slopes remain a surprisingly strong baseline, particularly for needle-in-a-haystack retrieval. Based on these findings we provide concrete recommendations on how to train models with ALiBi.

cs.CL↗

arXiv2026

Data Annotations as Pedagogical Hints: From Subjective Labels to Critical Thinking

Machine learning courses often use pre-labeled datasets, hiding the subjectivity of human annotation. This creates students with an overly trusting view of AI data and models, undervaluing interpretive diversity. We investigated whether manual data annotation tasks teach students about subjective labeling. Study Design: An annotation activity was implemented at two universities: Fontys (Netherlands) and IT University Copenhagen (Denmark). Students annotated skin lesion images for hair coverage on a 3-point scale. Surveys were collected from 43 participants measuring their understanding of annotation ambiguity, data quality, bias, fairness, implementation barriers, and pedagogical effectiveness. Key Findings: Self-reported familiarity with course content increased substantially across all concepts. Most students recognised that personal interpretation affects annotations. Students rated the activity as more effective than traditional lectures for understanding bias. Participants were motivated to learn more. Main Drawbacks: Emotional unease from viewing medical images was the primary issue. Many students still requested clearer guidelines to reduce disagreement, suggesting they hadn't internalised that disagreement from different perspectives is a learning feature, not a bug. Recommendations for Future Iterations: Ensure sufficient interpretive ambiguity in materials. Reduce repetitive annotation workload. Mitigate emotional unease from sensitive content. Explicitly frame disagreement as a learning opportunity rather than a problem to solve. Manual data annotations effectively teach students that human judgment shapes model behavior and that disagreement reflects domain complexity, not just noise.

cs.CY↗

arXiv2026

A Large-Scale, Cross-Disciplinary Corpus of Systematic Reviews

Existing benchmarks for systematic reviewing remain limited either in scale or in disciplinary coverage, with some collections comprising only a modest number of topics and others focusing primarily on biomedical research. We present Webis-SR4ALL-26, a large-scale, cross-disciplinary corpus of 301,871 systematic reviews spanning all scientific fields as covered by OpenAlex. Using a multi-stage pre-processing pipeline, we link reviews to resolved OpenAlex metadata and reference lists and extract, when explicitly reported, structured method artifacts relevant to retrieval and screening. These artifacts include reported search strategies (Boolean queries or keyword lists) that we normalize into executable approximations, as well as reported inclusion and exclusion criteria. Together, these layers support cross-domain benchmarking of retrieval and screening components against review reference lists, training and evaluation of extraction methods for review artifacts, and comparative meta-science analyses of systematic review practices across disciplines and time. To demonstrate one concrete use case, we report large-scale baseline retrieval signals by executing normalized search strategies in OpenAlex and comparing retrieved sets to resolved reference lists. We release the corpus and the pre-processing pipeline, along with code used for extraction validation and the retrieval demonstration.

cs.IR↗

arXiv2026

Detecting RAG Advertisements Across Advertising Styles

Large language models (LLMs) enable a new form of advertising for retrieval-augmented generation (RAG) systems in which organic responses are blended with contextually relevant ads. The prospect of such "generated native ads" has sparked interest in whether they can be detected automatically. Existing datasets, however, do not reflect the diversity of advertising styles discussed in the marketing literature. In this paper, we (1) develop a taxonomy of advertising styles for LLMs, combining the style dimensions of explicitness and type of appeal, (2) simulate that advertisers may attempt to evade detection by changing their advertising style, and (3) evaluate a variety of ad-detection approaches with respect to their robustness under these changes. Expanding previous work on ad detection, we train models that use entity recognition to exactly locate an ad in an LLM response and find them to be both very effective at detecting responses with ads and largely robust to changes in the advertising style. Since ad blocking will be performed on low-resource end-user devices, we include lightweight models like random forests and SVMs in our evaluation. These models, however, are brittle under such changes, highlighting the need for further efficiency-oriented research for a practical approach to blocking of generated ads.

cs.IR↗

arXiv2026

Representation Fidelity:Auditing Algorithmic Decisions About Humans Using Self-Descriptions

This paper introduces a new dimension for validating algorithmic decisions about humans by measuring the fidelity of their representations. Representation Fidelity measures if decisions about a person rest on reasonable grounds. We propose to operationalize this notion by measuring the distance between two representations of the same person: (1) an externally prescribed input representation on which the decision is based, and (2) a self-description provided by the human subject of the decision, used solely to validate the input representation. We examine the nature of discrepancies between these representations, how such discrepancies can be quantified, and derive a generic typology of representation mismatches that determine the degree of representation fidelity. We further present the first benchmark for evaluating representation fidelity based on a dataset of loan-granting decisions. Our Loan-Granting Self-Representations Corpus 2025 consists of a large corpus of 30 000 synthetic natural language self-descriptions derived from corresponding representations of applicants in the German Credit Dataset, along with expert annotations of representation mismatches between each pair of representations.

cs.CL↗

arXiv2026

Overview of PAN 2026: Voight-Kampff Generative AI Detection, Text Watermarking, Multi-Author Writing Style Analysis, Generative Plagiarism Detection, and Reasoning Trajectory Detection

The goal of the PAN workshop is to advance computational stylometry and text forensics via objective and reproducible evaluation. In 2026, we run the following five tasks: (1) Voight-Kampff Generative AI Detection, particularly in mixed and obfuscated authorship scenarios, (2) Text Watermarking, a new task that aims to find new and benchmark the robustness of existing text watermarking schemes, (3) Multi-author Writing Style Analysis, a continued task that aims to find positions of authorship change, (4) Generative Plagiarism Detection, a continued task that targets source retrieval and text alignment between generated text and source documents, and (5) Reasoning Trajectory Detection, a new task that deals with source detection and safety detection of LLM-generated or human-written reasoning trajectories. As in previous years, PAN invites software submissions as easy-to-reproduce Docker containers for most of the tasks. Since PAN 2012, more than 1,100 submissions have been made this way via the TIRA experimentation platform.

cs.CL↗

arXiv2025

The German Commons - 154 Billion Tokens of Openly Licensed Text for German Language Models

Large language model development relies on large-scale training corpora, yet most contain data of unclear licensing status, limiting the development of truly open models. This problem is exacerbated for non-English languages, where openly licensed text remains critically scarce. We introduce the German Commons, the largest collection of openly licensed German text to date. It compiles data from 41 sources across seven domains, encompassing legal, scientific, cultural, political, news, economic, and web text. Through systematic sourcing from established data providers with verifiable licensing, it yields 154.56 billion tokens of high-quality text for language model training. Our processing pipeline implements comprehensive quality filtering, deduplication, and text formatting fixes, ensuring consistent quality across heterogeneous text sources. All domain subsets feature licenses of at least CC-BY-SA 4.0 or equivalent, ensuring legal compliance for model training and redistribution. The German Commons therefore addresses the critical gap in openly licensed German pretraining data, and enables the development of truly open German language models. We also release code for corpus construction and data filtering tailored to German language text, rendering the German Commons fully reproducible and extensible.

cs.CL↗

arXiv2025

Investigating Counterclaims in Causality Extraction from Text

Many causal claims, such as "sugar causes hyperactivity," are disputed or outdated. Yet research on causality extraction from text has almost entirely neglected counterclaims of causation. To close this gap, we conduct a thorough literature review of causality extraction, compile an extensive inventory of linguistic realizations of countercausal claims, and develop rigorous annotation guidelines that explicitly incorporate countercausal language. We also highlight how counterclaims of causation are an integral part of causal reasoning. Based on our guidelines, we construct a new dataset comprising 1028 causal claims, 952 counterclaims, and 1435 uncausal statements, achieving substantial inter-annotator agreement (Cohen's $\kappa = 0.74$). In our experiments, state-of-the-art models trained solely on causal claims misclassify counterclaims more than 10 times as often as models trained on our dataset.

cs.CL↗

arXiv2025

Overview of the Plagiarism Detection Task at PAN 2025

The generative plagiarism detection task at PAN 2025 aims at identifying automatically generated textual plagiarism in scientific articles and aligning them with their respective sources. We created a novel large-scale dataset of automatically generated plagiarism using three large language models: Llama, DeepSeek-R1, and Mistral. In this task overview paper, we outline the creation of this dataset, summarize and compare the results of all participants and four baselines, and evaluate the results on the last plagiarism detection task from PAN 2015 in order to interpret the robustness of the proposed approaches. We found that the current iteration does not invite a large variety of approaches as naive semantic similarity approaches based on embedding vectors provide promising results of up to 0.8 recall and 0.5 precision. In contrast, most of these approaches underperform significantly on the 2015 dataset, indicating a lack in generalizability.

cs.CL↗

arXiv2025

Topic-Specific Classifiers are Better Relevance Judges than Prompted LLMs

The unjudged document problem, where systems that did not contribute to the original judgement pool may retrieve documents without a relevance judgement, is a key obstacle to the reuseability of test collections in information retrieval. While the de facto standard to deal with the problem is to treat unjudged documents as non-relevant, many alternatives have been proposed, such as the use of large language models (LLMs) as a relevance judge (LLM-as-a-judge). However, this has been criticized, among other things, as circular, since the same LLM can be used as the ranker and the judge. We propose to train topic-specific relevance classifiers instead: By finetuning monoT5 with independent LoRA weight adaptation on the judgments of a single assessor for a single topic's pool, we align it to that assessor's notion of relevance for the topic. The system rankings obtained through our classifier's relevance judgments achieve a Spearmans' $\rho$ correlation of $>0.94$ with ground truth system rankings. As little as 128 initial human judgments per topic suffice to improve the comparability of models, compared to treating unjudged documents as non-relevant, while achieving more reliability than existing LLM-as-a-judge approaches. Topic-specific relevance classifiers are thus a lightweight and straightforward way to tackle the unjudged document problem, while maintaining human judgments as the gold standard for retrieval evaluation. Code, models, and data are made openly available.

cs.IR↗

arXiv2025

Simplified Longitudinal Retrieval Experiments: A Case Study on Query Expansion and Document Boosting

The longitudinal evaluation of retrieval systems aims to capture how information needs and documents evolve over time. However, classical Cranfield-style retrieval evaluations only consist of a static set of queries and documents and thereby miss time as an evaluation dimension. Therefore, longitudinal evaluations need to complement retrieval toolkits with custom logic. This custom logic increases the complexity of research software, which might reduce the reproducibility and extensibility of experiments. Based on our submissions to the 2024 edition of LongEval, we propose a custom extension of ir_datasets for longitudinal retrieval experiments. This extension allows for declaratively, instead of imperatively, describing important aspects of longitudinal retrieval experiments, e.g., which queries, documents, and/or relevance feedback are available at which point in time. We reimplement our submissions to LongEval 2024 against our new ir_datasets extension, and find that the declarative access can reduce the complexity of the code.

cs.IR↗

arXiv2025

Ask, Retrieve, Summarize: A Modular Pipeline for Scientific Literature Summarization

The exponential growth of scientific publications has made it increasingly difficult for researchers to stay updated and synthesize knowledge effectively. This paper presents XSum, a modular pipeline for multi-document summarization (MDS) in the scientific domain using Retrieval-Augmented Generation (RAG). The pipeline includes two core components: a question-generation module and an editor module. The question-generation module dynamically generates questions adapted to the input papers, ensuring the retrieval of relevant and accurate information. The editor module synthesizes the retrieved content into coherent and well-structured summaries that adhere to academic standards for proper citation. Evaluated on the SurveySum dataset, XSum demonstrates strong performance, achieving considerable improvements in metrics such as CheckEval, G-Eval and Ref-F1 compared to existing approaches. This work provides a transparent, adaptable framework for scientific summarization with potential applications in a wide range of domains. Code available at https://github.com/webis-de/scolia25-xsum

cs.CL↗

arXiv2025

The Viability of Crowdsourcing for RAG Evaluation

How good are humans at writing and judging responses in retrieval-augmented generation (RAG) scenarios? To answer this question, we investigate the efficacy of crowdsourcing for RAG through two complementary studies: response writing and response utility judgment. We present the Crowd RAG Corpus 2025 (CrowdRAG-25), which consists of 903 human-written and 903 LLM-generated responses for the 301 topics of the TREC RAG'24 track, across the three discourse styles 'bulleted list', 'essay', and 'news'. For a selection of 65 topics, the corpus further contains 47,320 pairwise human judgments and 10,556 pairwise LLM judgments across seven utility dimensions (e.g., coverage and coherence). Our analyses give insights into human writing behavior for RAG and the viability of crowdsourcing for RAG evaluation. Human pairwise judgments provide reliable and cost-effective results compared to LLM-based pairwise or human/LLM-based pointwise judgments, as well as automated comparisons with human-written reference responses. All our data and tools are freely available.

cs.IR↗

arXiv2025

AiReview: An Open Platform for Accelerating Systematic Reviews with LLMs

Systematic reviews are fundamental to evidence-based medicine. Creating one is time-consuming and labour-intensive, mainly due to the need to screen, or assess, many studies for inclusion in the review. Several tools have been developed to streamline this process, mostly relying on traditional machine learning methods. Large language models (LLMs) have shown potential in further accelerating the screening process. However, no tool currently allows end users to directly leverage LLMs for screening or facilitates systematic and transparent usage of LLM-assisted screening methods. This paper introduces (i) an extensible framework for applying LLMs to systematic review tasks, particularly title and abstract screening, and (ii) a web-based interface for LLM-assisted screening. Together, these elements form AiReview-a novel platform for LLM-assisted systematic review creation. AiReview is the first of its kind to bridge the gap between cutting-edge LLM-assisted screening methods and those that create medical systematic reviews. The tool is available at https://aireview.ielab.io. The source code is also open sourced at https://github.com/ielab/ai-review.

cs.IR↗

arXiv2025

Counterfactual Query Rewriting to Use Historical Relevance Feedback

When a retrieval system receives a query it has encountered before, previous relevance feedback, such as clicks or explicit judgments can help to improve retrieval results. However, the content of a previously relevant document may have changed, or the document might not be available anymore. Despite this evolved corpus, we counterfactually use these previously relevant documents as relevance signals. In this paper we proposed approaches to rewrite user queries and compare them against a system that directly uses the previous qrels for the ranking. We expand queries with terms extracted from the previously relevant documents or derive so-called keyqueries that rank the previously relevant documents to the top of the current corpus. Our evaluation in the CLEF LongEval scenario shows that rewriting queries with historical relevance feedback improves the retrieval effectiveness and even outperforms computationally expensive transformer-based approaches.

cs.IR↗

arXiv2024

Ranking Generated Answers: On the Agreement of Retrieval Models with Humans on Consumer Health Questions

Evaluating the output of generative large language models (LLMs) is challenging and difficult to scale. Many evaluations of LLMs focus on tasks such as single-choice question-answering or text classification. These tasks are not suitable for assessing open-ended question-answering capabilities, which are critical in domains where expertise is required. One such domain is health, where misleading or incorrect answers can have a negative impact on a user's well-being. Using human experts to evaluate the quality of LLM answers is generally considered the gold standard, but expert annotation is costly and slow. We present a method for evaluating LLM answers that uses ranking models trained on annotated document collections as a substitute for explicit relevance judgements and apply it to the CLEF 2021 eHealth dataset. In a user study, our method correlates with the preferences of a human expert (Kendall's $\tau=0.64$). It is also consistent with previous findings in that the quality of generated answers improves with the size of the model and more sophisticated prompting strategies.

cs.IR↗

arXiv2024

Learning Effective Representations for Retrieval Using Self-Distillation with Adaptive Relevance Margins

Representation-based retrieval models, so-called bi-encoders, estimate the relevance of a document to a query by calculating the similarity of their respective embeddings. Current state-of-the-art bi-encoders are trained using an expensive training regime involving knowledge distillation from a teacher model and batch-sampling. Instead of relying on a teacher model, we contribute a novel parameter-free loss function for self-supervision that exploits the pre-trained language modeling capabilities of the encoder model as a training signal, eliminating the need for batch sampling by performing implicit hard negative mining. We investigate the capabilities of our proposed approach through extensive experiments, demonstrating that self-distillation can match the effectiveness of teacher distillation using only 13.5% of the data, while offering a speedup in training time between 3x and 15x compared to parametrized losses. All code and data is made openly available.

cs.IR↗

arXiv2024

Systematic Evaluation of Neural Retrieval Models on the Touch\'e 2020 Argument Retrieval Subset of BEIR

The zero-shot effectiveness of neural retrieval models is often evaluated on the BEIR benchmark -- a combination of different IR evaluation datasets. Interestingly, previous studies found that particularly on the BEIR subset Touch\'e 2020, an argument retrieval task, neural retrieval models are considerably less effective than BM25. Still, so far, no further investigation has been conducted on what makes argument retrieval so "special". To more deeply analyze the respective potential limits of neural retrieval models, we run a reproducibility study on the Touch\'e 2020 data. In our study, we focus on two experiments: (i) a black-box evaluation (i.e., no model retraining), incorporating a theoretical exploration using retrieval axioms, and (ii) a data denoising evaluation involving post-hoc relevance judgments. Our black-box evaluation reveals an inherent bias of neural models towards retrieving short passages from the Touch\'e 2020 data, and we also find that quite a few of the neural models' results are unjudged in the Touch\'e 2020 data. As many of the short Touch\'e passages are not argumentative and thus non-relevant per se, and as the missing judgments complicate fair comparison, we denoise the Touch\'e 2020 data by excluding very short passages (less than 20 words) and by augmenting the unjudged data with post-hoc judgments following the Touch\'e guidelines. On the denoised data, the effectiveness of the neural models improves by up to 0.52 in nDCG@10, but BM25 is still more effective. Our code and the augmented Touch\'e 2020 dataset are available at \url{https://github.com/castorini/touche-error-analysis}.

cs.IR↗