Jan Strich

Hans Ole Hatzel, Sebastian Steindl, Jan Strich

LLM-generated reviews for scientific papers are gaining considerable traction and are even being officially piloted by major conferences. We have to assume that not only reviewers are using LLM-assistance, but also that authors use LLMs to revise their papers before submitting. In this work, we perform empirical experiments on papers from the 2025 ACL Rolling Review (ARR) to evaluate LLM reviews from both the author and the reviewer perspective. First, we identify a limited alignment of LLM reviews with human ones. In the best-case scenario, the alignment is reasonable. However, we also find that LLM-human alignment varies substantially across prompts and models. Finally, we investigate the scenario in which the author uses an iterative draft-revise workflow to improve the submission according to the LLM review. We find that this "gaming" of LLM reviews can be effective in specific scenarios, leading to a statistically significant increase of overall scores for up to 35\% of papers. We publish our code: https://github.com/uhh-hcds/reviewarcade.

Narges Baba Ahmadi, Jan Strich, Martin Semmann et al.

Large language models (LLMs) are increasingly used to access legal information. Yet, their deployment in multilingual legal settings is constrained by unreliable retrieval and the lack of domain-adapted, open-embedding models. In particular, existing multilingual legal corpora are not designed for semantic retrieval, and PDF-based legislative sources introduce substantial noise due to imperfect text extraction. To address these challenges, we introduce LEMUR, a large-scale multilingual corpus of EU environmental legislation constructed from 24,953 official EUR-Lex PDF documents covering 25 languages. We quantify the fidelity of PDF-to-text conversion by measuring lexical consistency against authoritative HTML versions using the Lexical Content Score (LCS). Building on LEMUR, we fine-tune three state-of-the-art multilingual embedding models using contrastive objectives in both monolingual and bilingual settings, reflecting realistic legal-retrieval scenarios. Experiments across low- and high-resource languages demonstrate that legal-domain fine-tuning consistently improves Top-k retrieval accuracy relative to strong baselines, with particularly pronounced gains for low-resource languages. Cross-lingual evaluations show that these improvements transfer to unseen languages, indicating that fine-tuning primarily enhances language-independent, content-level legal representations rather than language-specific cues. We publish code\footnote{\href{https://github.com/nargesbh/eur_lex}{GitHub Repository}} and data\footnote{\href{https://huggingface.co/datasets/G4KMU/LEMUR}{Hugging Face Dataset}}.

Klejda Alushi, Jan Strich, Chris Biemann et al.

Conversational question answering increasingly relies on retrieval-augmented generation (RAG) to ground large language models (LLMs) in external knowledge. Yet, most existing studies evaluate RAG methods in isolation and primarily focus on single-turn settings. This paper addresses the lack of a systematic comparison of RAG methods for multi-turn conversational QA, where dialogue history, coreference, and shifting user intent substantially complicate retrieval. We present a comprehensive empirical study of vanilla and advanced RAG methods across eight diverse conversational QA datasets spanning multiple domains. Using a unified experimental setup, we evaluate retrieval quality and answer generation using generator and retrieval metrics, and analyze how performance evolves across conversation turns. Our results show that robust yet straightforward methods, such as reranking, hybrid BM25, and HyDE, consistently outperform vanilla RAG. In contrast, several advanced techniques fail to yield gains and can even degrade performance below the No-RAG baseline. We further demonstrate that dataset characteristics and dialogue length strongly influence retrieval effectiveness, explaining why no single RAG strategy dominates across settings. Overall, our findings indicate that effective conversational RAG depends less on method complexity than on alignment between the retrieval strategy and the dataset structure. We publish the code used.\footnote{\href{https://github.com/Klejda-A/exp-rag.git}{GitHub Repository}}

Anshul Singh, Chris Biemann, Jan Strich

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in interpreting visual layouts and text. However, a significant challenge remains in their ability to interpret robustly and reason over multi-tabular data presented as images, a common occurrence in real-world scenarios like web pages and digital documents. Existing benchmarks typically address single tables or non-visual data (text/structured). This leaves a critical gap: they don't assess the ability to parse diverse table images, correlate information across them, and perform multi-hop reasoning on the combined visual data. We introduce MTabVQA, a novel benchmark specifically designed for multi-tabular visual question answering to bridge that gap. MTabVQA comprises 3,745 complex question-answer pairs that necessitate multi-hop reasoning across several visually rendered table images. We provide extensive benchmark results for state-of-the-art VLMs on MTabVQA, revealing significant performance limitations. We further investigate post-training techniques to enhance these reasoning abilities and release MTabVQA-Instruct, a large-scale instruction-tuning dataset. Our experiments show that fine-tuning VLMs with MTabVQA-Instruct substantially improves their performance on visual multi-tabular reasoning. Code and dataset (https://huggingface.co/datasets/mtabvqa/MTabVQA-Eval) are available online (https://anonymous.4open.science/r/MTabVQA-EMNLP-B16E).

Mohamed Rissal Hedna, Jan Strich, Martin Semmann et al.

Large language models (LLMs) are systematically overconfident: they routinely express high certainty on questions they often answer incorrectly. Existing calibration methods either require labeled validation data, degrade under distribution shifts, or incur substantial inference costs. Recent work has shown that LLMs already contain a better-calibrated signal than the one they verbalize: the token probability of "True" when the model is asked "Is this answer correct?" ($P(\text{True})$) consistently outperforms their stated confidence, a gap that is theoretically grounded as generative error is lower-bounded by roughly twice the corresponding discriminative error. We introduce $\textbf{SECL}$ ($\textbf{SE}$lf-$\textbf{C}$alibrating $\textbf{L}$anguage Models), a test-time training (TTT) pipeline that exploits this gap as label-free self-supervision, requiring no labeled data or human supervision. SECL adapts only when the input distribution shifts, training on just 6--26% of the question stream at lower cost than the baseline it distills from. Across four small language models from three model families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56--78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods. SECL is the first method to apply TTT to calibration; seven ablations covering signal quality, gating strategy, weight accumulation, loss design, domain ordering, hyperparameter sensitivity, and layer selection confirm that each component is crucial and robust across configurations. Code: https://anonymous.4open.science/r/secl-emnlp26-submission-C890

Jan Strich, Adeline Scharfenberg, Chris Biemann et al.

We introduce EncouRAGe, a comprehensive Python framework designed to streamline the development and evaluation of Retrieval-Augmented Generation (RAG) systems using Large Language Models (LLMs) and Embedding Models. EncouRAGe comprises five modular and extensible components: Type Manifest, RAG Factory, Inference, Vector Store, and Metrics, facilitating flexible experimentation and extensible development. The framework emphasizes scientific reproducibility, diverse evaluation metrics, and local deployment, enabling researchers to efficiently assess datasets within RAG workflows. This paper presents implementation details and an extensive evaluation across multiple benchmark datasets, including 25k QA pairs and over 51k documents. Our results show that RAG still underperforms compared to the Oracle Context, while Hybrid BM25 consistently achieves the best results across all four datasets. We further examine the effects of reranking, observing only marginal performance improvements accompanied by higher response latency.