Calibration of language models – the alignment between expressed or implicit confidence and empirical correctness – is a well-studied subfield within NLP. Methods to measure it already exist. The problem is adoption: outside this subfield, NLP research regularly introduces new models, datasets, and benchmarks without checking whether the model’s confidence scores are meaningful. We argue that this adoption gap is a major obstacle to trustworthy LLM evaluation. Miscalibration causes problems in two distinct areas: at deployment, where overconfident mistakes cause real harm, and inside the research pipeline, where methods like LLM-as-a-judge, synthetic data generation, and active learning rely on calibrated confidence without verifying it. Standard calibration metrics only require two inputs per example: a confidence score and a correctness judgment. Most benchmarks in use today already provide both, meaning calibration can be reported immediately. For open-ended generation, however, defining these two inputs is still an open challenge. We argue that each NLP subfield should pair its main performance metric with a calibration score and call for treating calibration as an essential property of every model rather than a niche topic.

Reasoning traces improve large language models (LLMs), but current models are trained to reason mostly in English. It has been shown that forcing a model to reason in another language degrades accuracy, even when the reasoning language matches the language of the prompt – but only for a setting where the model reasons over a short prompt. Here, we ask whether the same holds for retrieval-augmented generation (RAG), where the model must read and integrate a large amount of retrieved evidence in the target language. To study this, we build a fully monolingual German RAG question-answering testbed over the fictional world of the tabletop role-playing game The Dark Eye, a domain that is richly documented in German but too niche for the model to answer from memory, so that it has to rely on retrieval. Varying the forced reasoning language of an agentic RAG system on this testbed, we find that aligning the reasoning language with the language of the query and the retrieved documents helps. Forced German reasoning outperforms forced French, although the model benchmarks higher in French, so the benefit comes from alignment and not from language proficiency. The advantage grows when the retrieved context is richer and structure-aware. However, forced German only reaches the level of the model’s native, unconstrained English reasoning without surpassing it, showing that native multilingual reasoning is needed. We publicly release the testbed and QA benchmark.

Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template’s effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an “ownership bias” – models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model’s answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.

Should multilingual LLMs answer medical questions consistently across input languages, or adapt responses to cultural cues? Existing multilingual medical benchmarks usually assume that medically correct answers should remain consistent across languages and treat cross-lingual variation as model error. In contrast, cultural adaptation research argues that appropriate medical answers may legitimately differ across contexts. We review the multilingual medical NLP literature through these two perspectives, we identify three gaps: limited stakeholder perspectives (e.g., of medical professionals), a lack of empirical evidence on which approach better serves users, and no benchmarks capable of distinguishing universally correct from culture-specific cases. To address the first gap, we survey 356 participants across three stakeholder groups (medical, NLP, and anthropology professionals) in three countries (Germany, Spain, and the United States). Anthropologists consistently favor adaptation, while medical and NLP respondents remain divided, with notable divergence between U.S. and European medical professionals. LLMs prompted with profession and country personas fail to reproduce this variation, overestimating cross-lingual consistency preference among NLP and medical personas. We conclude that neither consistency nor adaptation can currently be considered clearly preferable, highlighting the need for empirical evidence on which approach better serves users across cultural contexts.

This paper presents the JGU Mainz submission to the WMT25 Shared Task on LLMs with Limited Resources for Slavic Languages: Machine Translation and Question Answering, focusing on Ukrainian, Upper Sorbian, and Lower Sorbian. For each language, we jointly fine-tune a Qwen2.5-3B-Instruct model for both tasks with parameter-efficient finetuning. Our pipeline integrates additional translation and multiple-choice question answering (QA) data. For Ukrainian QA, we further use retrieval-augmented generation. We also apply ensembling for QA in Upper and Lower Sorbian. Experiments show that our models outperform the baseline on both tasks.

In-context learning (ICL) has transformed the use of large language models (LLMs) for NLP tasks, enabling few-shot learning by conditioning on labeled examples without finetuning. Despite its effectiveness, ICL is prone to errors, especially for challenging examples. With the goal of improving the performance of ICL, we propose corrective in-context learning (CICL), an approach that incorporates a model’s incorrect predictions alongside ground truth corrections into the prompt, aiming to enhance classification accuracy through self-correction. However, contrary to our hypothesis, extensive experiments on text classification tasks demonstrate that CICL consistently underperforms standard ICL, with performance degrading as the proportion of corrections in the prompt increases. Our findings indicate that CICL introduces confusion by disrupting the model’s task understanding, rather than refining its predictions. Additionally, we observe that presenting harder examples in standard ICL does not improve performance, suggesting that example difficulty alone may not be a reliable criterion for effective selection. By presenting these negative results, we provide important insights into the limitations of self-corrective mechanisms in LLMs and offer directions for future research.

Large language models (LLMs) are powerful zero- and few-shot learners. However, when predicting over a set of candidate options, LLMs suffer from label biases, and existing calibration methods overlook biases arising from multi-token class labels. We tackle an issue we call label length bias, where labels of different lengths are treated inconsistently, even after standard length normalization. To mitigate it, we propose normalized contextual calibration (NCC), an effective method that normalizes and calibrates predictions at the full-label level. NCC achieves statistically significant improvements over prior approaches across multiple datasets and models, with gains of up to 10% F1. Moreover, NCC extends bias mitigation to broader tasks such as multiple-choice question answering. Our analysis shows that, when combined with in-context learning, NCC is less sensitive to few-shot example selection, requires fewer examples for competitive performance, and produces more reliable confidence estimates. These findings highlight the importance of mitigating full-label biases to improve the performance and robustness of LLM-based methods, particularly in real-world applications where class labels naturally consist of multiple tokens.

When evaluating large language models (LLMs) with multiple-choice question answering (MCQA), it is common to end the prompt with the string “Answer:” to facilitate automated answer extraction via next-token probabilities. However, there is no consensus on how to tokenize the space following the colon, often overlooked as a trivial choice. In this paper, we uncover accuracy differences of up to 11% due to this (seemingly irrelevant) tokenization variation as well as reshuffled model rankings, raising concerns about the reliability of LLM comparisons in prior work. Surprisingly, we are able to recommend one specific strategy – tokenizing the space together with the answer letter – as we observe consistent and statistically significant performance improvements. Additionally, it improves model calibration, enhancing the reliability of the model’s confidence estimates. Our findings underscore the importance of careful evaluation design and highlight the need for standardized, transparent evaluation protocols to ensure reliable and comparable results.

Large Language Models (LLMs) have demonstrated capabilities for natural language formulations of molecular property prediction tasks, but little is known about how performance depends on the representation of input molecules to the model; the status quo approach is to use SMILES strings, although alternative chemical notations convey molecular information differently, each with their own strengths and weaknesses. To learn more about molecular string representation preferences in LLMs, we compare the performance of four recent models—GPT-4o, Gemini 1.5 Pro, Llama 3.1 405b, and Mistral Large 2—on molecular property prediction tasks from the MoleculeNet benchmark across five different molecular string representations: SMILES, DeepSMILES, SELFIES, InChI, and IUPAC names. We find statistically significant zero- and few-shot preferences for InChI and IUPAC names, potentially due to representation granularity, favorable tokenization, and prevalence in pretraining corpora. This contradicts previous assumptions that molecules should be presented to LLMs as SMILES strings. When these preferences are taken advantage of, few-shot performance rivals or surpasses many previous conventional approaches to property prediction, with the advantage of explainable predictions through chain-of-thought reasoning not held by task-specific models.

This paper presents JGU Mainz’s winning system for the BLP-2025 Shared Task on Code Generation from Bangla Instructions. We propose a multi-agent-based pipeline. First, a code-generation agent produces an initial solution from the input instruction. The candidate program is then executed against the provided unit tests (pytest-style, assert-based). Only the failing cases are forwarded to a debugger agent, which reruns the tests, extracts error traces, and, conditioning on the error messages, the current program, and the relevant test cases, generates a revised solution. Using this approach, our submission achieved first place in the shared task with a Pass@1 score of 95.4. We also make our code public.