@inproceedings{nagl-etal-2026-benger,
title = "{B}en{GER}: Benchmarking {LLM} Systems on Subsumption-Based Legal Reasoning in {G}erman Law",
author = "Nagl, Sebastian and
Mayrhofer, Ann-Kristin and
Heidebach, Martin and
Ko{\c{c}}ak, Aleyna and
Zettelmeier, Anne and
Breu, Elly and
Greiner, Angelina and
Milijas, Sofija and
Grabmair, Matthias",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1266/",
doi = "10.18653/2026.findings-emnlp.1266",
pages = "26759--26794",
ISBN = "979-8-89176-462-0",
abstract = "We introduce BenGER (Benchmark for German Law), a benchmark and dataset for subsumption-based legal reasoning in German law. It combines 596 exam-style free-text case tasks across several levels of legal education with 531 short doctrinal reasoning tasks, and a controlled subset adds timed human solutions written both unaided and in human-AI co-creation. We evaluate 12 contemporary LLM systems (closed flagship, efficiency-oriented, and open-weight) with a rubric-aligned LLM-as-a-Judge, which we cross-validate against a multi-rater human layer (three blind reviews per solution) and across six LLM judges. Our results show that closed-flagship systems lead across all three corpora, that human-AI co-creation measurably improves on unaided human work, and that the judge tracks human grading at Pearson r=0.76 and Cohen{'}s {\ensuremath{\kappa}}=0.60, with rankings stable across judges and two judges from independent providers clearing the Calderon single-reviewer replacement bar on human-authored solutions."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nagl-etal-2026-benger">
<titleInfo>
<title>BenGER: Benchmarking LLM Systems on Subsumption-Based Legal Reasoning in German Law</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sebastian</namePart>
<namePart type="family">Nagl</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ann-Kristin</namePart>
<namePart type="family">Mayrhofer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Martin</namePart>
<namePart type="family">Heidebach</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aleyna</namePart>
<namePart type="family">Koçak</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anne</namePart>
<namePart type="family">Zettelmeier</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Elly</namePart>
<namePart type="family">Breu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Angelina</namePart>
<namePart type="family">Greiner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sofija</namePart>
<namePart type="family">Milijas</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matthias</namePart>
<namePart type="family">Grabmair</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>We introduce BenGER (Benchmark for German Law), a benchmark and dataset for subsumption-based legal reasoning in German law. It combines 596 exam-style free-text case tasks across several levels of legal education with 531 short doctrinal reasoning tasks, and a controlled subset adds timed human solutions written both unaided and in human-AI co-creation. We evaluate 12 contemporary LLM systems (closed flagship, efficiency-oriented, and open-weight) with a rubric-aligned LLM-as-a-Judge, which we cross-validate against a multi-rater human layer (three blind reviews per solution) and across six LLM judges. Our results show that closed-flagship systems lead across all three corpora, that human-AI co-creation measurably improves on unaided human work, and that the judge tracks human grading at Pearson r=0.76 and Cohen’s \ensuremathąppa=0.60, with rankings stable across judges and two judges from independent providers clearing the Calderon single-reviewer replacement bar on human-authored solutions.</abstract>
<identifier type="citekey">nagl-etal-2026-benger</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1266</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1266/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>26759</start>
<end>26794</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T BenGER: Benchmarking LLM Systems on Subsumption-Based Legal Reasoning in German Law
%A Nagl, Sebastian
%A Mayrhofer, Ann-Kristin
%A Heidebach, Martin
%A Koçak, Aleyna
%A Zettelmeier, Anne
%A Breu, Elly
%A Greiner, Angelina
%A Milijas, Sofija
%A Grabmair, Matthias
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F nagl-etal-2026-benger
%X We introduce BenGER (Benchmark for German Law), a benchmark and dataset for subsumption-based legal reasoning in German law. It combines 596 exam-style free-text case tasks across several levels of legal education with 531 short doctrinal reasoning tasks, and a controlled subset adds timed human solutions written both unaided and in human-AI co-creation. We evaluate 12 contemporary LLM systems (closed flagship, efficiency-oriented, and open-weight) with a rubric-aligned LLM-as-a-Judge, which we cross-validate against a multi-rater human layer (three blind reviews per solution) and across six LLM judges. Our results show that closed-flagship systems lead across all three corpora, that human-AI co-creation measurably improves on unaided human work, and that the judge tracks human grading at Pearson r=0.76 and Cohen’s \ensuremathąppa=0.60, with rankings stable across judges and two judges from independent providers clearing the Calderon single-reviewer replacement bar on human-authored solutions.
%R 10.18653/2026.findings-emnlp.1266
%U https://aclanthology.org/2026.findings-emnlp.1266/
%U https://doi.org/10.18653/2026.findings-emnlp.1266
%P 26759-26794
Markdown (Informal)
[BenGER: Benchmarking LLM Systems on Subsumption-Based Legal Reasoning in German Law](https://aclanthology.org/2026.findings-emnlp.1266/) (Nagl et al., Findings 2026)
ACL
- Sebastian Nagl, Ann-Kristin Mayrhofer, Martin Heidebach, Aleyna Koçak, Anne Zettelmeier, Elly Breu, Angelina Greiner, Sofija Milijas, and Matthias Grabmair. 2026. BenGER: Benchmarking LLM Systems on Subsumption-Based Legal Reasoning in German Law. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 26759–26794, Budapest, Hungary. Association for Computational Linguistics.