We cordially invite you to attend the symposium and network meeting for our research project ‘Constrained Retrieval Augmented Language Models (CORAL)’ in March 2026, entitled:

Next Level Language Models
How language models are becoming more efficient, legally compliant and reliable.
Symposium and network meeting of the CORAL project

Location & time:
17–18 March 2026 (lunch-to-lunch plus optional workshops)
in Leipzig

Aims and schedule of the event:
In addition to presentations by project participants on the results of CORAL's work to date (see below for details), there will be short presentations (pitches) by the network partners. We can also look forward to keynotes by Leonie Weißweiler and Daniel Steinigen. On the afternoon of 18 March, there will be additional workshops on the generation, evaluation and use of language models.

Agenda CORAL network meeting 2026

Registration:
Please register by 12 March 2026. Please note that the Leipzig Book Fair 2026 will begin on Thursday, 19 March, which may be of interest for individual travel planning.

Language policy:
The conference languages are English and German. Most of the content will be presented in English, and materials will mainly be available in English. Data sets and the performance of language models will be considered primarily or at least equally for the German language. The titles of the presentations and workshops indicate the primary language of interaction. Questions may be asked in either German or English. No translation service will be provided. For discussions and interactions, we recommend an approach that is as inclusive as possible based on the composition of the audience on site.

Background:
The CORAL project explores methods for developing, evaluating and using language models (LLMs) under legal, technical and qualitative constraints. The focus is on the traceability of generated texts and the use of protected data through constrained training and retrieval-augmented generation. CORAL uses data from the German National Library, the Internet Archive, Common Crawl and the Leipzig Vocabulary, as well as partner-specific data, e.g. from the financial sector. The aim is to develop legally sound methods for obfuscating and using this data. Research questions concern robust training methods, resource-efficient models, effective obfuscation and the transparency of generated texts. The CORAL project is funded by the BMFTR (grant number 16IS24077[A-D]) from October 2024 to September 2027. It strengthens the establishment of a German market for language models.
Further information: https://coral-nlp.github.io/