Simon Razniewski

Research agenda

I investigate what language models know, how that knowledge can be measured and structured, and how it can be used to build more reliable knowledge systems. I am Professor of Knowledge-Aware AI at ScaDS.AI Dresden/Leipzig and TU Dresden and lead the KAAI research group.

My work combines empirical analysis with method and system building across language models, knowledge bases, factuality, knowledge extraction and consolidation. Explore our research, browse selected publications, or contact me about collaboration.

Selected current work

Epistemic twins

Epistemic twins are a core KAAI idea: discrete, comprehensive and comprehensible models of the factual knowledge stored inside an LLM. By building epistemic twins, we enable a symbolic science of language model knowledge.

GPTKB and LLMpedia are the two flagship projects that turn this idea into public research artifacts.

GPTKB

GPTKB transforms parametric LLM knowledge into knowledge bases in triple form, providing a basis for systematic statistical analysis, querying, and comparison of LLM knowledge. The current version, GPTKB 2.0, extracts more than 38 M triples from GPT 5.1.

LLMpedia

LLMpedia transforms parametric LLM knowledge into text form, in the style of Wikipedia, providing a richer and more natural format for human readers, albeit less accessible to statistical tools. The current version, LLMpedia, builds more than 1 M articles from GPT-5-mini.

Biography and professional service

Simon Razniewski was previously a research scientist at the Bosch Center for AI (2023-2024), senior researcher at the Max Planck Institute for Informatics (2017-2021), where he was heading the Knowledge Base Construction and Quality area, and assistant professor at the Free University of Bozen Bolzano¹ (2014-2017). He holds a PhD from the Free University of Bozen-Bolzano (2014), and a Diplom (MSc.) from TUD Dresden University of Technology (2010; not this Dresden). He spent time as visitor at the Max-Planck Institute for Informatics (2016), the University of Queensland (2015), AT&T Labs-Research (2013), the University of California, San Diego (2012), and has previous industrial experience from Globalfoundries (2010) and Siemens IT (2009). Simon Razniewski has published >20 papers at premier² conferences in the area of artificial intelligence (and more than 110 papers in total).

Acquired grants (selection)

• DFG individual research grant of 312,000 € for compiling negative knowledge at web scale, 2021
• Diffbot grant of $30,000 for KB construction in fiction, 2020
• Google Cloud Credit grant of $5,000 for research on commonsense knowledge, 2019
• NVIDIA hardware grant (~$1,100) for research on information extraction, 2018
• Free University of Bozen-Bolzano, acquired projects grants Recall (20,000 €, 2016-17), TQTK (20,000 €, 2016-17), TaDaQua (50,000 €, 2016-18)
• Province of South Tyrol, open research call, coauthored project MAGIC (250,000 €, 2013-16)

Other grant proposals

• ERC consolidator grant 2025: Passed to stage 2, overall grade B
• ERC starting grant 2020: Proposal received grade A but not funded due to budget constraints

Organization

• Co-organizer, AKBC workshop at EMNLP 2026
• Co-organizer, Wikidata Workshop at ISWC 2026, 2025, 2023, 2022 and 2021
• Guest editor of SWJ special issue on Wikidata (2023), on commonsense knowledge (2021)
• Co-organizer, KBC-LM workshop at ISWC 2023
• Co-organizer, LM-KBC challenge at ISWC 2024, 2023 and 2022
• Website chair, ESSLLI 2016 summer school

Reviewing

• 2026: Area chair for ACL RR, PC member of ESWC, SemEval Task Proposals, ISWC Doctoral Consortium, ISWC P&D
• 2025: Area chair for ACL RR, senior PC member of IJCAI, PC member of ACL RR, AAAI, WSDM Industry track, RepL4NLP, KI, KG-Star@ESWC
• 2024: senior PC member of CIKM, PC member of ACL RR, ISWC P/D, IJCAI demos, KaLLM workshop@ACL, KI, AKR³ workshop, TGDK
• 2023: Area chair of EMNLP and EACL, senior PC member of CIKM, ESWC, PC member of ACL ARR, neurIPS, VLDB demos, IJCAI demos, LDK, VLDB tutorials, board member of TGDK, reviewer for TODS
• 2022: Senior PC member of WWW, CIKM, PC member of WSDM, ESWC, EMNLP, VLDB demos, reviewer for neurIPS
• 2021: Area chair of IJCAI, senior PC member of CIKM, PC member of WSDM, ACL, NAACL, EMNLP, ESWC, WWW, TKDE, CSKGs workshop@AAAI, LDK, QOD
• 2020: Senior PC member of CIKM, IJCAI, ISWC, PC member of WSDM, ACL, ESWC, QOD, Wikidata workshop@ISWC, SemDeep-6@IJCAI
• 2019: PC member of WWW, ACL, AAAI, NAACL, ISWC, K-CAP, AKBC, LDK, Data Quality in Wikidata workshop, Quality of Open Data workshop, Commonsense Inference workshop@EMNLP
• 2018: PC member of ACL, CIKM, EMNLP, DASFAA, SemDeep-4, DL4KGs, QOD, RoD. Reviewer for VLDB Journal
• 2017: PC member of CIKM, WebDB. Reviewer for SWJ
• 2011-2016: External reviewer for AAMAS 2016, SEBD 2015, CIKM 2014, COOPIS 2013, TbiLLC 2013, CIKM 2011, BNCOD 2011.

Grant reviewing

• • Niedersachsen.next
  • DFG (2026, 2025, 2024, 2023, 2020)
  • Jönköping University (2024)
  • GACR (2021)

Membership

• • ELLIS (since 2025)

Media appearances

• “Forscher aus Dresden baut Wikipedia-Klon und entlarvt, was ChatGPT wirklich kann,” DNN
• “Knowledge expansion for computers,” Max Planck Society, 2021
• “Was ist wichtig, was ist richtig,” Südtiroler Wirtschaftszeitung, April 2018 (in German)

Demo systems

• • Ascent (2021): Advanced semantics for commonsense knowledge

• • Wikinegata (2021): Negative knowledge for Wikidata

• • Entyfi (2020): Entity typing in fiction

• • Sandi (2020): Story-image alignment

• • Dice (2020): Joint reasoning for multifaceted commonsense knowledge

• • CounQER (2020): Counting queries and entity valued relations

• • ReCoin (2017): A user script for adding relative completeness annotations to Wikidata. Developed by Vevake Balaraman and Albin Ahmeti in the context of the TaDaQua project (Video in Danish)

• • COOL-WD (2017): A completeness tool for Wikidata. Developed by Radityo Eko Prasojo and Fariz Darari

• • MAGIC (2012): A tool for reasoning about the completeness of relational databases, developed by Ognjen Savkovic, Paramita Mirza and Alex Tomasi

• • Lookslikescanned: A website to make PDFs look appear like scanned

Other scripts and datasets

• • Cinex: Code, experiment data and SPARQL endpoint with Wikipedia results for counting quantifier extraction

• • Property Ranking: Dataset of 350 (entity, property1, property2) pairs for humans in Wikidata, along with a preference judgment [O18]

• • A dataset of about 2000 crowdsourced completeness assertions for YAGO and Wikidata.

Other

• • I love flying, I hold (inactive) glider and paraglider licenses, and once flew in the German Gliding Bundesliga (for Aeroclub Pirna), and the OLC Alps League (Aeroclub Bolzano). I also circled the globe four times as passenger, and was on world’s longest 2017 commercial flight (QR921).

• • Did you know that the extent of Indonesia (main islands only) at 5300 km is more than the distance from Germany to China, Kenya, or the North Pole (4630, 5200, 3920 km, respectively)?