TLDR:
we are releasing an open corpus and toolkit that teach AI to explore Wikidata step by step, and to show its sources, built by Pleias and funded by Wikimedia Switzerland!

Ask a museum curator what they would like to know from Wikidata and the answer is rarely a SPARQL query. It is a question: which works in our collection were made by artists born in the same city? Where do the English and Chinese records disagree? Wikidata could answer many of these questions. Today, almost nobody outside a small technical community can ask them.

The first phase of AI-BRIDGES, a research programme at the Digital Humanities Research Hub, University of London, sets out to change that. Funded by the European Commission’s Marie Curie Postdoctoral Research Fellowship Program, additionally by Wikimedia Switzerland and delivered by French AI lab Pleias, AI-BRIDGES releases three open resources: a Synthetic Reasoning Corpus of AI exploration traces over Wikidata, the open-source harness that produced them, and a technical report documenting the method and its limits.

Look, don’t remember

Most attempts to connect Large Language Models (LLMs) to knowledge bases pour as much data as possible into the model’s context window and hope it copes. It doesn’t, for long. Accuracy drops and the model begins quoting statements it never actually read.
AI-BRIDGES and Pleias took a different approach — using a Recursive Language Model (RLM), the system explores Wikidata one step at a time, following links, reading what it finds and citing the exact statements behind its answer. Working across a six-language snapshot of Wikidata (English, French, German, Chinese, Arabic and Russian), an open-weight model running on a single GPU answered 74 of 100 evaluation questions correctly, the best score in the evaluation, against 61 of 100 for a leading commercial model in the same harness, and 49 for that same model using conventional tool calling.

The results are reported with their difficulty on show: 46 of 50 on single-step questions, 28 of 50 on the hardest multi-hop ones. Even there, the system reaches the right entity in 39 of 50 runs, against 27 under conventional tool calling, so the gain comes from carrying the search further. As Dr. Anastasia Stasenko, CEO at Pleias, explains:

“Wikidata is the largest open knowledge graph in the world, and asking it a complex question still takes someone who can write SPARQL. We have shown that an open model, running on an institution’s own hardware, can explore the graph step by step, quote its evidence, and answer these questions better than a leading commercial system. Knowledge that volunteers built in the open should not have to be rented back to them through a proprietary model.”

A gift to Wikidata editors

Walking the graph this carefully turned up real data-quality issues as a by-product: a film dated 2005 in three languages and 2006 in a fourth, nationality recorded in descriptions but missing from statements, a source link pointing to a book about fish. None of it was edited automatically. The agent flags and cites; the community decides. As Ilario Valdelli, Innovation Programme Lead at Wikimedia Switzerland shares:

“How AI systems treat open knowledge is being decided now. By funding work that is released openly, with its methods and limitations visible, we make sure the Wikimedia movement is shaping that answer rather than waiting for it.”

What comes next

Phase 2 will use the corpus and harness to train a family of open Wikidata search models at three sizes, small enough to run and open enough to inspect. The goal is simple to state: a natural-language question in, a cited and checkable answer out, with no SPARQL and no commercial API required. A third phase, planned for 2027, will put the models in front of humanities and GLAM scholars working on real research questions. Dr. Shani Evenstein Sigalov, lead researcher and director of AI-BRIDGES shares:

“What I care about is that people with real questions can get verifiable, cited answers, can tell when an answer is wrong, and do it in an open platform that serves the public, rather than the interest of for-profit companies. That’s what this technical development phase set out to test, and the results show it can be done in the open. Phase 2 will turn it into something people can actually use.”

Get involved!

AI-BRIDGES is collecting real questions from people working in galleries, libraries, archives, museums and the Humanities, the questions you would ask Wikidata if the technical barrier disappeared. They will shape how Phase 2 is evaluated. Please add queries directly to the doc, or send them to: contact@ai-bridges.org.

Access the release

* Harness, code and benchmark (GitHub): https://github.com/aibridges/RLM_Wikidata

* Corpus (Hugging Face): https://huggingface.co/datasets/AI-BRIDGES/wikidata-search-traces

* Technical report (ArXiv):
https://arxiv.org/abs/2610.06650

Acknowledgements and thanks

We thank the European Commission’s Marie Curie Postdoctoral Fellowship program for supporting AI-BRIDGES; Wikimedia Switzerland for funding this phase, Pleias for delivering it so brilliantly, and Wikimedia Deutschland for their ongoing support of the programme. We also thank our additional core-partners, Wikimedia UK & Wikimedia Brasil, as well as additional partners: individuals, organizations and institutions supporting the AI-BRIDGES efforts and being part of its network. A special thank you is dedicated to Prof. Jane Winters, Director of the Digital Humanities Research Hub, University of London, for her ongoing support. 

About AI-BRIDGES

AI-BRIDGES is a research programme at the Digital Humanities Research Hub, School of Advanced Study, University of London, creating bridges between institutional data, Linked Open Data platforms like Wikidata & Wikibase, and AI systems.