Dear Wiki Research community, Sharing initial results from the first phase of technical developemtn for AI-BRIDGES here as well, as it might be of interest to people.
Best, Shani. ---------- Forwarded message --------- From: Shani Evenstein Sigalov <[email protected]> Date: Mon, Oct 12, 2026 at 4:07 PM Subject: [Release] Synthetic Reasoning Corpus and harness for natural-language exploration of Wikidata (CC0) To: Wikimedia-l <[email protected]> Hello all, I'm happy to share the first open release from AI-BRIDGES <https://ai-bridges.org>, funded by Wikimedia Switzerland <https://meta.wikimedia.org/wiki/Wikimedia_CH/Grant_apply/AI-BRIDGES:_Synthetic_Reasoning_Corpus> and delivered by Pleias <https://pleias.ai/>, with support from Wikimedia Deutschland. We built a harness that lets a language model explore a frozen, six-language Wikidata snapshot step by step, reading entities and citing the statements it used, rather than stuffing large amounts of data into its context, taking a Recursive Language Model approach. From it we generated a Synthetic Reasoning Corpus of exploration traces, released under CC0. A few results from the technical report: - 74 of 100 questions answered correctly by Qwen3.8-27B, an open-weight model on a single GPU, vs 61 for gpt-6-luna in the same harness - 10,235 solving traces released under CC0, over single-entity and multi-hop questions - The harness lifts both models we ran under both interfaces, with the gain mostly on multi-hop questions Of particular interest to editors: exploration surfaced cross-language inconsistencies and misplaced information (for example, nationality appearing in descriptions but not in statements). Nothing was edited automatically; the findings are in the report for community review. - 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: https://arxiv.org/abs/2610.06650 Next, we plan to train open Wikidata search models on this corpus and we would appreciate your help: if you work with GLAMs, Education, Research, especially in the Humanities, please send us real questions you'd want to ask Wikidata in plain language or in SPARQL <https://docs.google.com/document/d/1ipIzwSwf-xCtU7R7ewNCCND7vk0yha6byWMNs3OyIEc/edit?tab=t.0>, or via email: [email protected]. I'd like to take a moment to thank not just Pleias, WMCH & WMDE, but also our other core partners, WMUK & WM Brasil, as well as other partners -- organizations, institutions and so many wonderful individuals from this community -- for being part of the AI-BRIDGES journey and supporting it. Feedback, questions and critique are all very welcome on-list, on wiki <https://meta.wikimedia.org/wiki/Research:AI-BRIDGES> or privately. All best wishes, Shani. ----------------------- Dr. Shani Evenstein Sigalov Marie Curie Postdoctoral Research Fellow & AI-BRIDGES <https://ai-bridges.org/> lead Digital Humanities Research Hub, School of Advanced Study, University of London. _______________________________________________ Wiki-research-l mailing list -- [email protected] To unsubscribe send an email to [email protected]
