Grounding an LLM against a knowledge base looks solved until you look at the tail. This EMNLP 2026 paper (arXiv) makes the tail measurable: instead of ranking entity rarity by Wikipedia pageviews, it scores rarity with knowledge-graph structural metrics — how well-documented and connected an entity actually is. That reframing surfaces rare entities popularity never flags, and on those slices state-of-the-art accuracy drops 15.4–39.9%. If your eval buckets retrieval quality by traffic, you’re averaging over exactly the queries that break in production.
The result I keep coming back to is the ablation. A reasoning-capable vision-language model that iteratively searches and reasons over Wikipedia beats prior state of the art by 6.9% overall and up to 23.3% on rare slices — but the pieces don’t stand alone. Reasoning by itself barely moves rare-entity accuracy. Retrieval without reasoning lifts the rare cases but hurts overall accuracy. Only the combination wins.
That maps cleanly onto how I’ve watched RAG systems fail. Teams bolt a reasoning model onto a retriever and expect it to “figure it out,” or they crank retrieval recall and quietly regress on the head. The paper’s loop — retrieve, then reason over what you pulled, iterate — is the pattern that holds, and it’s training-free, so you can test it against your own KG grounding this week. The HF paper page has the MERLIN benchmark across five languages (Hindi, Indonesian, Japanese, Tamil, Vietnamese) if you want to see how much wider the gap opens in the multilingual case.
The open question for me: does KG-connectedness-as-rarity carry over to enterprise knowledge graphs, where “documented and connected” really measures how much your own org has curated an entity — or does every schema need its own rarity metric before the slice means anything?