Xiong (2026), ICLR 2026 — ARXIV-2603.01227
Standing in the record
Held in both records under ADR-013. The anthology holds it as ANTH-LIT-460, read for concept geometry, with its account ANTH-THEORY-034. It is filed here at the owner’s request of 2026-09-27, to be re-read against van Rijsbergen’s The Geometry of Information Retrieval (LIT-262). Formal concept analysis, Galois connections and the logic of lattices of classes are that book’s second chapter, and questions this record asks that the anthology does not.
It was filed Deferred, unread. NOTE-240 is the close reading of 2026-09-27, and it placed the work: Proposed — promising as a bridge between probe geometry and formal concept analysis, and unproven as a hypothesis about LLMs. It would be settled by three things. The attributes would have to come from something other than another LLM’s annotations. The region-level join would have to be defined correctly, as the Galois closure of the shared intent. And lattice identities (closure, absorption, meet/join agreement with the recovered context) would have to be tested, not pairwise subsumption and candidate ranking alone.
A derivation of this paper’s premise has since entered the record. THEORY-185 (Korchinski et al., LIT-863) shows that when binary attributes multiply co-occurrence independently, each is a linear direction of the PMI embedding and thresholding recovers the incidence exactly. But that independence makes the incidence the full hypercube and its concept lattice Boolean, so it supports the step this paper assumes and not the lattice it claims. Whether linear directions survive attributes that imply one another is QUESTION-025.