Source
Xiong (2026), LIT-460 — ARXIV-2603.01227.
What was actually shown
The construction and its evidence are LIT-460’s: it derives the half-space lattice, with meet and join, from linear attribute directions plus thresholds, and reports LLM embeddings encoding that lattice on WordNet sub-hierarchies.
The construction is short enough to state completely.
The Linear Representation Hypothesis says an attribute is a direction in embedding space. Add a separating threshold and the attribute becomes a half-space. A concept defined by several attributes is then the intersection of their half-spaces — a region rather than a vector.
Regions order by inclusion, and the resulting partial order is a complete lattice: any two concepts have a greatest lower bound (meet) and a least upper bound (join). Meet is intersection of the half-spaces — conceptual refinement. Join is approximated by the conic hull of the combined attribute directions — conceptual generalisation. A canonical form exists when the attribute directions are linearly independent, and soft versions using fuzzy projection profiles handle embeddings that do not sit cleanly on one side of a threshold.
The empirical part: on WordNet sub-hierarchies, LLM embeddings are reported to encode these lattices and their logical structure.
The threshold is the whole move. A direction gives a scalar; a direction with a threshold gives a region; regions compose and scalars do not. That is why this says something the Linear Representation Hypothesis does not.
What this does not say
The evidence is the friendliest test available. WordNet is a hand-built ontology with exactly the clean hierarchical structure the hypothesis predicts. Recovering a lattice from concepts somebody constructed as a lattice is weak evidence that the lattice is in the embeddings rather than in the choice of concepts. The severe test — cluster a model’s own representations and look for the structure where nobody put it — is cheap and is not run.
It assumes the Linear Representation Hypothesis and does not re-establish it. Everything here is downstream of that premise, which the record now holds as THEORY-090. That account’s own evidence stops at the output space, and LIT-526’s null is against one of its uses.
Thresholds are assumed to exist and separate. The soft version concedes they do not, and the softening is a definitional choice rather than something derived.
Join is an approximation. The conic hull is the least subsuming region under that construction, not the least subsuming region.
Nothing downstream is demonstrated. The structure is shown, not used. The obvious application — constructing a steering vector for a compound concept by meet or join rather than searching for one — is available and untried, and until somebody runs it this is a description of a geometry rather than a capability.
And linear independence is a condition on the concepts, not the space. The canonical form holds for sets of attributes that happen to be independent; it says nothing about arbitrary concept collections.
Why explains: is empty
It underwrites no practice in this record and should not pretend to. The paper establishes a structure on a curated ontology; no recommendation about training, serving or evaluating follows from it yet. A finding that underwrites nothing yet is still a finding, and saying which is the honest version.
The trunk underneath
Both this and THEORY-036, filed in the same contribution, stand on the
Linear Representation Hypothesis. When they were filed the record held no
document for it and said so here. It now holds Park, Choe and Veitch
(LIT-606) and their account (THEORY-090), which this one
extends: that account makes an attribute a direction, and this one adds the
threshold. The lattice paper builds on that account’s causal inner product
(NOTE-210), under which separable attributes are orthogonal. That is what
makes the canonical form’s linear-independence condition more than a
convenience. The two accounts share their weak point: neither is tested
beyond curated concepts.