nucleation

LIT-1820: Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops

3 min read

Properties4
StatusActive — read in full 2026-10-29 (NOTE-1331), after the owner brought it to the exchange "Paper assessment" asking whether it is "on to something or if this is AI slop nonsense". Worth keeping for its conditional theorem, which is sound and is QUESTION-130's construction carried further, and for its measured caution about its own experiments. Its framing outruns both.
Tagscompositionality, mathematics, network-science, representation-learning, anthology-candidate
Cited as “paper” byNOTE-1331 — Sarkar, Mukherjee and Sato — Attention as lattice gauge theory
Cited as “source” byTHEORY-1140 — If every entity's O(d) frame may change independently and relation transports change as Tᵢⱼ ↦ gᵢTᵢⱼgⱼ⁻¹, then functions of the transports alone carry no invariant information except around cycles, a triangle's invariants are exactly the class functions of its holonomy H = TᵢⱼTⱼₖTₖᵢ, and an equivariant update of an edge from one cycle through it is Ψ(H)Tᵢⱼ with Ψ conjugation-equivariant; the symmetry fixes a family of updates, not one mechanism, and says nothing about whether a task has the symmetry, THEORY-1158 — On synchronization of S4 relations with outliers, an untrained mechanism that composes relations through third vertices gains 26–45 points over the single direct path, and weighting the candidate paths by the Wilson loops of squares that avoid the missing edge adds 0.6–7.8 points, most where clean two-hop paths become a minority; it stays below spectral synchronization with one reweighting pass on every setting and within 2.2 points of plain spectral synchronization either way; its 0.99 AUC for corrupted triangles is the ceiling that exactly flat clean data set, not a measure of detection power

Graph View

Sarkar, A., Mukherjee, A., & Sato, M. (2026). Attention as lattice gauge theory: Relations interact only through Wilson loops. ML4PS 2026 workshop, NeurIPS 2026. ARXIV-2610.37656

Key takeaways

  • Under per-entity O(d) frame freedom, fixing the gauge along a spanning tree shows that the transports carry no invariant information except around cycles; a triangle’s invariants are the class functions of its holonomy H = TᵢⱼTⱼₖTₖᵢ (THEORY-1140).
  • An equivariant update of an edge from one cycle through it is Ψ(H)Tᵢⱼ with Ψ(gHg⁻¹) = gΨ(H)g⁻¹. Ψ is a family, not one function, so “the unique form the symmetry permits” fixes a design space, not a mechanism.
  • One loop through a corrupted edge detects the corruption but cannot say which of its edges is wrong; scoring a candidate by squares that avoid it can (THEORY-1158).
  • On the S4 benchmark, composition through cycles is the capability and holonomy routing a small refinement; Singer’s spectral method ties or edges it, as the paper says.

Standing in the record

The construction is the one QUESTION-130 wrote down on 2026-10-18 for evaluative transport: local frames, covariant transports, holonomy around a loop, conjugacy class as the invariant. The paper adds the forced form of an equivariant update. It does not touch the question’s open part: its transports are given, as they are in every formalism QUESTION-282 found.

The paper’s premise is weaker than its framing. Its companion paper (LIT-1803) calls per-entity frame freedom “an inductive bias, not a claim that fresh names logically imply a physical symmetry” (p. 2); this paper calls it “the task’s, not a network’s” (p. 2). The companion’s own CLUTRR result shows orthogonal transports costing 64.6 points out of distribution, where its StepGame result shows them helping, so whether the premise holds is a property of the task.

anthology-candidate: an architecture paper in machine learning, which the anthology may want as well. It is filed here for its mathematics and for what it bears on in QUESTION-130 and CASE-179.

Note of 2026-10-30: plain spectral synchronization trails holonomy at p = 0.45

“Singer’s spectral method ties or edges it”, in the takeaways above, and the summary’s “it does not beat spectral synchronization” hold against spectral synchronization with one reweighting pass, which is ahead on every row of Table 1. They do not hold against plain spectral synchronization. Holonomy weighting beats it on three of four settings: 0.988 against 0.987, 0.937 against 0.933, and at p = 0.45 0.700 against 0.678. It trails only on the sparse graph, 0.886 against 0.897 (NOTE-1331’s table). Found by the status review of this date in narrowing THEORY-1158.

Backlinks