Sarkar, A., Mukherjee, A., & Sato, M. (2026). Relational composition as parallel transport: What per-entity frames force in attention. Preprint. ARXIV-2610.34945
Key takeaways
- Per-entity frame freedom is adopted “as an inductive bias”, and the paper says it is not derived (pp. 2, 12–13).
- Under it, invariant scores are ϕ(‖xᵢ‖, ‖xⱼ‖, ⟨xᵢ, Tᵢⱼxⱼ⟩) and equivariant messages lie in span{xᵢ, Tᵢⱼxⱼ} (Theorem 1). A covariant transport computed from the endpoints alone has rank at most the endpoints’ rank (Theorem 2), so one vector per entity can never give an orthogonal transport: “relations must be supplied”.
- Which transports a task needs is “decided by the algebra of its relations” (p. 3). Orthogonality helps on StepGame and costs 64.6 points on CLUTRR, where the paper’s motivation is that a nonidentity idempotent has no invertible faithful representation (p. 9, App. B.4).
- With matched anchors B = QA, an orthogonal transport is identified only on the anchored subspace (p. 9).
Standing in the record
It is the source of the premise that LIT-1820 calls “the task’s”, and
its own CLUTRR control is the plainest evidence that the premise is
task-relative. That bears on CLAIM-098 and CLAIM-035, which hold that
translation is generally not a group action. Only the parts named in
NOTE-1326 were read. anthology-candidate, as for its companion.