Thanks @algomagic for this lovely illustration!
How should AI compose relations? If all we have are rotations or reflections between independent local frames, symmetry allows interaction only through closed loops.
Accepted at ML4PS 2026, happy to chat!
https://t.co/oF5bnoZOA6
Attention as Lattice Gauge Theory: A new framework proves that edge attention via Wilson loops is the unique form permitted by symmetry. It achieves 0.88 accuracy on S4 relations with zero training, vastly outperforming commutative architectures. https://t.co/setqca1hx2
Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops
https://t.co/15CM57KIZL [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸 𝚙𝚑𝚢𝚜𝚒𝚌𝚜.𝚌𝚘𝚖𝚙-𝚙𝚑]
💬Accepted at the 9th Workshop on Machine Learning and the Physical Sciences (ML4PS 2026), co-located with NeurIPS 2026
Relational Composition as Parallel Transport: What Per-Entity Frames Force in Attention
Aakash Sarkar, Ananya Mukherjee, Motoshige Sato
https://t.co/0pUlRusg4p [𝚌𝚜.𝙻𝙶 𝚌𝚜.𝙰𝙸 𝚌𝚜.𝙵𝙻]
Our companion paper views standard attention as the shared-frame (flat) case. Independent frames require transport before comparison - pointing toward relation operators learned from context.
https://t.co/scfyb2ZaqY with @always_ananya & @Motoshige_Sato.
https://t.co/wOqDmi6PV0
Relational Composition as Parallel Transport: What Per-Entity Frames Force in Attention
Per-entity coordinate symmetry constrains attention, sets width limits for exact relation composition, and improves multihop reasoning.
https://t.co/BqbhWzcP7k
Thanks @algomagic for this lovely illustration!
How should AI compose relations? If all we have are rotations or reflections between independent local frames, symmetry allows interaction only through closed loops.
Accepted at ML4PS 2026, happy to chat!
https://t.co/oF5bnoZOA6
Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops
Independent entity-frame rotations force edge attention through closed-loop holonomies; the resulting model infers missing relations and flags corrupted triangles.
https://t.co/AyJ7RimPip
I just read Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops. If every node keeps an independent local frame, the math forces relational attention through Wilson loops: the only gauge-invariant triangle score is a class function of H = Tij Tjk Tki, and the only equivariant edge update is U = Ψ(H)Tij. Their no-training holonomy attention infers unseen S4 relations on sparse graphs at 0.88 with 20% edges corrupted, slightly below specialist synchronization algorithms; commutative architectures cap at 0.083. It flags corrupted triangles at AUC 0.99 versus 0.76 for commuting scores.
Discord https://t.co/HzZUF4ElGB
#gaugetheory #holonomyattention
Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops
Independent entity-frame rotations force edge attention through closed-loop holonomies; the resulting model infers missing relations and flags corrupted triangles.
https://t.co/AyJ7RimPip
Relational Composition as Parallel Transport: What Per-Entity Frames Force in Attention
Per-entity coordinate symmetry constrains attention, sets width limits for exact relation composition, and improves multihop reasoning.
https://t.co/BqbhWzcP7k