Mechanistic interpretability analysis reveals serialization order confounds spatial attention in MeshXL, suggesting attention tracks sequence recency rather than 3D geometry.
Autoregressive mesh transformers look like an unusually convenient target for mechanistic interpretability: the token stream is the geometry, so every attention weight has a literal spatial address and can be painted onto the surface being generated. We show that this convenience is a trap. In MeshXL's canonical serialization, faces arrive sorted by height, so a face's index in the token stream predicts its vertical position at Spearman rho = 0.951 across all 2820 training chairs, and a cubic in normalized sequence position explains 86% of height variance. Any attention map read spatially is therefore confounded with plain recency. The confound is structural rather than incidental: it is a property of the preprocessing, its median across four object categories ranges only from +0.914 to +0.968, and re-ordering faces to break it raises cross-entropy fortyfold on every mesh tested. Controlling for sequence position on unmodified inputs, attention's partial correlation with 3D proximity is +0.194 (95% CI [+0.165, +0.206]) against +0.430 for sequence proximity; the spatial estimate replicates at +0.180 ([+0.167, +0.199]) on a disjoint held-out sample. No head in the network is primarily geometric, none reaching rho_spatial = 0.50 while rho_seq reaches +0.951, and layers 0-11 carry no geometric selectivity at all. Testing mesh topology, a geometric feature nearly independent of the confound, attention does not track it (+0.041, no head above +0.15), while sensitivity to repeated coordinate tokens between non-touching faces is twice as strong. Separately, MeshXL inherits OPT-1.3b's transformer body and reinitializes only its embeddings; we find its attention sinks bear no relation to OPT's (rho = -0.146) while being sevenfold stronger, so they form under mesh training rather than arriving with the weights. All confirmatory claims follow rules fixed before any attention was extracted. Of three such arms, one returned inconclusive, one failed its own validity gate, and one was refuted; all three are reported as they landed.
No takes yet. Share an insight, caveat, or question.
Minwoo Choi (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: