In my view, the conceptual contribution is too incremental and insufficiently significant for ICLR.
The pipeline combines familiar ingredients: local graph-consistency residuals, k-hop neighborhood overlap, greedy matching, connected-component cleanup, geodesic Voronoi labels, neighborhood averaging, and nearest-vertex snapping. The paper does not identify a new learning principle, optimization formulation, representation, or general mechanism arising from this combination. Relative to local structural methods such as 2D-GEM and HOPE, the main contribution appears to be a carefully engineered sequence of heuristics and thresholds rather than a clear conceptual advance. Such an engineering component could be valuable inside a broader geometry-processing system or as a focused workshop contribution, but I do not see sufficient novelty or general interest for the main track of ICLR.
The significance is further limited by the problem formulation and evaluation regime. The experiments concern full pointwise correspondence on established small mesh benchmarks, while more difficult and currently consequential aspects of shape correspondence (including substantial partiality, topology or genus changes, severe cross-category variation, and robust correspondence under realistic geometric corruption) are either excluded or only weakly represented. The paper itself states that large partiality and genus changes are outside scope. As a result, the work does not establish why this restricted refinement problem should be considered a timely ICLR contribution in 2027 rather than an incremental geometry-processing utility.
One important aspect I noticed is that the theoretical section does not validate the implemented algorithm, and the paper/code specification contains several inconsistencies. I took my time going through the code and theorems, and whereas the propositions are mostly correct, they concern elementary or idealized facts: row equality makes a residual vanish; nearest-anchor assignment has a pointwise geodesic property but does not prove label contiguity; greedy matching is a standard 1/2-approximation for score weight; and symmetric averaging lowers a Dirichlet energy. The reported method, however, uses asymmetric directed weights followed by discrete snapping, so the Dirichlet result explicitly does not apply. Cleanup uses approximate heat distances and does not modify the correspondence, and its proposition explicitly does not establish a connected output label. The matching result controls only the score of the selected matching within the movable set B, not geodesic accuracy, residual decrease, global injectivity, or the collision-containing completed map. Collectively, the results provide no recovery, convergence, monotonicity, stability, or final-contiguity guarantee for LOGOS.
There are also concrete specification mismatches. Appendix B.7 says zero-score edges are included in the greedy scan, whereas both the Python and CUDA implementations filter them out. A positive-score column can consequently remain unmatched after its positive rows have been consumed, after which the implementation's argmax maps the all-zero assignment column to source index 0. The implementation uses a strict greater-than test where Algorithm 1 says "at or above" the polishing quantile, implements the residual denominator with a clamp rather than the stated addition of delta, and uses max(1,T//2) polishing steps rather than ceil(T/2) for nondefault odd T. The hop-shell definition also needs an explicit zero-diagonal convention to exclude two-step return walks. Most of these discrepancies are individually repairable and do not necessarily invalidate the stored results at T=32, but together they weaken confidence that Algorithm 1 is a precise, self-contained description of the released method.