»Forward Pass.

10 papers, 1 industry update — Two papers probe TFM reliability from opposite ends: TabPFN/TabICL/TabDPT all fail basic Bayesian self-consistency checks, while frontier LLMs uniquely lose accuracy as tabular input dimensionality grows.

10 papers, 1 industry update — A 9-model benchmark (TabPFN v2/v2.5/v2.6/v3, TabFM, Mitra, TabICL/v2, LimiX) finds all tabular FMs degrade under distribution shift, with real-data pretraining helping in-distribution fit more than true OOD robustness.

9 papers — TabPFN's internal representation topology (persistent homology) tracks when its in-context predictions become unreliable.

8 papers, 1 industry update — Two papers push TabPFN into new territory: as an in-context head for discrete-choice/marketing panels (2607.13314) and as a calibration-boosting classifier head on frozen multimodal embeddings (2607.11007).

1 paper — Quiet day for tabular ML: 49 candidates screened, only one cleared the relevance bar.

2 papers — GRAFT reframes table retrieval in data lakes as graph matching, gaining ~8-11% over prior retrievers on Spider/BIRD.

4 papers — Graph foundation models are borrowing tabular ICL backbones: GTAlign aligns graph structure to a TabPFNv2.5/LimiX-16M input format for in-context node/graph classification.