Bayesian model selection couples a discrete model indicator with a model-specific continuous parameter space. We introduce structured dimension-matched variational transdimensional inference (SM-VTI) for finite enumerable model spaces. A rooted construction graph expresses a model as a sequence of local stop/child decisions. Each typed edge compiles a declared scientific parent-child edit into an exact native-coordinate dimension-matching lifting; an edge-conditioned flow then learns the residual continuous transport. The resulting local policy and conditional flow define one direct joint variational distribution, without embedding every model in a saturated maximum-dimensional surrogate. We derive its exact path density and optimize the joint reverse-KL objective. On a controlled 15-model target, SM-VTI-Joint recovers terminal masses, local actions, and nonlinear conditional geometry. On a 128-model misspecified robust variable-selection problem, a 10-data-set nearly parameter-matched affine comparison with AVTI shows stronger early model-mass recovery and competitive final joint accuracy under the same target-evaluation budget.
Siddhartha R Dalal, Vishal Misra, Abhay Parekhcs.LG stat.ML
Prior work has shown that transformers can perform exact Bayesian filtering within a fixed hypothesis class. Can they also perform Bayesian model selection -- identifying the correct hypothesis class from data? We introduce model-selection Bayesian wind tunnels: controlled environments where ground-truth posteriors over hypothesis classes are available in closed form. Using fixed-point-free involutions -- whose defining property f(f(x))=x is purely relational -- a 2.8M-parameter transformer achieves 0.01-bit entropy agreement with the Bayesian optimum (3 seeds), with both integer tokens and opaque symbols whose meanings change every episode. This extends to non-nested comparisons: involutions vs. 3-cycles (where neither class is a subset of the other) achieve class-posterior MAE under 0.001, demonstrating genuine model selection beyond simplicity/subset bias. We then identify a sharp perceptual access condition: when the discriminative statistic requires arithmetic -- modular addition (rotations) or multiplication (f(x)=cx mod p) -- model selection succeeds with integer tokens but fails completely with opaque symbols, and this boundary persists under 112x scaling (2.8M to 316M parameters). A stationarity control confirms the operative factor: opaque tokens with a fixed relabeling succeed (0.009-bit MAE), showing that stable semantics, not integer identity, enable circuit compilation. Header subtask diagnostics localize the failure to the composition of header inversion with arithmetic rather than header parsing itself. Probing frontier LLMs on the same tasks shows qualitative Bayesian behavior but a large calibration gap (~55x), measured through lossy probes and therefore directional rather than exact.