BEYOND ARGMAX: A MECHANISTIC STUDY OF SEMANTIC RETENTION IN FROZEN FOUNDATION-MODEL COMPOSITION FOR GENERALIZED FEW-SHOT 3D SEGMENTATION

Classical classifier-combination work distinguishes score-level fusion from hard decision voting. We revisit this distinction in a modern regime where independently pretrained, frozen foundation models are composed at inference time for generalized few-shot 3D segmentation. Our contribution is not a new sum rule; we ask a mechanistic question: *how much useful semantic information is lost when heterogeneous sources are collapsed to one class before interaction?* We construct a same-input top-$k$ retention intervention that freezes model weights, point clouds, masks, geometry, class vocabularies, scene lists, and fusion rules while varying only the semantic support retained before fusion. On 156 held-out ScanNet200 scenes, full retention reaches $34.87$ harmonic-mean (HM) IoU versus $28.47$ for the frozen top-1 control ($+6.40$, 95% paired scene-bootstrap CI $[+5.24,+7.64]$). The effect independently replicates on all 50 ScanNet++ validation scenes ($26.50$ vs. $23.02$ HM; $+3.48$, CI $[+1.64,+5.93]$). Most of the gain appears immediately after top-1, while moderate top-$k$ can slightly outperform the unrestricted tail. The conclusion survives a GroundingDINO–SAM2.1 sparse-source replacement, alternative fusion operators, tie-policy changes, and source-weight variation. Calibration diagnostics show severe but opposite raw miscalibration of the dense and sparse sources, yet calibration alone does not eliminate the retention advantage. We therefore identify *premature semantic collapse* as a repeatable information bottleneck in heterogeneous frozen-model composition.