Chapter 4 Gate-Factorial Grounding Results

This page replaces the previous grounding recap with a direct comparison of the four mechanisms used in the pin/no-pin and keep-gate checks. Bars show matched grounded - direct deltas, averaged across available seeds; error bars are sample SD across seeds and black points are individual seeds.

Clean pinned/raw source root: gw_setup/outputs/ch4_gate_factorial_sixcell_clean_20260829
Reused dynamic comparator root: gw_setup/outputs/ch4_gate_factorial_activated_20260828

Incomplete cells:
Raw pinned + softmax keep: Symbolic -> RepTok / ImageNet 1000 x 10 has n=0

Figures

BGE -> RepTok on COCO-1000
BGE -> RepTok on COCO-1000
BGE -> RepTok on balanced ImageNet-1000
BGE -> RepTok on balanced ImageNet-1000
Symbolic -> RepTok external bbox evaluation
Symbolic -> RepTok external bbox evaluation
RepTok -> SONAR BLEU
RepTok -> SONAR BLEU
DINO -> symbolic classification
DINO -> symbolic classification
All metric deltas
All metric deltas

Downloads

per_seed_results.csv and aggregate_seed_results.csv contain the plotted values. Existing per-task bootstrap confidence intervals are carried through in the per-seed CSV when present; the bar error bars are seed variability.

Aggregate Table

DirectionDatasetMetricMechanismnDirectGrounded mean +/- SDDelta mean +/- SDDirectionality
BGE -> RepTokCOCO-1000CLIP similarityActivated dynamic + sigmoid keep30.286470.291675 +/- 0.000239+0.00520567 +/- 0.000239higher is better
BGE -> RepTokCOCO-1000CLIP similarityActivated dynamic + softmax keep30.286470.290629 +/- 0.000851+0.00415929 +/- 0.000851higher is better
BGE -> RepTokCOCO-1000CLIP similarityActivated pinned + sigmoid keep30.286470.293101 +/- 0.000112+0.00663126 +/- 0.000112higher is better
BGE -> RepTokCOCO-1000CLIP similarityActivated pinned + softmax keep30.286470.293127 +/- 0.000527+0.00665744 +/- 0.000527higher is better
BGE -> RepTokCOCO-1000CLIP similarityRaw pinned + sigmoid keep30.286470.292705 +/- 5.71e-05+0.00623528 +/- 5.71e-05higher is better
BGE -> RepTokCOCO-1000CLIP similarityRaw pinned + softmax keep30.286470.292329 +/- 0.000295+0.00585931 +/- 0.000295higher is better
BGE -> RepTokCOCO-1000FIDActivated dynamic + sigmoid keep375.614669.2706 +/- 0.324-6.34405 +/- 0.324lower is better
BGE -> RepTokCOCO-1000FIDActivated dynamic + softmax keep375.614669.6307 +/- 0.802-5.98393 +/- 0.802lower is better
BGE -> RepTokCOCO-1000FIDActivated pinned + sigmoid keep375.614668.697 +/- 0.121-6.91765 +/- 0.121lower is better
BGE -> RepTokCOCO-1000FIDActivated pinned + softmax keep375.614668.0537 +/- 0.259-7.5609 +/- 0.259lower is better
BGE -> RepTokCOCO-1000FIDRaw pinned + sigmoid keep375.614668.7221 +/- 0.121-6.89254 +/- 0.121lower is better
BGE -> RepTokCOCO-1000FIDRaw pinned + softmax keep375.614669.0655 +/- 0.0842-6.54913 +/- 0.0842lower is better
BGE -> RepTokCOCO-1000LPIPSActivated dynamic + sigmoid keep30.6971830.69602 +/- 0.000304-0.00116261 +/- 0.000304lower is better
BGE -> RepTokCOCO-1000LPIPSActivated dynamic + softmax keep30.6971830.695985 +/- 0.000241-0.00119813 +/- 0.000241lower is better
BGE -> RepTokCOCO-1000LPIPSActivated pinned + sigmoid keep30.6971830.697211 +/- 6.63e-05+2.78354e-05 +/- 6.63e-05lower is better
BGE -> RepTokCOCO-1000LPIPSActivated pinned + softmax keep30.6971830.69724 +/- 0.000408+5.72602e-05 +/- 0.000408lower is better
BGE -> RepTokCOCO-1000LPIPSRaw pinned + sigmoid keep30.6971830.696936 +/- 9.32e-05-0.000246982 +/- 9.32e-05lower is better
BGE -> RepTokCOCO-1000LPIPSRaw pinned + softmax keep30.6971830.69697 +/- 0.000115-0.000213166 +/- 0.000115lower is better
BGE -> RepTokImageNet-1000CLIP similarityActivated dynamic + sigmoid keep30.3242030.326046 +/- 0.00014+0.00184281 +/- 0.00014higher is better
BGE -> RepTokImageNet-1000CLIP similarityActivated dynamic + softmax keep30.3242030.325806 +/- 0.000292+0.00160271 +/- 0.000292higher is better
BGE -> RepTokImageNet-1000CLIP similarityActivated pinned + sigmoid keep30.3242030.326694 +/- 0.000168+0.00249083 +/- 0.000168higher is better
BGE -> RepTokImageNet-1000CLIP similarityActivated pinned + softmax keep30.3242030.326585 +/- 0.000177+0.00238241 +/- 0.000177higher is better
BGE -> RepTokImageNet-1000CLIP similarityRaw pinned + sigmoid keep30.3242030.326339 +/- 0.000298+0.00213664 +/- 0.000298higher is better
BGE -> RepTokImageNet-1000CLIP similarityRaw pinned + softmax keep30.3242030.326415 +/- 9.6e-05+0.00221173 +/- 9.6e-05higher is better
BGE -> RepTokImageNet-1000FIDActivated dynamic + sigmoid keep351.6450.6868 +/- 0.125-0.953152 +/- 0.125lower is better
BGE -> RepTokImageNet-1000FIDActivated dynamic + softmax keep351.6450.7774 +/- 0.16-0.862517 +/- 0.16lower is better
BGE -> RepTokImageNet-1000FIDActivated pinned + sigmoid keep351.6450.4729 +/- 0.00393-1.16708 +/- 0.00393lower is better
BGE -> RepTokImageNet-1000FIDActivated pinned + softmax keep351.6450.4728 +/- 0.17-1.16719 +/- 0.17lower is better
BGE -> RepTokImageNet-1000FIDRaw pinned + sigmoid keep351.6450.5645 +/- 0.156-1.07546 +/- 0.156lower is better
BGE -> RepTokImageNet-1000FIDRaw pinned + softmax keep351.6450.6411 +/- 0.13-0.998834 +/- 0.13lower is better
BGE -> RepTokImageNet-1000LPIPSActivated dynamic + sigmoid keep30.6475360.646692 +/- 0.000103-0.000843028 +/- 0.000103lower is better
BGE -> RepTokImageNet-1000LPIPSActivated dynamic + softmax keep30.6475360.647548 +/- 0.000159+1.2219e-05 +/- 0.000159lower is better
BGE -> RepTokImageNet-1000LPIPSActivated pinned + sigmoid keep30.6475360.646077 +/- 0.000292-0.00145831 +/- 0.000292lower is better
BGE -> RepTokImageNet-1000LPIPSActivated pinned + softmax keep30.6475360.64568 +/- 0.00013-0.00185593 +/- 0.00013lower is better
BGE -> RepTokImageNet-1000LPIPSRaw pinned + sigmoid keep30.6475360.645413 +/- 0.000147-0.00212284 +/- 0.000147lower is better
BGE -> RepTokImageNet-1000LPIPSRaw pinned + softmax keep30.6475360.645638 +/- 0.000314-0.00189779 +/- 0.000314lower is better
DINO -> symbolicImageNet-50kTop-1 accuracyActivated dynamic + sigmoid keep30.8360.836027 +/- 6.43e-05+2.66433e-05 +/- 6.43e-05higher is better
DINO -> symbolicImageNet-50kTop-1 accuracyActivated dynamic + softmax keep30.8360.799627 +/- 0.0413-0.0363734 +/- 0.0413higher is better
DINO -> symbolicImageNet-50kTop-1 accuracyActivated pinned + sigmoid keep30.8360.83598 +/- 0.00014-2.00272e-05 +/- 0.00014higher is better
DINO -> symbolicImageNet-50kTop-1 accuracyActivated pinned + softmax keep30.8360.81284 +/- 0.0203-0.02316 +/- 0.0203higher is better
DINO -> symbolicImageNet-50kTop-1 accuracyRaw pinned + sigmoid keep30.8360.830207 +/- 0.00108-0.00579335 +/- 0.00108higher is better
DINO -> symbolicImageNet-50kTop-1 accuracyRaw pinned + softmax keep30.8360.79168 +/- 2e-05-0.04432 +/- 2e-05higher is better
RepTok -> SONARImageNet-50kCorpus BLEU-4Activated dynamic + sigmoid keep311.811712.3648 +/- 0.423+0.553028 +/- 0.423higher is better
RepTok -> SONARImageNet-50kCorpus BLEU-4Activated dynamic + softmax keep311.811711.5165 +/- 1.31-0.295276 +/- 1.31higher is better
RepTok -> SONARImageNet-50kCorpus BLEU-4Activated pinned + sigmoid keep311.811711.2083 +/- 0.941-0.603406 +/- 0.941higher is better
RepTok -> SONARImageNet-50kCorpus BLEU-4Activated pinned + softmax keep311.811711.3173 +/- 1.57-0.494454 +/- 1.57higher is better
RepTok -> SONARImageNet-50kCorpus BLEU-4Raw pinned + sigmoid keep311.811712.3779 +/- 0.453+0.566166 +/- 0.453higher is better
RepTok -> SONARImageNet-50kCorpus BLEU-4Raw pinned + softmax keep311.811711.7102 +/- 1.19-0.101519 +/- 1.19higher is better
Symbolic -> RepTokImageNet 1000 x 10ResNet target-class accuracyActivated dynamic + sigmoid keep30.95250.957567 +/- 0.00177+0.00506669 +/- 0.00177higher is better
Symbolic -> RepTokImageNet 1000 x 10ResNet target-class accuracyActivated dynamic + softmax keep30.94990.9546 +/- 0.00329+0.00469995 +/- 0.00329higher is better
Symbolic -> RepTokImageNet 1000 x 10ResNet target-class accuracyActivated pinned + sigmoid keep30.95050.951633 +/- 0.00142+0.00113332 +/- 0.00142higher is better
Symbolic -> RepTokImageNet 1000 x 10ResNet target-class accuracyActivated pinned + softmax keep30.94680.956033 +/- 0.00311+0.00923334 +/- 0.00311higher is better
Symbolic -> RepTokImageNet 1000 x 10ResNet target-class accuracyRaw pinned + sigmoid keep30.95090.402733 +/- 0.115-0.548167 +/- 0.115higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO IoUActivated dynamic + sigmoid keep30.4600040.458865 +/- 0.00115-0.00113982 +/- 0.00115higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO IoUActivated dynamic + softmax keep30.4597040.448074 +/- 0.00171-0.0116296 +/- 0.00171higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO IoUActivated pinned + sigmoid keep30.460510.476182 +/- 0.00497+0.0156717 +/- 0.00497higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO IoUActivated pinned + softmax keep30.4580740.476833 +/- 0.00159+0.018759 +/- 0.00159higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO IoURaw pinned + sigmoid keep30.4594110.463827 +/- 0.0234+0.00441679 +/- 0.0234higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO detection rateActivated dynamic + sigmoid keep30.98470.9822 +/- 0.000889-0.00249996 +/- 0.000889higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO detection rateActivated dynamic + softmax keep30.9860.978367 +/- 0.000379-0.00763335 +/- 0.000379higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO detection rateActivated pinned + sigmoid keep30.98230.978333 +/- 0.000586-0.00396659 +/- 0.000586higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO detection rateActivated pinned + softmax keep30.98330.972867 +/- 0.00159-0.0104334 +/- 0.00159higher is better
Symbolic -> RepTokImageNet 1000 x 10YOLO detection rateRaw pinned + sigmoid keep30.98220.991833 +/- 0.00196+0.00963328 +/- 0.00196higher is better