BLEU Grounding
BLEU: 11.90 -> 12.63
Canonical Activated-Start Grounding Results
Direct GW decodes are compared with the corresponding activated-start iterative grounder. The raw-start bbox result is excluded from the canonical claim.
DirectGrounded
TTI Grounding
COCO: LPIPS: 0.6972 -> 0.6953; CLIPScore: 0.2864 -> 0.2916; FID: 75.6 -> 68.9
ImageNet heldout: LPIPS: 0.6475 -> 0.6469; CLIPScore: 0.3242 -> 0.3262; FID: 51.6 -> 51.0
BBox IoU Grounding
YOLO11n IoU: 0.4494 -> 0.4433
Symbolic Grounding
accuracy_pct: 83.5940 -> 83.5920
Training Notes
- BLEU grounding. RepTok-to-SONAR uses an activated-start iterative attention grounder over the five GW modality hypotheses. The selected run used a fixed BLEU probe with a joint probe-reward plus MSE objective.
- TTI grounding. BGE-to-RepTok uses the activated-start short_T5_keep iterative attention recipe trained on the ImageNet/GW training split with a RepTok MSE objective.
- BBox IoU grounding. Symbolic-to-RepTok uses the activated-start bbox proxy grounder trained against a frozen RepTok-bbox head. The raw-start improvement is intentionally excluded from this canonical table.
- Symbolic grounding. DINO-to-symbolic uses an activated-start CE grounder with sigmoid keep gating, optimized on symbolic classification accuracy/cross-entropy.
Metric Table
| Task | Eval | Metric | Direct | Grounded | Delta | Direction |
|---|---|---|---|---|---|---|
| BLEU | ImageNet heldout | BLEU | 11.9035 | 12.6323 | +0.728723 | higher is better |
| TTI | COCO val2017 | LPIPS | 0.697177 | 0.695347 | -0.00182959 | lower is better |
| TTI | COCO val2017 | CLIPScore | 0.286432 | 0.291646 | +0.00521456 | higher is better |
| TTI | COCO val2017 | FID | 75.6099 | 68.9286 | -6.6813 | lower is better |
| TTI | ImageNet heldout | LPIPS | 0.64754 | 0.646865 | -0.000674901 | lower is better |
| TTI | ImageNet heldout | CLIPScore | 0.324205 | 0.326213 | +0.00200819 | higher is better |
| TTI | ImageNet heldout | FID | 51.6413 | 51.0234 | -0.617939 | lower is better |
| BBox IoU | ImageNet heldout bbox subset | YOLO11n IoU | 0.449428 | 0.443313 | -0.0061145 | higher is better |
| Symbolic | ImageNet heldout | accuracy_pct | 83.594 | 83.592 | -0.002 | higher is better |