Link the paper page; SigLIP2 SO400M head is now the best epoch (13 peak heads)

README.md

@@ -10,16 +10,20 @@ library_name: pytorch
 
 # Efficient Probing heads
 
+**Paper:** [Attention, Please! Revisiting Attentive Probing Through the Lens of Efficiency](https://huggingface.co/papers/2506.10178)
+(ICLR 2026) · [arXiv:2506.10178](https://arxiv.org/abs/2506.10178) ·
+[code & leaderboard](https://github.com/billpsomas/efficient-probing)
+
 Trained EP (efficient probing) heads for all 37 frozen encoders of the standing
-ImageNet-1k benchmark at [billpsomas/efficient-probing](https://github.com/billpsomas/efficient-probing)
-(ICLR 2026). Each head is a multi-query cross-attention pooling + BatchNorm + linear
-classifier, trained for 90 epochs with LARS on frozen features. **No backbone weights
-are included** -- each `config.json` records the exact `--model` / loader flags to
+ImageNet-1k benchmark at [billpsomas/efficient-probing](https://github.com/billpsomas/efficient-probing).
+Each head is a multi-query cross-attention pooling + BatchNorm + linear classifier,
+trained for 90 epochs with LARS on frozen features. **No backbone weights are
+included** -- each `config.json` records the exact `--model` / loader flags to
 rebuild the encoder from its original source.
 
 Why two kinds of head: the training code initially kept only a rolling per-epoch
 checkpoint, so a finished run left its **final** epoch behind rather than its best.
-12 early-peaking models were later re-run with best-epoch checkpointing, so their
+13 early-peaking models were later re-run with best-epoch checkpointing, so their
 heads are the **peak** and reproduce the leaderboard number exactly. Every file's
 metadata records both its own accuracy at the saved epoch and the table's
 best-epoch figure, so nothing has to be taken on trust.
@@ -55,7 +59,7 @@ python tools/eval_reimagenet.py predict <backbone flags from config.json> \
 | MetaCLIP2 ViT-bigG/14-378 | ep | 88.12 | 6 | peak |
 | EVA02-CLIP E-14-plus | ep | 87.98 | 6 | peak |
 | EVA02-CLIP E-14 | ep | 87.70 | 6 | peak |
-| SigLIP2 SO400M/14 | ep | 85.64 | 29 | final epoch |
+| SigLIP2 SO400M/14 | ep | 87.68 | 6 | peak |
 | PE-Core L-14/336 | ep | 87.25 | 12 | peak |
 | MetaCLIP2 ViT-bigG/14 | ep | 87.11 | 6 | peak |
 | SigLIP2 ViT-L/16 | ep | 87.06 | 6 | peak |
@@ -92,3 +96,15 @@ python tools/eval_reimagenet.py predict <backbone flags from config.json> \
 Full provenance (training logs, exact commands, the leaderboard itself) lives in the
 [GitHub repo](https://github.com/billpsomas/efficient-probing). Heads were trained on
 ImageNet-1k; use accordingly.
+
+## Citation
+
+```bibtex
+@inproceedings{psomas2026attention,
+  title     = {Attention, Please! Revisiting Attentive Probing Through the Lens of Efficiency},
+  author    = {Bill Psomas and Dionysis Christopoulos and Eirini Baltzi and Ioannis Kakogeorgiou and Tilemachos Aravanis and Nikos Komodakis and Konstantinos Karantzalos and Yannis Avrithis and Giorgos Tolias},
+  booktitle = {The Fourteenth International Conference on Learning Representations},
+  year      = {2026},
+  url       = {https://openreview.net/forum?id=PXo0gtT7Al}
+}
+```

manifest.json

@@ -387,10 +387,10 @@
    "cls_features": "ep",
    "ep_queries": 32,
    "d_out": 1,
-   "head_epoch": 29,
-   "val_acc1_at_head_epoch": 85.64,
+   "head_epoch": 6,
+   "val_acc1_at_head_epoch": 87.68,
    "table_best_acc1": 87.68,
-   "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+   "note": "best-epoch head: saved at the run's peak, the number the table reports.",
    "training_log": "logs/siglip2_so400m/ep.txt",
    "params_incl_bn_stats": 2519273,
    "file": "siglip2_so400m/ep_head.pth",

siglip2_so400m/config.json

@@ -11,9 +11,9 @@
  "cls_features": "ep",
  "ep_queries": 32,
  "d_out": 1,
- "head_epoch": 29,
- "val_acc1_at_head_epoch": 85.64,
+ "head_epoch": 6,
+ "val_acc1_at_head_epoch": 87.68,
  "table_best_acc1": 87.68,
- "note": "final-epoch head, not best-epoch: checkpoints overwrite per epoch. table_best_acc1 is the peak of the same run; the difference is the post-peak decline of a frozen-feature probe, largest on VLM encoders.",
+ "note": "best-epoch head: saved at the run's peak, the number the table reports.",
  "training_log": "logs/siglip2_so400m/ep.txt"
 }

siglip2_so400m/ep_head.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6215dc7c33f7a960c045c8d3989a1a5590ab816adc511c722a0ad507b5275a68
-size 10080785
+oid sha256:af8d28e35f4a9f481023540f60e59e63edd64a9fa534c1b5932dbde36d9b9f70
+size 10080657