diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-04-08 12:22:58 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-04-08 12:22:58 -0500 |
| commit | c201cb31018b35bf88482f7dc768b8f7a057703b (patch) | |
| tree | d12e4640b4d0abef34c73f0f667f8a0eb026f794 /experiments/snapshot_exploitability.py | |
| parent | 35be969067396306c19a3caac2d887bcde48c5d0 (diff) | |
Round 41 (Appendix L): add per-block drift diagnostic reinforcing cos-vs-acc hypothesis
Extracted from existing round 38 JSON data without running new compute.
The drift field (||W_final - W_init||_F / ||W_init||_F) is produced by
cifar_resmlp.py's feature_drift() and was already saved but not used in
the paper.
Key finding: CB+penalty has LARGER block updates than SB+penalty
(per-block w2 drift 19.3x vs 14.3x; embed drift 44.6x vs 7.1x) yet
9.3 pp LOWER accuracy. This rules out 'CB just has smaller updates'
as an alternative explanation for the cos-vs-acc dissociation.
Added 2 sentences to Appendix L paragraph 2 noting this supporting
evidence for the 'angular agreement does not certify functional
forward-state content' mechanism hypothesis in ยง4.
Main content still 9 pages exactly within E&D budget.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Diffstat (limited to 'experiments/snapshot_exploitability.py')
0 files changed, 0 insertions, 0 deletions