Add repository index and remove generated bytecode
Browse filesLink the repository card to the detailed adaptive-skip analysis and remove an accidental local pycache artifact.
README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
tags:
|
| 4 |
+
- image-generation
|
| 5 |
+
- diffusion
|
| 6 |
+
- adaptive-computation
|
| 7 |
+
- layer-skipping
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# SRA-B Experiment Results
|
| 11 |
+
|
| 12 |
+
This repository contains evaluation artifacts and analysis for SRA-B/LARA
|
| 13 |
+
experiments.
|
| 14 |
+
|
| 15 |
+
## Official LARA 400k adaptive-skip landscape
|
| 16 |
+
|
| 17 |
+
The complete 35,840-observation transition study, plots, configuration, logs,
|
| 18 |
+
reproduction script, and detailed interpretation are available here:
|
| 19 |
+
|
| 20 |
+
**[Open the full adaptive-skip study](experiments/adaptive_skip_landscape/official_lara_400k/README.md)**
|
| 21 |
+
|
| 22 |
+
Main conclusion: the checkpoint has substantial oracle adaptive-depth potential
|
| 23 |
+
(38.2% theoretical backbone reduction at suffix-error threshold 0.10), but
|
| 24 |
+
direct-vs-composed disagreement is a weak standalone verifier (Spearman 0.147;
|
| 25 |
+
AUROC 0.577).
|
| 26 |
+
|