[Non-record] Meta-Learned TTT + Error-Guided Adaptation Analysis (val_bpb=1.1645)#294

Closed

sseanliu wants to merge 12 commits intoopenai:mainfrom

Commits on Mar 19, 2026

Seed Crystallization: ALBERT-style weight sharing with 3 wide blocks x 4 recycled = 12 effective layers
sseanliu
committed
Refactor shared block forward pass for torch.compile compatibility
sseanliu
committed
Add README for Seed Crystallization submission
sseanliu
committed

Commits on Mar 20, 2026

Add autoresearch framework for autonomous experiment loop
sseanliu
committed
Fix data paths to resolve relative to repo root
sseanliu
committed
Improve autoresearch program.md with depth frontier strategy and lessons learned
sseanliu
committed
Add early-vs-late reversal warning to program.md
sseanliu
committed
Add comprehensive literature review for meta-learned TTT strategy
sseanliu
committed
Add SmearGate, BigramHash, SWA, Reptile meta-learning, and TTT eval to MetaTTT submission
sseanliu
committed
Add MetaTTT v2 research submission: Reptile meta-learning + test-time training
sseanliu
committed
MetaTTT v2: int5 quantization fix, TTT eval stride fix, 1xH100 validation results
sseanliu
committed

Commits on Mar 21, 2026

Add MetaTTT v2 research submission: Reptile meta-learning + error-guided TTT analysis
sseanliu
committed