Record: SP8192 + Polar Express NS + Multi-Phase Global TTT — val_bpb 1.0771 (3-seed mean)#1802
Open
aamodbhatt wants to merge 1 commit intoopenai:mainfrom
Open
Record: SP8192 + Polar Express NS + Multi-Phase Global TTT — val_bpb 1.0771 (3-seed mean)#1802aamodbhatt wants to merge 1 commit intoopenai:mainfrom
aamodbhatt wants to merge 1 commit intoopenai:mainfrom
Conversation
…1.0771 (3-seed mean)
8 tasks
sunnypatneedi
pushed a commit
to sunnypatneedi/parameter-golf
that referenced
this pull request
Apr 24, 2026
…ai#1787 Polar Express NS new base; PR openai#1795 PPM 1.01252; Issue openai#1604 deadline passed; Session 20 - Merged SOTA 1.0810 confirmed Day 15 (README not updated despite Scylla record commit) - Scylla 0.9485 committed to track_10min_16mb/ on Apr 23 (PR openai#1184) but byte accounting disputed by PR openai#1271 (corrected ~1.1289 bpb); treat merged SOTA as 1.0810 - PR openai#771 CLOSED/REJECTED confirmed; PR openai#727 CLOSED (illegal); PR openai#758 open but dead; PR openai#731 still awaiting seeds 1337+2024 - Issue openai#1604 (CaseOps ruling): NO @valerio-oai response in 11 days; self-deadline Apr 24 passed; proceed with clean legal stack immediately - NEW: PR openai#1787 (nprime06, 1.06335) — new community-consensus clean base with Polar Express Newton-Schulz (arXiv:2505.16932, ICLR 2026) + MIN_LR=0.10 warmdown floor - NEW: PR openai#1795 (OE-GOD, 1.01252) — byte-level PPM order-4 adaptive mixture; gate legality concern fixed; await organizer ruling before implementing - NEW: PR openai#1797 (dexhunter, 1.06157) — PR openai#1787 + SmearGate + LQER Asym; new dexhunter best - NEW: PR openai#1802 (aamodbhatt, 1.0771) — Polar Express NS + Multi-Phase Global TTT - TECHNIQUE: Polar Express NS (arXiv:2505.16932) and Gram NS (Dao-AILab) added to table - TECHNIQUE: MIN_LR=0.10 warmdown floor added to best-stack approach - Updated competition strategy: stop waiting for CaseOps, implement clean stack with Polar Express NS + MIN_LR immediately (6 days to deadline) https://claude.ai/code/session_01JZ3FiS937NwLHt3Fv9WHPD
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
val_bpb = 1.0771 (3-seed mean, std 0.0005) | ~15.99 MB | 8xH100 SXM
3-Seed Results
Merged SOTA (PR #1493): 1.0810 BPP. Delta: -0.0039 BPP.
Key Innovations
Multi-Phase Global TTT (Novel) — Instead of per-chunk score-then-train, score ALL windows globally, train ALL chunks, repeat for 3 phases. SGD (lr=0.015, momentum=0.9), cosine LR decay across chunks. Gives -0.0040 BPB improvement from TTT (vs -0.0017 for standard per-chunk TTT). Total eval time: ~440s.
Polar Express Newton-Schulz Coefficients — 5 per-iteration minimax-optimal NS tuples replacing fixed
(3.4445, -4.775, 2.0315). Higher-quality polar factor with same compute budget. (PR Record: SP4096 + Polar Express + MuonEq-R + Depth Recurrence — 1.0923 BPB (3-seed) #1344)MIN_LR=0.10 Warmdown Floor — Floors LR at 10% of peak during warmdown, enabling meaningful gradient updates throughout. +~70 extra training steps.
Compliance (Issue #1017, Track B)
Credits