Recursive Transformer 4B/7L + VE + QAT + TTT — val_bpb 1.1696 (3-seed mean)

[Tonyy1977]

Summary

Recursive transformer: 4 shared blocks × 7 loops (7× weight reuse) at dim=1024, with ValueEmbedding, int6 QAT from step 0, and score-first TTT+sliding window eval.

3-seed mean: 1.1696 BPB | ~15.85MB artifact | 600s on 8xH100 SXM

Seed	TTT+Sliding BPB	Post-quant BPB	Artifact
1337	1.1698	1.1952	15,749,104
42	1.1697	1.1949	15,778,257
2024	1.1693	1.1947	15,750,116
Mean	1.1696	1.1949

Key novelty

Unlike other depth recurrence submissions that repeat 1-2 layers on top of 10-11 unique blocks (~1.2× reuse), this uses 4 shared blocks looped 7 times (7× reuse). This enables dim=1024 (2× wider than standard 512) while staying under 16MB.

Architecture highlights

• U-Net encoder-decoder skip connections across loops
• Int6 QAT from step 0 (essential for recursive models — without it, quantization error compounds through loops)
• ValueEmbedding reinjects token identity at late loops
• SmearGate + BigramHash + XSA on last 4 loops
• Score-first TTT + sliding window eval (stride=64)

See README.md in the submission folder for full details and negative results.

[MatoTeziTanka]

Community Review — Recursive Transformer 4B/7L + VE + QAT + TTT — val_bpb 1.1696 (3-seed mean)

BPB: 1.1696 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 64490b14a165, file records/track_10min_16mb/2026-03-26_RecursiveTransformer_4B7L_VE_QAT_TTT/train_gpt.py):

The TTT path at line 918 implements the score-first-per-chunk pattern: each chunk is scored under torch.no_grad() / inference_mode() before the base_model.train() + SGD adaptation runs on that same chunk, with an is_last_chunk guard so the final chunk gets no adaptation pass. This is the structural shape the legal frontier uses (PRs #1416 erichroepke, #1423 aryanbhosale).

Per Issue #402 and Issue #677, TTT is legal when each token is scored before the adapter updates on it, and that's what the code does here — chunk ci is scored under weights adapted only on chunks 0..ci-1. No prequant_ttt_adapt_adamw(val_tokens, ...) multi-epoch fine-tune, no scored-region SLOT, no target-in-key n-gram cache.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.07s, dim=1024, layers=, vocab=1024, code=68750 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending standard checks (3-seed validation, 16MB artifact cap, 10-min wallclock on 8×H100 SXM). The compliance picture matches the legal reference frontier and no flags were raised by the classification pass.

Auto-classification caveat: this review was drafted by the AST-based classifier against a template derived from manually-reviewed cluster PRs (#1420, #1450, #1487, #1541, #1529, #1533, #1518). If I've misread a subtlety in your eval path — e.g., multi-epoch TTT that I mistook for single-pass, or a target-in-key lookup I missed in a helper function — please flag it and I'll re-run the audit manually.

---

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.07s, dim=1024, layers=, vocab=1024, code=68750 B, SMOKE_TEST_PASS. Classification via deterministic AST-based classify_prs.py (pattern bank derived from ~65 manually-reviewed PRs earlier in the 2026-04-11 sweep). This review was auto-drafted from a template and spot-checked before posting — if the template misread your code, please call it out so I can iterate the classifier.