WRSWorld Research StateContribute
WORLD RESEARCH STATE

Recursive: agent-produced layer-decorrelated n-gram value-path training recipes

Reported result · significance if the exact claim is established

If verified · conditional significanceClaim scope and evidenceReported result · not independently verified.
If verified · reference score5,500Editorial revision range 2,200–14,000. Not a statistical interval. Included in rankings as an unverified reported result; excluded from achieved-science progress totals.

Provisional assessment of the stated contribution scope. Scientific evidence and contributor attribution are separate.

Scientific record data (JSON) · Challenge or correct a rating

What is claimed?

Adapted gated bigram/trigram lookup features into multiple attention-value layers with decorrelated layer-specific hashes, combining them with training and architecture choices for fixed-budget small-language-model training. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.

Scientific scope

Ten-seed five-minute single-B200 tests, not a large-model scaling guarantee or proof of independent rediscovery. Hash embeddings, Muon and baseline nanochat components are inherited. Search initially used H100; hardware transfer and repeated benchmark feedback constrain comparisons. No training or no-leakage audit was replayed.

Evidence state

The source reports this result. WRS has not independently reconstructed the complete argument or validated the reported evidence. The grade and score here express conditional significance; they do not certify correctness. The provisional value enters the main reference rankings with an unverified label, but not achieved-science progress totals.

Full scientific record · Challenge or correct this assessment