HELIX-56M-VS-TRANSFORMER

A same-size, same-data, same-protocol dense transformer will reveal how much perplexity HELIX pays for its O(L)-inference design.

A 57.36M-parameter Llama-style dense transformer trained on Google Colab (A100) under the identical WikiText-103 data protocol, step count, and schedule as HELIX-56M's wt103-complete run.

• ◆Architecture only (HELIX H-E-R vs. dense transformer)

• →HELIX-56M: val 26.91 / test 27.29 tok PPL (44.99 / 48.59 word).
• →Transformer: val 24.70 / test 25.10 tok PPL (41.48 / 44.82 word).
• →Gap: 0.084 nats test — roughly 8–9% perplexity, stable from step 7k onward.

• ·Run deliberately before scaling, not after — the lab wanted the tax measured while a bad number was still cheap to learn from.

HELIX pays a real, measured perplexity cost for its recurrent/O(L) design at 56M. The gap becomes the reference point every later generation is checked against.

Re-check the gap at 127M with a second matched control.