ISO-AdamW tested: 1000-question benchmark shows only 4 answers over AdamW

On October 3, 2026, a DEV Community write-up described a hands-on test of ISO-AdamW, a new geometric optimizer the author had implemented, benchmarked against a baseline AdamW on the same set of problems. The comparison used a 1,000-question held-out test. ISO-AdamW answered 758 questions correctly; the baseline AdamW answered 754. The gap was four answers.

The author called out that gap directly and imagined the more exciting headline it could have supported — something along the lines of "New geometric optimizer beats AdamW! Time to rewrite the RL pipeline!" — but did not actually make that claim. No conclusion that the new optimizer beats AdamW was offered, and no comparison beyond the single held-out set was reported.

The write-up also documents a separate piece of engineering work: fixing a painfully slow matrix operation encountered during the implementation. That fix is part of the same record, alongside the benchmark numbers.

As of this report, the work stops at the published measurement. The only figures given are the 758 versus 754 result on 1,000 held-out questions and the resulting four-answer difference. The author did not state whether testing would be expanded, whether the RL pipeline would be changed, or whether ISO-AdamW would be evaluated on any other benchmark or task.