Reflqt.
HOMERESEARCHPUBLICATIONSBLOGEXPLOREABOUTCONTACT →
§ art-021-eml-paradox · April 24, 2026 · PIVOT

The EML Paradox: Better Language Model, Worse Symbolic Regression

Refleqt Labs · 7 min read

An EML hybrid FFN beat BD4 on 125M-scale language modelling and failed to recover sin(x) in all 12 symbolic regression configurations we tested -- a clean demonstration that universality and trainability are different things.

inside the ffnemlsymbolic regressionnegative resultsstructured matrices
RELATED
APR 20The Case for Negative Results in ML Research
5 min→
APR 24Block-Diagonal Transformers Win 8 of 17 Downstream Tasks
7 min→
APR 24Structured Transformers in 2026: Where FFN Research Stands
8 min→
Reflqt.

Every claim, plotted. A research lab in public.

Research
All entriesPublicationsBlog
Explore
All demosAlgebraKnowledge Layer
Lab
About · Open letterContactGitHub
© 2026 REFLQT. LABSv2026.04 · plotted