Can NDA estimate coefficients in theories where multiple mass scales compete?

Yes, though the question's framing is a bit abstract, the preprints show that modern machine-learning-assisted methods can now handle theories with competing mass scales.

The most direct answer comes from effective field theory (EFT) inference. [AG-2026.03-1506] describes a differentiable framework that combines renormalization-group evolution (which tracks how couplings change across energy scales) with matching conditions at different mass thresholds—precisely the scenario where multiple scales compete. By making the entire multi-scale calculation differentiable, gradient-based inference can now estimate hundreds of EFT coefficients simultaneously. The breakthrough is practical: they demonstrate this on 374-parameter SMEFT (Standard Model Effective Field Theory) analyses, which would have been intractable a few years ago.

Why this matters: previously, when you had heavy new physics at one scale and Standard Model physics at another, you had to guess where to "match" them; now the fit itself learns the optimal matching scale and coefficient values together.

Beyond EFT, [AG-2024.02-531] shows neural networks can compute Yukawa couplings in string theory by numerically solving the underlying geometry at a given modulus value, circumventing the need for closed-form formulas when multiple mass hierarchies emerge from a single compactification. And [AG-2024.11-1468] proposes a more general uncertainty framework (theory nuisance parameters) that can be profiled in fits, allowing data to constrain theories even when their higher-order structure involves multiple energy scales.

The preprints do not directly discuss NDA (numerical derivative approximation or similar acronym) as a specific algorithm; if you meant a different method, that context would help.