MARMAM
Knowing in Advance When an Evolutionary Outer Loop Will Not Help
arXiv 2606.29119
Read on arXiv →What It Is
A pre-registered, cheap-baseline screening rule that decides — before any expensive evolutionary or population-based outer loop is built — whether that outer loop can possibly help. A single Phase-0 gate computes a recovery ratio R; when R ≥ 90%, the outer loop is predicted to add nothing and is skipped.
The Problem
Evolutionary and population-based outer loops over neural networks are costly to build and run, and they routinely appear to "win" only because a far simpler static baseline was never measured. MARMAM began as a heritable sparse-mutability-mask method for rescuing gradient-warmed LoRA adapters — but a one-week random-κ control showed κ-sparsity alone captured ~99% of the rescue effect. The generalizable contribution is the screening rule that catches this in advance.
Approach
- Defined the recovery ratio R = (best single-shot gradient / curvature-statistic gain) ÷ (best gain from cheap static methods), computed at a Phase-0 gate before the outer loop is implemented
- Pre-registered the decision threshold: R ≥ 90% ⇒ the evolutionary outer loop is predicted not to help and is skipped
- Validated the rule against pre-registered experiments, checking each screening verdict against the eventual full outer-loop result
- Anchored to two real lab post-mortems (UYIR, MARMAM) where the cheap baseline matched the mechanism the project was built around
Results
- The screening gate correctly flagged cases where cheap static computations matched the full evolutionary approach
- Gate cost ≈ 50–70 GPU-hours while preventing an estimated 400+ GPU-hours of outer-loop compute plus weeks of development
- Turns the lab's standing "falsify the cheap baseline first" discipline into a quantitative, pre-registered go/no-go rule