# Leader-Directed Differential Evolution for Adapter Training

- ID: 1582
- Canonical URL: https://synthcore.org/idea/1582/leader-directed-differential-evolution-for-adapter-training
- API JSON: https://synthcore.org/api/idea/1582.json
- API Markdown: https://synthcore.org/api/idea/1582.md
- Verification status: unverified
- Source: [arXiv:2607.29228](https://arxiv.org/abs/2607.29228)
- Category: optimization
- Solves: sample-efficiency, accuracy, stability
- ML areas: optimizer, fine-tuning, training-dynamics
- Math tags: probability, linear-algebra, optimization
- Ratings: usefulness 5/10; difficulty 5/10; novelty 5/10

## Idea description

Use differential evolution over adapter or prompt parameters, combining attraction to the current best parameter vector with a population-difference direction. Binomial crossover supplies coordinate-level exploration, while the operator-selection separation makes it possible to measure raw proposal geometry independently from parameter repair and noisy fitness selection.

## Mathematical statement

For target vector x_i in R^d, best vector x_best, two distinct population vectors x_r1 and x_r2, scale factor F, and diagonal crossover mask C = diag(C_1,...,C_d), the paper gives the DE/current-to-best/1/bin proposal u_i = x_i + C[F(x_best-x_i) + F(x_r1-x_r2)]. Each C_j is Bernoulli with crossover probability CR, except one forced coordinate is set to one so the offspring cannot equal x_i solely because of crossover. The first difference is leader attraction; the second is a population-difference direction. The paper's variation-repair-selection factorization requires analyzing raw u_i before feasibility repair because repair can change mean, covariance, support, and coordinate dependence. In neural adapter training, x_i is the flattened LoRA, prefix, or prompt vector and f is validation loss or a noisy minibatch loss. Optional per-layer normalization should be applied to population differences so large adapter tensors do not dominate the proposal.

## Key formulas

- $$\bm{u}\cdot i=\bm{x}\cdot i+\bm{C}\left[F(\bm{x}\cdot{\mathrm{best}}-\bm{x}\cdot i)+F(\bm{x}\cdot{r_{1}}-\bm{x}\cdot{r_{2}})\right]$$
- $$u_i=x_i+C\left[F(x_{\mathrm{best}}-x_i)+F(x_{r_1}-x_{r_2})\right],\qquad C=\operatorname{diag}(C_1,\ldots,C_d)$$
- $$C_j\sim\operatorname{Bernoulli}(CR),\qquad C_{j_0}=1\text{ for one uniformly selected forced coordinate }j_0$$
- $$\text{state}\xrightarrow{\mathcal{V}\,\mathrm{(DE)}}\text{raw }u_i\xrightarrow{\mathcal{R}_{\Omega}}\text{feasible adapter}\xrightarrow{\mathcal{S}_{f}}\text{next population}$$

## Implementation notes

Integrate this at the optimizer level for a frozen pretrained network with trainable LoRA, prefix-tuning, or soft-prompt parameters. Flatten each candidate's trainable parameters into x_i, while retaining a layer map so the differential vector can optionally be normalized separately per tensor. For layer l, compute d_l = (x_r1,l - x_r2,l)/(RMS(x_r1,l-x_r2,l)+epsilon), and use the corresponding normalized leader difference if per-layer scale control is enabled. At every generation, identify x_best using a fixed validation minibatch or an averaged score over two minibatches. For each target, sample distinct r1 and r2, sample C_j from Bernoulli(CR), force one random coordinate or tensor block to be active, calculate u_i = x_i + C elementwise-multiplied by [F(x_best-x_i) + F(x_r1-x_r2)], project into LoRA norm bounds, evaluate the candidate, and retain elites plus the best offspring. Use population size 16 to 32, F in [0.3,0.8], and CR in [0.5,0.9] for the first sweep. Record raw proposal norms, changed-coordinate fraction, coordinate coverage, repaired-coordinate fraction, and loss improvement per forward evaluation. The first experiment should use a small ViT or 100M-parameter transformer with rank-4 or rank-8 adapters on CIFAR-100 or SST-2, comparing against random search, isotropic evolutionary strategies, and gradient-based LoRA under an equal forward-pass budget. The predicted signal is faster validation improvement than isotropic perturbations when population differences capture useful adapter directions; a negative result is equal or worse loss descent after accounting for population evaluations.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.