Evolution-Proxy Approach: A Methodological Strategy

• Evolution-Proxy Approach is a strategy that replaces direct evaluations with proxy objects that preserve essential ordering and diagnostic sensitivity.
• It underpins diverse applications, from machine learning training and macro placement to medical imaging and physical sciences diagnostics.
• Its effectiveness depends on the proxies’ fidelity and calibrated mapping to true outcomes, while misalignment can lead to suboptimal or incorrect results.

Searching arXiv for the cited papers and topic terminology to ground the article in current literature. The expression Evolution-Proxy Approach does not denote a single canonical algorithm across the literature represented here. Rather, it denotes a recurring methodological pattern in which an expensive, inaccessible, or unstable target evaluation is replaced by a proxy object that preserves enough structure to guide iterative refinement, selection, or diagnosis. In machine learning engineering, the proxy is a low-fidelity training-and-scoring loop; in macro placement, a deterministic greedy probe; in medical image segmentation, proxy data and proxy networks; in multi-agent reinforcement learning, an evolutionary reward aligned with replicator abundance; in condensed-matter and helioseismic inference, transfer-matrix and entropy-like observables; and in AMO physics, the notion of a proxy field appears as a cautionary example of a proxy that can be useful in limited regimes but is not physically equivalent to the target system (Yuan et al., 6 Nov 2025, Mo et al., 8 Jun 2026, Nath et al., 2021, Abrantes et al., 2020, Pi et al., 2021, Buldgen et al., 11 Jun 2025, Reiss, 2016).

1. Conceptual structure and scope

Taken together, these works suggest a common architecture. A candidate, policy, model, or physical state is not judged directly by the full objective; instead, a cheaper or more accessible surrogate is constructed, and the search or inference loop evolves with respect to that surrogate. The essential requirement is not exact equality between proxy and target, but preservation of a useful ordering, phase structure, parity, or diagnostic sensitivity. When that preservation holds, the proxy can accelerate exploration, broaden the searchable design space, and make otherwise intractable evaluations operational. When it fails, the search can become misaligned or even physically incorrect.

A concise taxonomy is useful.

Context	Proxy object	Role in the loop
Automated ML engineering	Low-fidelity proxy training and aggregated proxy score	Rank candidates before sparse full training
Macro placement	Deterministic greedy HPWL probe	Filter ordering policies during LLM evolution
Medical HPO	Proxy data and proxy networks	Preserve hyper-parameter ranking at lower cost
Multi-agent RL	Evolutionary reward based on kin abundance	Let RL stand in for outer evolutionary search
Finite-temperature topology	Transfer-matrix eigenphases, proxy EGP, proxy index	Diagnose mixed-state topology
Solar helioseismology	Entropy proxy $S=P/\rho^{\Gamma_1}$	Infer BCZ thermodynamics from seismic data
AMO laser theory	Dipole-approximation proxy field	Useful in limited regimes, but not gauge-equivalent to a laser field

This broad usage also clarifies a central distinction. In some domains, the proxy is intentionally a computational surrogate for an objective that remains directly evaluable, albeit expensively. In others, the proxy is a diagnostic observable for a quantity that is not directly measured or is structurally obscured. The transfer-matrix proxy index for finite-temperature topology and the revised entropy proxy for the solar convective envelope belong to the latter class, whereas proxy training, proxy data, and proxy networks belong to the former (Pi et al., 2021, Buldgen et al., 11 Jun 2025).

2. Proxy-guided optimization and evolutionary search

In automated ML engineering, "ArchPilot: A Proxy-Guided Multi-Agent Approach for Machine Learning Engineering" organizes the approach around three agents: an Orchestration Agent, a Generation Agent, and an Evaluation Agent. The Orchestration Agent maintains the search tree and memory, runs an MCTS-inspired controller with decaying exploration, enforces global compute budgets, and manages restart policies when the proxy aggregator updates. The Generation Agent operates in Draft, Improve, and Debug modes; Draft returns a plan and a complete PyTorch script, Improve applies exactly one atomic, testable change, and Debug minimally repairs failing pipelines. The Evaluation Agent runs proxy training exactly one epoch on 10% of the training data, executes a small set of proxies, and aggregates them into a fidelity-aware score

$$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$

The proxy set includes one-epoch validation, noisy validation, and feature-dropout validation, and the proxy weights are refit after at least $k=5$ labeled pairs by ridge-regularized least squares followed by simplex projection. If $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$, the tree is restarted and reseeded from top-$k$ verified nodes (Yuan et al., 6 Nov 2025).

In chip physical design, "Order Matters: Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution" treats the placement order itself as the optimization object. The instance is $I=(M,E,G)$, the sequence $s$ is a permutation of macros, and the target is

$s^*=\arg\min_s \mathrm{HPWL}(A(s)),$

where $A$ is a deterministic greedy placer. The proxy is not the final full placement flow but a lightweight greedy probe with wirelength masks. Candidate ordering policies are emitted as code by an LLM, either as static scoring functions or dynamic context-dependent prioritizers. Proxy fitness is the mean HPWL across randomized probes,

$$\mathrm{fitness}(s)=\frac{1}{|V'|}\sum_{i\in V'} \mathrm{HPWL}(A(s;\mathrm{seed}_i)),$$

with $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$0, $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$1 workers, and per-strategy timeout $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$2. The evolution loop keeps top-$$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$3 elites, generates $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$4 new strategies per generation, and runs for $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$5 generations by default (Mo et al., 8 Jun 2026).

The reinforced evolution-based load balancing method applies the same pattern to a strict discrete feasibility regime. Assignments $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$6 must satisfy

$$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$7

for every node $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$8 and resource $$s(c)=\frac{\sum_{i=1}^{m}\lambda_i d_i x_i(c)}{\sum_{i=1}^{m}\lambda_i}, \qquad \sum_{i=1}^{m}\lambda_i=1,\quad \lambda_i\ge 0.$$9, and the objective is to minimize migration cost

$k=5$0

Because fewer than 4% of genotypes are feasible in the reported experiments, the method abandons a purely classical GA schema in favor of m-ary encoding, aggressive elimination of unstable genotypes, termination of weak individuals, and feasible-only migration as an analogue of random genetic drift. Population size is maintained in the interval $k=5$1, crossover selects 25% of individuals, mutation 5%, only 10% of infeasible individuals survive the feasibility termination step, and migration fills the deficit with randomly generated stable genotypes (Sliwko, 6 Nov 2025).

"Mimicking Evolution with Reinforcement Learning" pushes the logic further by making reinforcement learning itself the proxy for the outer evolutionary loop. Evolution via Evolutionary Reward defines kinship by

$k=5$2

and the immediate reward by the abundance of one’s replicators in the next population state,

$k=5$3

At death, a terminal reward aggregates discounted future abundance. The resulting RL objective equals expected evolutionary fitness, and the algorithm uses Evolutionary Value-Decomposition Networks to perform family-weighted temporal-difference learning over full trajectories rather than scalar episode-level fitness summaries (Abrantes et al., 2020).

3. Fidelity, calibration, and empirical performance

The proxy is only useful if it preserves the relevant ordering or diagnostic relation. ArchPilot makes that requirement explicit. The Evaluation Agent maintains a calibrated mapping from proxy vectors to true scores via regularized least squares, and sparse full training is reserved for calibration and confirmation rather than routine evaluation. On 75 Kaggle-style MLE-Bench tasks under a 2.5 GPU-hours-per-task budget, ArchPilot achieved Valid Submission 0.893, Above Median 0.293, Bronze+ 0.187, Silver+ 0.147, Gold+ 0.120, and Mean Normalized Ranking 0.6149, compared with AIDE’s 0.787, 0.240, 0.173, 0.133, 0.107, and 0.6953, and ML-Master’s 0.867, 0.267, 0.173, 0.147, 0.107, and 0.6535. The reported budget-sensitivity analysis further shows that improvements are most pronounced on high-difficulty tasks, where proxy-guided filtering reduces wasted full trainings (Yuan et al., 6 Nov 2025).

In OrderPlace, proxy fidelity is assessed against final greedy-placement outcomes. The paper states that the proxy mean HPWL strongly correlates with final HPWL and reports that the proxy correctly ranks the best strategy on several datasets. On ISPD 2005 benchmarks, OrderPlace reduces wirelength by 34.04% relative to WireMask-EA and 14.08% relative to EGPlace, achieves the best HPWL on 7 of 8 circuits, and attains the best average rank of 1.17. Against EGPlace, it is significantly better on 7 circuits under a Wilcoxon rank-sum test with $k=5$4. Strategy analysis shows that among 32 top-4 slots across datasets, 24 are LLM-generated, and all 8 top-1 strategies are dynamic and LLM-generated (Mo et al., 8 Jun 2026).

"The Power of Proxy Data and Proxy Networks for Hyper-Parameter Optimization in Medical Image Segmentation" makes ranking preservation the central design criterion. Proxy data is constructed by computing pairwise mutual information or normalized local cross-correlation within a task-specific label-cropped ROI and selecting samples with the lowest mean similarity under a budget. Proxy networks preserve the UNet inductive bias while reducing capacity by fixing the first encoder block to 4 channels, using 1 residual block per level, and setting the depth $k=5$5. For CT spleen segmentation, MI+Labelcrop achieved external BTCV Dice $k=5$6 versus $k=5$7 for random selection, and Pearson correlation to full-data training on external validation $k=5$8 versus $k=5$9 for random selection. Across proxy networks, validation Dice against the full network showed Pearson correlation $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$0 for all depth variants. In AutoML, proxy networks alone yielded up to $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$1 speed-up, and combined proxy data plus proxy network yielded up to $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$2 speed-up; for spleen, GPU-hours decreased from 1056 to 240, and for prostate from 772 to 248 (Nath et al., 2021).

The RL-based evolutionary setting reports a different form of fidelity: not rank preservation against a fixed target, but superior evolutionary outcomes relative to a classical evolutionary algorithm. In the asexual environment, E-VDN produced higher population sizes, longer lifespans, and higher birth rates than CMA-ES, and in direct competition CMA-ES families were consistently driven to extinction by E-VDN families. The sexual-environment experiments likewise showed consistent improvements in founders’ gene survival and reproduction, although the paper does not present formal sample-efficiency or convergence proofs (Abrantes et al., 2020).

The reinforced load-balancing method provides a further empirical pattern: even under strict feasibility and sparse feasible solutions, the proxy-evolutionary scheme remains competitive. In the smaller instance where FULLSCAN found the optimal cost 5 in under two minutes, EVOLVE achieved median 11 and average 12.95 and found the optimum in at least one run within 30 seconds. In the larger instance where FULLSCAN required about 53 hours, EVOLVE achieved median 19 and average 21.90 and found the optimum in one run within 10 minutes. In the largest instance where FULLSCAN did not finish within 7 days, EVOLVE reached median 26 and average 26.75, with best cost 19 found within 60 minutes (Sliwko, 6 Nov 2025).

4. Proxy diagnostics in physical sciences

In finite-temperature topological matter, the proxy is not a cheaper simulation but a structurally informative observable built directly from the transfer matrix $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$3, the ensemble Wilson loop. For a mixed state $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$4, the ensemble geometric phase is

$\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$5

but for time-reversal invariant insulators this quantity can become insensitive because $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$6 is real. The proposed transfer-matrix proxies therefore act directly on the spectrum of $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$7. The proxy EGP is

$\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$8

and the proxy index is determined by the parity or winding of the eigenphases $\|\bm{\lambda}^{*}-\bm{\lambda}\|_1>\varepsilon$9 between TRIM. In the BHZ and Kane–Mele models, these proxies remain quantized and topologically nontrivial for all finite temperatures and change only in the limit $k$0, in contrast to the Uhlmann phase, which in the BHZ model exhibits a finite-temperature topological transition (Pi et al., 2021).

In solar helioseismology, the revised entropy proxy

$k$1

is constructed to mimic the specific entropy plateau in the deep, nearly adiabatic convective envelope while avoiding the large surface amplitude of the earlier proxy $k$2. The inversion is performed with SOLA on the structural pair $k$3. The proxy is used to diagnose thermodynamical conditions at the base of the convective zone, with helioseismic $k$4 as a key constraint. The overshooting series from $k$5 to $k$6 deepens the convective zone and changes $k$7 and $k$8, but lowers the entropy plateau by only about 3% and conflicts with helioseismic BCZ and light-element constraints. By contrast, localized opacity increases of order $k$9 around $I=(M,E,G)$0 with widths $I=(M,E,G)$1 lower the plateau efficiently while keeping $I=(M,E,G)$2. The paper therefore states that the revised entropy proxy allows one to invalidate adiabatic overshooting as a solution to the solar modelling problem and strongly points towards the need for revised opacities (Buldgen et al., 11 Jun 2025).

AMO physics supplies the negative case. "Laser fields and proxy fields" argues that the dipole approximation replaces a source-free transverse laser field by a spatially uniform longitudinal proxy field $I=(M,E,G)$3 with $I=(M,E,G)$4. The remaining Maxwell equation implies a virtual source current,

$I=(M,E,G)$5

so the proxy field is not a gauge rewriting of a laser field. The gauge- and Lorentz-invariant fingerprints differ: a plane wave satisfies $I=(M,E,G)$6 and $I=(M,E,G)$7, whereas the proxy field has $I=(M,E,G)$8. The paper therefore concludes that numerical solution of the TDSE is exact for proxy fields, but not for laser fields, and that low-frequency magnetic corrections added inside the proxy framework cannot restore the physics of a propagating plane wave (Reiss, 2016).

5. Failure modes, misconceptions, and control mechanisms

A persistent failure mode is proxy misalignment. ArchPilot identifies several mechanisms: proxies may correlate weakly with true performance; weight fitting can become unstable in noisy or sparse regimes; over-reliance on proxies can misguide the search; and restarts can discard promising but underexplored branches. The mitigation stack is correspondingly explicit: hard-zero policy for unreliable proxies, ridge regularization, conservative initialization, sparse full-training calibrations, and restart when the weight update exceeds the threshold $I=(M,E,G)$9 (Yuan et al., 6 Nov 2025).

Medical HPO exhibits a closely related issue under a different guise. Proxy data is selected by diversity within a label-defined ROI, but if labels are noisy, unrepresentative, or exclude the discriminative context, the proxy can bias model selection. The paper therefore recommends external validation on withheld datasets and notes that very shallow proxy networks reduce correlation, even though $s$0 remains observed at $s$1. OrderPlace makes a parallel point for physical-design optimization: proxy mean HPWL is strongly correlated with final HPWL, but timing-driven or extreme designs may require richer proxies such as timing or congestion terms, and the present experiments optimize HPWL only (Nath et al., 2021, Mo et al., 8 Jun 2026).

The condensed-matter and AMO cases expose two common misconceptions about proxies. In finite-temperature topology, using $s$2 naively in a time-reversal invariant setting can obscure the topological content because the determinant is real; the eigenvalue spectrum of the ensemble Wilson loop must be interrogated directly. In laser theory, the proxy-field misconception is more severe: the dipole field is sometimes serviceable, but it is not gauge-equivalent to a transverse plane wave, has no Poynting flux, and acquires an unphysical static-field intuition as $s$3, whereas a real laser field propagates at $s$4 at all frequencies (Pi et al., 2021, Reiss, 2016).

The proxy can also shift computational burden rather than eliminate it. In EvER, the reward requires explicit genomes and kinship computation, and the family-weighted joint value induces $s$5 aggregations per environment step if implemented naively. In reinforced load balancing, the feasibility filter is so strict that migration dominates wall-clock time; the paper reports that migration consumed more than 70% of execution time because the empirical feasible acceptance rate was below 4% (Abrantes et al., 2020, Sliwko, 6 Nov 2025).

6. Relation to evolutionary methodology and future directions

The relation to classical evolutionary algorithms is heterogeneous. ArchPilot explicitly states that its search augments a tree search rather than classic evolutionary algorithms, but is evolutionary in spirit because it iteratively mutates and refines candidates while selecting via proxy scores. OrderPlace is population-based and LLM-driven: elite code policies are retained, mutated, and recombined through prompting, while a deterministic proxy probe provides selection pressure. The reinforced load-balancing method is structurally closer to a classical GA but reinforces it with feasibility-aware migration and aggressive culling. EvER goes in the opposite direction by dispensing with an outer evolutionary loop and using RL to optimize a reward explicitly aligned with replicator abundance (Yuan et al., 6 Nov 2025, Mo et al., 8 Jun 2026, Sliwko, 6 Nov 2025, Abrantes et al., 2020).

A plausible unifying interpretation is that the Evolution-Proxy Approach becomes most effective when the true objective is either too expensive to evaluate repeatedly or too indirect to observe directly, but the problem still admits a cheaper signal that preserves the relevant ordering or structure. The computational cases make this condition concrete: full training runs are expensive, macro placement is path-dependent, and exhaustive combinatorial search is infeasible. The diagnostic cases make the same point in inverse form: mixed-state topology is encoded more robustly in transfer-matrix spectral flow than in a naive scalar phase, and BCZ thermodynamics is more directly constrained by an entropy-like plateau than by less targeted observables (Yuan et al., 6 Nov 2025, Mo et al., 8 Jun 2026, Pi et al., 2021, Buldgen et al., 11 Jun 2025).

The reported future directions preserve this logic while extending the proxy repertoire. ArchPilot notes that crossover of architectural motifs could be integrated while retaining proxy-guided evaluation and MCTS-based allocation. OrderPlace identifies timing- and power-aware objectives, richer proxy signals, and integration with full PPA flows or end-to-end learning as next steps. The entropy proxy is proposed as a diagnostic for solar-like stars when combined with asteroseismic BCZ measurements, and ensemble-Wilson-loop proxies are suggested for other symmetry classes such as mirror, chiral, or particle–hole symmetries (Yuan et al., 6 Nov 2025, Mo et al., 8 Jun 2026, Buldgen et al., 11 Jun 2025, Pi et al., 2021).

In that sense, the Evolution-Proxy Approach is best understood not as a single algorithmic family but as a disciplined strategy for replacing unattainable direct evaluation with a proxy that is cheap enough to be operational, structured enough to preserve the relevant invariants or rankings, and explicitly monitored for drift. Its successes arise when those conditions are met; its failures, equally instructive, arise when the proxy alters the problem’s governing semantics rather than merely approximating its evaluation.