PaperlayerView exact revision on arXiv

Source-linked AI summary

ADAPT for Science: A Multi-Agent System for Self-Evolving Scientific Optimization

Zhaoran Feng, Yuchen Ni, Jianan Liu, Lemao Liu, Dejing Dou, Kairui Feng

arXiv:2610.30851v1cs.AIcs.MAphysics.optics

TL;DR

Scientific optimization must find feasible, high-performing designs with costly evaluations, sparse feasible regions, and interacting parameters, while general-purpose methods do not jointly account for scientific priors, physical constraints, and limited budgets. ADAPT uses historical designs to guide where to search and execution feedback to evolve how to search. It achieves lower loss than the strongest classical and general-purpose agentic evolution baselines across eight held-out detector topologies.

  • Problem

    Scientific optimization faces costly evaluations, interacting parameters, sparse feasible solutions, and strict physical constraints, while existing general-purpose approaches do not explicitly combine scientific priors, feasibility, and limited budgets.

    Source: §1 · p. 1

  • Method

    ADAPT uses historical designs to adapt target-specific initialization and resource allocation, then uses execution feedback to evolve the optimization procedure.

    Source: §1 · p. 2

  • Results

    ADAPT achieves a mean loss of 0.204±0.101, reducing mean loss by 54.9% versus the strongest classical optimizer and 62.2% versus the strongest general-purpose agentic evolution baseline, and has the lowest loss on all eight held-out topologies.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The results support adapting the scientific search policy itself, with ablations indicating contributions from adaptive initialization, optimizer evolution, and science-specific guidance.

    Source: §6 · p. 9

    The evaluation establishes effectiveness in gravitational-wave detector design but does not demonstrate universality across scientific disciplines.

    Source: §5 · p. 8

Abstract

from arXiv · show

Scientific discovery often requires identifying a small number of high-performing designs in vast, constrained search spaces. This is difficult when evaluations are costly, feasible solutions are sparse, parameters are strongly coupled, and physical constraints are strict. Recent AI for AI methods can automatically improve algorithms and optimization procedures. However, general purpose approaches are not designed around scientific priors, physical constraints, and limited evaluation budgets. We introduce ADAPT for Science, a multi-agent framework that treats scientific optimization as two coupled decisions: where to search and how to search. ADAPT uses historical designs to identify promising and diverse initial conditions, progressively concentrates computation on favorable search directions, and evolves the optimization procedure from execution feedback. We evaluate ADAPT on the design of gravitational wave detectors using a differentiable physics simulator and eight unseen detector topologies. Under matched scientific evaluation budgets, ADAPT achieves a mean loss of 0.204 +/- 0.101, reducing loss by 54.9% relative to the best classical optimizer and by 62.2% relative to the best general purpose agentic evolution baseline. ADAPT achieves the lowest loss on all unseen topologies among the evaluated baselines. ADAPT provides a step toward scientific optimization systems that continuously learn from prior experience and their own search outcomes.

1 Introduction

Scientific optimization must choose not only what to evaluate but where to begin and how to search, especially when evaluations are costly and feasible designs are scarce. ADAPT jointly adapts these search decisions using historical experience and execution feedback.

1 Introduction

  • Costly evaluations, strongly interacting parameters, sparse feasible solutions, and strict constraints make choosing what to try next a central challenge.A poor initialization or unsuitable optimization strategy can consume substantial computation before reaching a useful detector design.Source: §1 · p. 1
  • Existing approaches typically adapt scientific candidates while fixing the search procedure, or evolve procedures without explicitly modeling reusable scientific experience and target-dependent initialization.The gap also includes physical feasibility and limited scientific evaluation budgets.Source: §1 · p. 1
  • ADAPT optimizes the complete scientific search policy as two coupled decisions: where to search and how to search.Historical designs guide promising, diverse starting points and evaluation allocation, while execution feedback evolves the optimization procedure.Source: §1 · p. 2
  • 54.9% and 62.2% mean-loss reductions: ADAPT outperforms the strongest classical optimizer and general-purpose agentic evolution baseline, respectively, under matched budgets.It achieves the lowest loss on all eight held-out detector topologies among evaluated baselines.Source: §1 · p. 2

2 Related Work

Prior work uses observations to select scientific candidates and increasingly explores improving algorithms themselves. ADAPT connects these directions by combining target-specific scientific priors with procedure adaptation.

2 Related Work

  • Bayesian optimization, active learning, foundation models, and autonomous agents use observations, historical data, planning, or tool use to support scientific search and workflows.Examples include chemical reaction optimization and autonomous experimentation [20].Source: §2 · p. 2
  • Learned optimizers, symbolic program search, and language-model-driven evolution show that optimization rules and algorithms can be improved through experience or evaluator feedback.Examples include FunSearch, EoH, ReEvo, and MCTS-AHD [18] [33] [35].Source: §2 · p. 2
  • ADAPT treats the complete scientific search policy as the adaptation target, using historical designs for where to search and execution feedback for how to search.It jointly adapts initialization, resource allocation, and optimization dynamics under physical constraints and limited budgets.Source: §2 · p. 2

3 ADAPT FOR SCIENCE

ADAPT adapts scientific search through two coupled decisions: selecting where to begin and evolving how to optimize. Historical experience informs target-specific choices, while held-out validation tests the resulting policy under the scientific objective and constraints.

3 ADAPT FOR SCIENCE

  • ADAPT adapts a target-specific search policy through a selection evolver that decides where to search and an optimizer evolver that uses execution feedback to decide how.An independent validator evaluates the policy on held-out targets under the original objective and physical constraints.Source: §3 · p. 2, Figure 1 · p. 3

3.1 Problem Formulation

  • The search policy pairs W, which selects the initial design, with O, which generates the trajectory; its budgeted objective is the best feasible loss reached within cumulative evaluation cost B.Unlike conventional optimization with W and O fixed in advance, ADAPT adapts both to each target.Source: §3.1 · p. 3

3.2 Historical Scientific Experience

  • A reusable library contains 15,409 valid topology–design records and ten established optimization procedures, with held-out targets excluded from its construction and subsequent strategy development.Design records supply candidate initializations; the separate baseline library supplies reference search behavior for policy evolution.Source: §3.2 · p. 3

3.3 Target-Aware Retrieval and Initialization

  • For each target, ADAPT retrieves 64 distinct eligible historical sources using fixed-position topology distance, then transfers compatible parameters by semantic matching and uses frozen training-set medians for unmatched values.The distance assigns 0 to identical symbols, 0.25 to same-family orientation changes, and 1 to different component families; same-topology sources are excluded.Source: §3.3 · pp. 3–4

3.4 Selection Evolution

  • The Selection Evolver searches for a policy that chooses one initialization under a fixed budget; intermediate screening, pruning, and budget-allocation strategies are left to evolution.Policies can use historical attributes, target evaluations, feasibility, constraint violations, and optimization progress to decide what to evaluate, retain, or discard.Source: §3.4 · p. 4

3.5 Optimizer Evolution

  • The Optimizer Evolver revises executable optimization programs under the same scientific budget and numerical constraints, ranking candidates by quality, anytime performance, feasibility, and improvement over initialization.Reported results use the native scientific objective and feasibility criteria, rather than the strategy-evolution score.Source: §3.5 · p. 4

3.6 Independent Validation

  • Held-out targets are excluded from retrieval, baseline selection, policy evolution, and model selection, and their outcomes are never returned to either evolver.Comparisons use the same scientific budget and report native objective, physical feasibility, and anytime optimization performance.Source: §3.6 · p. 4

4 Results

The results show how ADAPT evolves selection and optimization procedures, then test their combined performance on eight held-out detector topologies and ablations of the system’s components.

4.1 Selection evolution discovers progressive search

  • The evolved selection policy reduces mean loss to 3.0054 and shifts from 64 starts to eight, then two, then one, while retaining a protected exploratory branch.It first combines four high-quality feasible candidates with four diversity-preserving candidates, then uses target-specific probes before concentrating the remaining budget.Source: §4.1 · p. 5, Figure 3 · p. 5

4.2 Optimizer evolution discovers budget-aware search dynamics

  • Optimizer evolution retains Adam’s basic structure while testing numerical safeguards, time-dependent updates, stopping behavior, and shorter optimizer memory under a finite evaluation horizon.The selected procedure uses β1 = 0.7, scales updates by the remaining budget, skips non-finite losses or gradients, and rejects non-finite proposals.Source: §4.2 · p. 6, Figure 4 · p. 6

4.3 Overall Performance on Held-Out Topologies

  • ADAPT reaches a mean loss of 0.204 ± 0.101, reducing loss by 54.9% versus NAAdam and 62.2% versus GEPA, and achieves the lowest loss on all eight held-out topologies.These comparisons use matched evaluation budgets.Source: §4.3 · pp. 6–7

4.4 Representative Detector Optimization

  • On a representative held-out detector, ADAPT reaches loss 0.155 versus 0.666 for Adam, a 76.7% reduction, while suppressing three optical branches whose optimized contributions approach zero.The remaining optical structure is preserved, and the comparison uses the same evaluation budget.Source: §4.4 · p. 7

4.5 Ablation of Where and How to Search

  • Removing evolved start selection raises mean loss to 0.423 ± 0.243, while replacing the evolved optimizer with Adam yields 0.256 ± 0.094 and removing science-specific guidance yields 0.303 ± 0.128.Full ADAPT reduces these respective losses by 51.8%, 20.3%, and 32.7%; the guidance ablation improves all eight targets.Source: §4.5 · pp. 7–8

5 Limitations and Future Work

The evidence is bounded to gravitational-wave detector design, while proposed future work extends evaluation to other scientific settings and addresses library coverage, computational overhead, and validation.

5 Limitations and Future Work

  • The study establishes ADAPT's effectiveness in gravitational-wave detector design, but does not demonstrate universality across scientific disciplines.The evaluation uses a differentiable simulator and a large historical design library.Source: §5 · p. 8
  • ADAPT depends on experience-library quality and coverage; retrieval may lose advantage for targets far outside the observed distribution, and overhead is justified when evaluations are costly.The authors recommend reducing overhead for cheaper optimization problems.Source: §5 · p. 8
  • Future work includes non-differentiable and experimental settings, online library updates, more efficient strategy evolution, and validation against real measurements.The proposed efficiency approaches include learned transfer, uncertainty-aware allocation, and meta-optimization.Source: §5 · p. 8

6 Conclusion

ADAPT optimizes the complete scientific search policy by combining historical experience with execution feedback. On eight held-out detector topologies, it lowers mean loss against both baseline classes under matched budgets.

6 Conclusion

  • ADAPT treats the complete scientific search policy as the optimization target, using historical experience for initialization and resource allocation and execution feedback to adapt the optimizer.The policy couples decisions about where and how to search.Source: §6 · p. 9
  • ADAPT achieves mean loss 0.204±0.101 on eight held-out topologies, reducing mean loss by 54.9% versus the strongest classical optimizer and 62.2% versus the strongest general-purpose agentic evolution baseline.It also achieves the lowest loss on all eight topologies among evaluated baselines.Source: §6 · p. 9
  • The results support a view of scientific automation in which the search procedure itself adapts, rather than scientific parameters being optimized under a fixed workflow.ADAPT combines prior experience with execution feedback to refine search initialization and conduct.Source: §6 · p. 9

AI use statement

  • Large language models propose and revise candidate search policies and optimization procedures using execution feedback, while scientific simulators and objective functions evaluate candidates.Source: AI use statement · p. 9
  • Generative AI tools were used for language editing and manuscript polishing, with authors reviewing edits and taking responsibility for the final manuscript.Source: AI use statement · p. 9

Reproducibility statement

  • The supplementary material provides the scientific objective, data construction, procedures, evaluation protocol, implementation settings, and further details for reproducing the study.Source: Reproducibility statement · p. 9

A Scientific Objective and Simulator Specification

This section specifies ADAPT’s finite-budget detector-design objective, simulator, constraints, and optimization coordinates, and explains how historical designs and simulator feedback shape target-specific selection policies and diverse starting points. It also describes how those policies balance promising regions, target-side evaluation, and exploration while allocating a shared evaluation budget, distinguishing policy development from frozen target-time execution.

A.1 Finite-Budget Scientific Optimization

  • Under finite evaluation budgets, ADAPT jointly chooses an initialization policy and optimizer, scoring the best feasible design found before the budget is exhausted.The recorded best feasible solution is monotonic with additional budget; this does not imply that every intermediate iterate improves.Source: §A.1 · p. 13

A.2 Detector Topology Representation

  • Each target pairs a size-three UIFO topology (nine internal and twelve boundary symbols) with topology-dependent continuous variables and bounds mapped to unconstrained coordinates for optimization.The active parameter dimension can vary by topology, and some physical properties share an optimization scalar.Source: §A.2 · p. 13, Table 4 · p. 14, §A.3 · p. 14

A.4 Differentiable Physics Simulator

  • A differentiable interferometer simulator evaluates quantum-noise, amplitude-modulation, and frequency-modulation responses on 100 log-spaced frequencies from 20 to 5000 Hz, enabling gradients for a broadband geometric-mean sensitivity objective.The objective averages logarithmic sensitivity ratios against a fixed reference curve.Source: §A.4 · pp. 14–15, §A.5 · p. 15

A.6 Physical Constraints and Power Penalty

  • The native objective adds differentiable penalties for optical-power violations, but final design validity is determined independently by physical feasibility constraints.The penalty guides search; it does not replace the feasibility test.Source: §A.6 · pp. 15–16

B.1 Historical Design Library

  • The historical library combines 9,342 official training representatives and 6,067 compatible GraviTune designs; 1,318 incompatible records cannot provide valid transferred initializations.The library serves as reusable prior scientific experience, not as a language-model training set.Source: §B.1 · p. 17

B.2 Target-Aware Topology Retrieval

  • For each target, the matcher retrieves 64 closest compatible designs using positionwise topology distance that distinguishes exact matches, orientation changes, and component-family substitutions.Exact target copies are excluded, and retrieval skips incompatible records.Source: §B.2 · p. 17

B.3 Topology-Compatible Semantic Transfer

  • Historical values transfer only when variable semantics and target bounds match, preserving coupled component groups; unmatched coordinates use frozen training-set normalized medians selected hierarchically.For coupled variables spanning component types, the component-specific prior is bypassed in favor of property-level or global statistics.Source: §B.3 · p. 18, §B.4 · p. 18

B.5 Target-Specific Screening of Retrieved Designs

  • After one budget-charged target evaluation, ADAPT ranks feasibility before loss and, when at least eight candidates are feasible, selects four lowest-loss starts plus four greedily diverse designs.This screening checks transferred designs on the target because historical performance may not carry over to a different topology.Source: §B.5 · pp. 18–19, §B.6 · p. 19

B.6 Quality–Diversity Construction of the Eight-Start Set

  • The eight-start set uses quality and diversity selection, retains feasible candidates first when fewer than eight are available, and terminates if fewer than eight numerically valid candidates exist.Source: §B.6 · p. 19

B.8 Isolation of Held-Out Targets

  • For held-out targets, topology queries permitted historical experience, while optimized designs, final losses, future trajectories, and validator outcomes remain excluded from historical supervision.Source: §B.8 · pp. 19–20

C.1 Selection Policy as a Search Object

C.2 Agentic Workflow Pool

  • During development, candidate selection rules are explored across four agentic workflow families, two language-model backends, and two fixed reference optimizers; the retained policy is frozen for held-out evaluation.Source: §C.2 · p. 21

C.3 From 64 Retrieved Designs to Eight Search Starts

  • Of 64 retrieved historical designs, only eight target-evaluated starts enter progressive optimization: four lowest-loss feasible candidates and four additional starts chosen for diversity when feasible candidates suffice.Source: §C.3 · p. 21

C.4 Progressive 8 →2 →1 Allocation

  • The scheduler screens eight historical starts and continues the two strongest trajectories, ranking feasible branches by objective and otherwise by physical violation before finite objective.The reported 21-minute continuation is cumulative active optimization time for one surviving trajectory, not elapsed time from the run's start.Source: §C.4 · p. 22, §C.7 · p. 23

C.5 Protected Exploratory Branch

  • A protected exploratory trajectory runs alongside historical branches and is not discarded for weaker early performance, preventing all pre-optimization budget from following early historical rankings.Source: §C.5 · p. 22

C.9 Parameter-Only Handoff to the Optimizer Evolver

  • Under one shared wall-clock deadline, ADAPT selects the best admissible observed vector from the winning historical branch and hands that parameter vector—not screening state—to the final optimizer.Source: §C.6 · p. 23, §C.8 · pp. 23–24, §C.9 · p. 24

C.10 Complete Selection Algorithm

  • Algorithm 3 summarizes target-time execution of the frozen selection policy, while Algorithm 2 separately summarizes its meta-level development.Source: §C.10 · p. 24

C.11 Why Progressive Allocation is Part of the Search Policy

  • The progressive schedule avoids both uniform compute across starts and immediate commitment to the historically best candidate, which would forfeit most target-side evaluation.It allocates computation among competing starting regions without claiming short runs perfectly predict long-horizon optimizer performance.Source: §C.11 · p. 24
  • Promising retrieval, diversity, short target-side probes, and protected exploration balance historical guidance against early collapse when selecting starting regions.The resulting initialization is a finite-budget decision, not simply the nearest historical design or the minimum-loss historical record.Source: §C.11 · p. 24

C.12 Held-Out Evaluation Boundary

This section defines the held-out evaluation protocol and reporting procedures for ADAPT-Adam, including its initialization and execution under shared deadlines. It separates development-time optimizer evolution from held-out evaluation, where the frozen selection strategy is applied without feeding test outcomes back into policy development.

C.12 Held-Out Evaluation Boundary

  • Held-out performance is not used to revise the frozen selection policy, keeping evaluation separate from strategy development.The target topology and simulator remain available, and target evaluations consume the declared scientific budget; outcomes are not returned to the Selection Evolver or added to the historical library.Source: §C.12 · p. 25

D.1 Optimizer Programs as Meta-Level Search Objects

  • ADAPT evolves executable optimizer programs as search objects, selecting them by measured performance on the scientific objective rather than by judgments of their source code.The Optimizer Evolver receives a target-specific initialization from the Selection Evolver; the search can change optimizer components while leaving the simulator, objective, physical constraints, and held-out protocol fixed.Source: D Optimizer Evolution: Learning How to… · p. 25, §D.1 · pp. 26–27, §D.2 · p. 27

D.2 Execution-Based Optimizer Evolution

  • Candidate optimizers are ranked using execution feedback on feasible quality, anytime performance, feasibility, and improvement over initialization.The composite score is for optimizer evolution, not the final scientific result, which is reported using the native detector objective and physical feasibility criteria.Source: §D.2 · p. 27

D.8 Transactional Invalid-State Protection

  • Invalid losses, gradients, or proposals restore the last valid parameters without committing moment changes, and the bias-correction counter advances only for accepted updates.This transactional handling prevents rejected iterations from contaminating optimizer state and keeps the counter aligned with updates that affect the scientific trajectory.Source: §D.3 · p. 27, §D.8 · p. 30, §D.9 · pp. 30–31

D.3 From Standard Adam to ADAPT-Adam

  • ADAPT-Adam remains an Adam-family local optimizer but conditions its finite-budget dynamics on the remaining horizon and numerical validity.Its changes include shorter first-moment memory, finite-horizon update scaling, and transactional protection against invalid states.Source: D Optimizer Evolution: Learning How to… · p. 25, §D.3 · p. 27

D.5 Short-Memory First-Moment Dynamics

  • ADAPT-Adam clips gradients globally before applying Adam moment recursions, and its shorter first-moment memory gives greater relative weight to recent gradients.The memory-kernel characterization does not imply a general convergence-rate guarantee.Source: §D.4 · p. 28, §D.5 · p. 29

D.7 Why Update Scaling is Applied After Adam

  • ADAPT-Adam scales the proposed Adam update—not the input gradient—so the moment estimator still summarizes clipped gradients while displacement becomes more conservative as budget runs down.The schedule’s analytical average is not a measured average over scientific evaluations, which need not be uniformly distributed in wall-clock time.Source: §D.6 · p. 29, §D.7 · p. 30

D.10 Warm Start from the Selection Evolver

  • The final optimizer warm-starts from the parameter vector selected by the Selection Evolver, without implicitly transferring the screening optimizer’s moment state.The transferred information is the selected detector parameters.Source: §D.10 · p. 31
  • The finite-horizon schedule’s remaining-time reference is initialized at the selection boundary, after the Selection Evolver has consumed part of the shared end-to-end budget.Source: §D.10 · p. 31

D.11 Best-Feasible Incumbent Tracking

  • The reported result is the best physically admissible evaluated design completed before the shared deadline, not necessarily the optimizer’s final parameter state.An unevaluated proposed update cannot become the reported result.Source: §D.11 · p. 31

D.12 Deadline Semantics

  • An evaluation counts for scientific comparison only if it completes before the global pipeline deadline, preventing results from being reported after the declared budget expires.The finite-horizon scale uses the final optimizer’s remaining time, while evaluations remain subject to the original deadline.Source: §D.12 · p. 32

D.13 Complete ADAPT-Adam Update

  • An ADAPT-Adam update is accepted only when its scientific loss, gradient, and proposed parameter vector are finite, with optimizer state committed alongside accepted parameters.Algorithm 4 gives the complete target-time optimizer used after the selection stage.Source: §D.13 · p. 32, §D.14 · p. 32

D.15 Meta-Level Optimizer Evolution

  • The final update rule is produced by optimizer evolution rather than being a manually selected modification to Adam.Source: §D.15 · p. 32

D.16 Separation Between Optimizer Generator and Evolved Optimizer

  • During development, an agentic workflow proposes code changes from execution feedback, but the resulting numerical optimizer—not a language model—produces the scientific trajectory.The distinction separates the optimizer generator from the optimizer executed against the differentiable simulator.Source: §D.16 · p. 33

D.17 What ADAPT-Adam Does Not Change

  • ADAPT-Adam changes how candidate parameter vectors are proposed, not the scientific problem or the criteria used to judge candidates.The detector topology, simulator, objective, physical bounds, feasibility constraints, evaluation definition, and held-out budget remain fixed.Source: §D.17 · p. 33

D.18 Held-Out Execution Boundary

  • Held-out evaluation runs a frozen optimizer from the frozen selection policy’s initialization, and held-out final losses are used only for reporting, not further optimizer evolution.The optimizer may use current loss, gradient, feasibility, and remaining budget during a run.Source: §D.18 · p. 33

D.19 Implementation Provenance and Reproducibility

This section documents ADAPT’s development and evaluation procedures to support auditability and reproducibility, including optimizer provenance, prompts, and constraints on valid execution and comparisons. It specifies code-completeness requirements and separates method development from held-out validation, with final procedures frozen before evaluation on the eight held-out detector topologies.

D.19 Implementation Provenance and Reproducibility

  • Each retained optimizer version is traceable through its source and hash, parent and revision, generation workflow, prompt and feedback, and evaluation record.The record also includes numerical configuration, development targets, execution score, acceptance decision, and the frozen program used for held-out validation.Source: §D.19 · p. 34
  • The final runs use archived development identifier GEPA GPT 4, while the evolved numerical method is reported as ADAPT-Adam.Exact prompts are in Appendix E; hardware, runtime, model, and optimizer configuration is in Appendix F.Source: §D.19 · p. 34

E Agent Prompts and Evolution Constraints

  • The two evolution stages have distinct edit scopes: selection improves the 64-to-1 initialization policy, while optimizer evolution changes the numerical procedure after initialization.Retrieval, parameter transfer, evaluator, and downstream optimizer remain fixed for selection; topology, simulator, objective, physical constraints, and protocol remain fixed for optimizer evolution.Source: E Agent Prompts and Evolution… · p. 34

E.3 Anonymous Review and Artifact Release

  • During anonymous review, prompts are provided for audit, but implementation code and executable artifacts are withheld; release is planned upon acceptance.The prompts are reproduced verbatim without rewriting or post-processing in the reported algorithm-evolution runs.Source: §E.3 · p. 34, E Agent Prompts and Evolution… · p. 34

E.1 Selection Evolver Prompt

  • Figure 7’s selection prompt defines the 64-candidate input, single-clock budget, probe interface, admissible information, state continuation, and leakage constraints.It specifies the frozen scientific contract and selector interface across development and held-out evaluation.Source: §E.1 · p. 34, Figure 7 · p. 35

E.3 Anonymous Review and Artifact Release

  • The selection prompt requires evidence-backed changes to one mechanism, preservation of the frozen evaluator contract, robust fallbacks, and no topology-specific or held-out information.It also requires state continuity, public-API use, and comparisons under matched conditions rather than unsupported claims.Source: §E.3 · pp. 36–37

E.2 Optimizer Evolver Prompt

  • Figure 8’s optimizer prompt fixes the scientific objective and evaluator while allowing changes to the executable optimization procedure within its interface and wall-clock budget.The prompt specifies the development objective, held-out protocol, optimizer interface, and permitted execution environment.Source: §E.2 · p. 34, Figure 8 · p. 38

E.3 Anonymous Review and Artifact Release

  • Optimizer evolution must preserve the global budget, valid evaluated incumbents, state continuity, data isolation, and offline execution; unmatched runs cannot support direct attribution.The prompt also cautions that fixed seeds do not establish negligible noise and calls for paired or repeated matched evaluations before significance or uncertainty claims.Source: §E.3 · p. 39
  • Parseability is a hard requirement: edits must leave the complete Python file runnable, because malformed output is discarded without measurement.The instructions require checking syntax before correctness and keeping the required entry point, signature, and imports intact.Source: §E.3 · p. 40
  • Responses must fit the wall-clock budget and focus on the chosen mechanism rather than restating the prompt or listing rejected alternatives.The code deliverable must be complete, with other blocks limited to a few sentences.Source: §E.3 · p. 40

F Experimental Protocol and Compute Accounting

  • Development uses development targets, while the eight held-out detector topologies are reserved for final evaluation after procedures are frozen.The appendix states that it introduces no additional experiments and that quantitative results correspond to experiments reported in the main paper.Source: F Experimental Protocol and Compute… · p. 41, Table 5 · p. 41

F.1 Development and Held-Out Evaluation Regimes

The development and held-out evaluation stages use distinct computational regimes. Held-out validation outcomes are kept separate from algorithm development.

F.1 Development and Held-Out Evaluation Regimes

  • The two evaluation stages operate under distinct computational regimes.Source: §F.1 · p. 41
  • H100 validation outcomes are not returned to either evolver and cannot influence algorithm development.Source: §F.1 · p. 41
  • The distinction between the regimes is summarized in Table 5.Source: §F.1 · p. 41

F.2 Held-Out Validation Compute

The appendix accounts for held-out and development compute separately, enforces a shared evaluation deadline, isolates held-out outcomes from development, and clarifies which objects ADAPT learns.

F.2 Held-Out Validation Compute

  • 32 H100 GPU-hours cover final held-out validation; concurrent execution on eight GPUs takes approximately four hours of validator wall-clock time, excluding overhead.Source: §F.2 · p. 41

F.3 Development Compute Accounting

  • Development compute accounting treats the selection-evaluation quantity as a budget upper bound, not measured runtime, and reports H100 and RTX 4090 compute separately.Individual executions can end before their maximum budget, and optimizer-evolution rounds add development computation.Source: §F.3 · pp. 41–42

F.4 Single-Clock Evaluation

  • All comparisons use matched seeds, hardware, precision, and total wall-clock budgets; selection consumes time from the shared deadline rather than adding a separate warmup budget.The final optimizer uses only the time remaining under that deadline.Source: §F.4 · p. 42

F.5 Held-Out Isolation

  • Held-out outcomes are excluded from policy and optimizer evolution, model and hyperparameter selection, and later revision; retrieval also excludes the target itself.The target topology is available at test time, but its held-out solution and validation outcome are not development information.Source: §F.5 · p. 42

F.6 What Is and Is Not Trained

  • ADAPT learns search procedures rather than a detector-surrogate model; frozen procedures run directly against the simulator without held-out LLM inference.The Selection Evolver develops a target-selection policy, while the Optimizer Evolver develops an executable numerical optimization algorithm.Source: §F.6 · p. 42

G Complete Held-Out Results

The appendix consolidates held-out results across eight detector topologies and clarifies how to interpret their reported variation.

G Complete Held-Out Results

G.3 Selection and Optimizer Ablations

The ablations attribute gains to adaptive initialization, optimizer evolution, and science-specific guidance, while appendix sections document held-out visual comparisons and reporting conventions. The appendix also specifies reproduction requirements, runtime information boundaries, and how reported results should be interpreted.

G.3 Selection and Optimizer Ablations

  • Adaptive initialization, optimizer evolution, and science-specific guidance each improve a component of the search, but gains are not uniform across targets.Fixed+Adam is lower-loss than ADAPT on T2, while Direct+GEPA is marginally lower on T6.Source: §G.3 · p. 44

H Complete Held-Out Detector Topology Atlas

H.1 Held-Out Topology Definitions

  • Topology identifiers map to strings specifying the 3 × 3 internal optical layout and boundary components.Source: Table 9 · p. 46

I.1 Reproduction Boundary

  • Reproduction requires the scientific objective and parameterization, historical retrieval, selection and optimizer policies, global wall-clock rules, and held-out reporting procedure.Source: §I.1 · p. 46

I.2 Information Available at Held-Out Runtime

  • At held-out runtime, the procedure receives the current parameter vector, objective and gradient, feasibility information, incumbent, completed-evaluation count, and remaining wall-clock budget.Source: §I.2 · pp. 46–47

I.3 Best-Feasible Reporting Rule

  • Reported results use the best physically feasible point actually evaluated and admitted before the deadline, excluding unevaluated updates and late results.Source: §I.3 · p. 47

I.4 Interpretation of Variability

  • Reported variability refers to differences across the eight held-out topologies, not repeated-seed uncertainty, standard error, or statistical significance.The paper bases conclusions on complete per-topology results as well as aggregate means.Source: §I.4 · p. 47

I.5 Scope and Limitations

The reported evidence is limited to differentiable gravitational-wave detector optimization, and several factors constrain how broadly its results can be interpreted. These include dependence on historical-library coverage, development-time overhead, simulation-only validation, and the absence of a global convergence guarantee for ADAPT-Adam.

I.5 Scope and Limitations

  • The evaluation establishes ADAPT in one domain: differentiable gravitational-wave detector design with reusable historical experience and explicit physical constraints.Source: §I.5 · p. 47
  • Retrieval benefits may decrease for targets far outside the observed design distribution because they depend on the coverage and quality of historical experience.Source: §I.5 · p. 47
  • Evolving both initialization and optimization procedures adds development-time computational overhead relative to using one fixed optimizer.Source: §I.5 · p. 47
  • Simulation-based validation does not establish performance under real experimental noise, hardware imperfections, or non-differentiable laboratory feedback.Source: §I.5 · p. 48
  • ADAPT-Adam is an empirically discovered finite-budget optimizer, and the work makes no claim of a global convergence guarantee for its update rule.Source: §I.5 · p. 48
Loading 2610.30851v1…