acceptodds
Under review as a conference paper at ICLR 2027

BayesianAgent: Guiding Agentic Evolution with Bayesian Optimization

Abstract

Agents that iteratively improve executable solutions must allocate a limited experimental budget across a growing set of candidates. Most existing systems use rule-based selection informed by observed performance and experimental history. However, these rules do not explicitly learn to predict the gain from the next expansion and quantify uncertainty in that prediction. We introduce **BayesianAgent**, a Bayesian optimization approach that learns where to invest further experiments from completed outcomes. Its central intuition is that an experiment can inform the improvement potential of related candidates, including those with little direct evidence. A Gaussian process learns expansion gains and their uncertainty, sharing evidence across candidates. Joint Thompson sampling then selects the next candidate to expand, balancing promising refinements with opportunities whose value remains uncertain. We compare all methods using Muse Spark 1.3 under matched experimental budgets. Across 15 AlphaEvolve mathematical optimization tasks and 22 MLE-bench Lite tasks, BayesianAgent outperforms the MLEvolve baseline. It achieves the best AlphaEvolve average rank (1.83), the highest MLE-bench Lite gold rate (**45.45%**), and a joint-highest medal rate (**72.73%**) among evaluated methods. Further analysis shows that its predictions rank observed expansion gains more accurately than MLEvolve's historical rewards. These results support learning from experimental outcomes to guide subsequent allocation in agentic evolution. Our code is available at https://anonymous.4open.science/r/BayesianAgent-anonymous-6EBB.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic AI, Large Language Model, Bayesian Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_k7qBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces BayesianAgent, which learns normalized expansion gains within the MLEvolve solution graph using a Gaussian-process surrogate informed by parent performance, refinement history, ancestry, and LLM potential ratings. It combines predicted gains with parent scores and uses joint Thompson sampling to select the next expansion. Evaluation covers AlphaEvolve, MLE-bench Lite, and supplementary AutoLab tasks.

Critical strengths
  1. Clear formulation and coherent modeling. The distinction between current parent quality, expansion gain, and expected child quality is well motivated. The resulting model is understandable and offers a concrete approach to a relevant allocation problem.
  2. Broad evaluation with encouraging held-out results. The study covers 15 mathematical tasks and 22 ML-engineering tasks, with broadly favorable MLE-bench scores against MLEvolve. Importantly, MLE-bench separates the validation feedback used during search from the held-out scores used for final evaluation (Tables 1–2; Section 4.1).
Critical weaknesses
  1. Insufficient investigation of the proposed Bayesian mechanism

W1 — Limited evidence for the acquisition design. Joint Thompson sampling is consistent with the GP posterior, but the paper provides little insight into when this particular allocation rule is preferable. The posterior-mean ablation does not distinguish generic exploration from the value of joint dependence or compare other acquisition rules (Section 4.4).

W2 — Prediction diagnostics do not validate allocation. Section 4.3 evaluates posterior means only on valid expansions selected by BayesianAgent, not posterior uncertainty calibration or decisions among alternative expansions. The virtually identical gain correlations of the full and zero-mean full-kernel GPs also leave the nonzero rating prior’s added value unclear (Table 3; Appendix C.4).

  1. Empirical robustness and attribution remain incomplete

W3 — Execution differences weaken the AlphaEvolve comparisons. Appendix B.2 reports changes to providers/CPU allocations and outdated resource descriptions during AlphaEvolve runs. These can affect throughput and potential ratings, so nominally matched time budgets do not establish controlled comparisons; both the main results and ablations remain difficult to attribute to allocation alone.

W4 — Run-to-run robustness is insufficiently characterized. The favorable MLE-bench results merit credit, but Tables 1–2 provide no per-task repeated-run uncertainty. The three-seed AutoLab extension has mixed task outcomes and only a small overall mean advantage over MLEvolve (0.6558 versus 0.6547), leaving the robustness and magnitude of the gains insufficiently established (Appendix F).

Questions for authors
  1. Regarding W1: With the same GP, candidate proposals, and budget, what evidence distinguishes joint Thompson sampling from independent marginal sampling or another acquisition rule?
  2. Regarding W2: What evidence supports reliable posterior uncertainty and useful cross-candidate information sharing beyond the selected-action mean-prediction analysis? Which modeling components yield a reproducible online benefit?
  3. Regarding W3: Which AlphaEvolve task–method runs were affected by the execution-condition changes? Do the main and ablation conclusions persist in comparisons with matched conditions?
  4. Regarding W4: Are repeated runs available for the main benchmarks, with paired variability in task scores and medal attainment, to distinguish reliable gains from run-to-run variation?
7 comments
AuthorsOpenReview ↗

Response 1/3: Contributions and Acquisition Design (W1)

We thank the reviewer for recognizing the formulation of expansion gain and the held-out MLE-bench evaluation. BayesianAgent contributes a framework for learning expansion outcomes and using them to allocate experiments. Its contributions comprise three connected parts:

  1. An expansion-outcome formulation of parent selection. The method learns the expected gain from a proposed expansion and combines it with measured parent quality to model expected child quality, Qt(p)=s(p)+κf(xt(p))Q_t(p)=s(p)+\kappa f(x_t(p)).
  2. A shared predictive surrogate. The GP relates experimental feedback to parent performance, expansion history, common roots, and LLM potential assessments, allowing outcomes to inform candidates with little direct evidence.
  3. An allocation and feedback loop with empirical validation. Predictions guide the next experiment, and completed outcomes update later predictions. Three-seed benchmark comparisons and offline decision studies evaluate the resulting allocation value.

The acquisition rule converts this learned posterior into an expansion decision. Replacing joint TS with posterior-mean selection, Independent-TS, or GP-UCB preserves the expansion-outcome formulation, shared surrogate, and feedback-learning loop. These alternatives test how to use the model's predictions and retain the central modeling and learning contributions.

1. Why joint sampling uses the modeled relationships

Joint sampling compares candidates under a common plausible realization of their expected child quality. For two candidates,

Var⁡(Q1−Q2)=Var⁡(Q1)+Var⁡(Q2)−2Cov⁡(Q1,Q2).\operatorname{Var}(Q_1-Q_2)=\operatorname{Var}(Q_1)+\operatorname{Var}(Q_2)-2\operatorname{Cov}(Q_1,Q_2).

The covariance determines uncertainty about their relative quality. Joint TS preserves this relationship when choosing an expansion. Independent-TS keeps the marginal means and variances and removes the sampling dependence, providing a direct comparison of how candidate correlations affect allocation.

2. Acquisition comparisons

Table 1. Acquisition rules within the BayesianAgent modeling framework.

Acquisition ruleOnline average rank ↓AUC (%) ↑NPV (%) ↑
BayesianAgent (joint TS)4.5463.3961.19
Posterior mean4.6463.4160.67
Independent-TS5.3062.7759.13
GP-UCB5.5262.9060.67

Online ranks are from the common nine-method comparison across 15 tasks and three seeds. AUC uses 48 proposals and 14 positive-reference tasks; NPV uses fixed menus and 12 positive-reference tasks. Bold marks the best value among the acquisition rules shown.

The acquisition controls use the same proposal construction, GP fitting procedure, and resource configuration. GP-UCB selects by posterior mean child quality plus one latent posterior standard deviation. At fixed menus, the rules share the same fitted GP, recorded history, and available actions. Online and replay histories subsequently evolve according to each rule's choices. Joint TS achieves the best online rank and highest fixed-menu NPV among these acquisition controls. Its higher NPV than Independent-TS supports preserving candidate correlations when comparing expansions; its AUC is comparable to posterior-mean selection.

These experiments position joint TS as an effective use of BayesianAgent's joint posterior and demonstrate the framework's support for alternative acquisition functions. Section 4.5 and Appendices E and J provide the complete acquisition and component comparisons.

AuthorsOpenReview ↗

Response 2/3: Decision Evidence and Component Contributions (W2)

The revised evidence connects three levels of the method: predicting gains from decision-time feedback, choosing among alternative expansions under a common history, and improving solutions over a search budget. We also provide concrete posterior updates and targeted component controls to explain how the surrogate supports these decisions.

1. Direct decision evaluation

The replay simulator starts policies from the same valid recorded drafts and reveals a recorded expansion outcome after its action is chosen. Each expansion attempt counts toward the proposal budget. Fixed-menu evaluation holds the history and available actions constant, using repeated outcomes to score the choices of different policies.

The recorded expansion histories cover 45 source graphs from all 15 AlphaEvolve tasks. The fixed menus contain 122 snapshots, 1,365 state–action pairs, and 2,947 recorded expansion outcomes, including 2,460 valid recorded expansion outcomes for alternative actions. Known failures receive zero positive incumbent improvement. The normalization reference is available to the evaluator, while policy decisions use the recorded information available at each state.

Table 1. Offline allocation comparison.

PolicyAUC (%) ↑NPV (%) ↑
BayesianAgent (joint TS)63.3961.19
Greedy61.4753.22
MCGS (offline adapter)57.8534.27
Random56.9749.45
Agent-rating58.3553.90

AUC measures normalized incumbent improvement over 48 proposals across 14 positive-reference tasks. NPV measures choices at fixed menus across 12 positive-reference tasks. MCGS denotes the offline adaptation of MLEvolve. Section 4.4 presents the main results.

BayesianAgent attains the highest AUC and NPV among these external policies. The NPV result directly supports better expansion choices with the same recorded history and candidate set, while AUC measures how allocation translates into progress over a proposal budget.

2. How feedback informs related candidates

Appendix A.5 gives recorded examples of information sharing through posterior updates. The NOMAD example shows how an observation changes the inferred root contribution, with a larger response near the observed parent's performance. In Leaf Classification, one additional observation changes the predicted normalized gain of an unchanged candidate from another root from 0.6247 to 0.5988, with the preceding observations, candidate input, normalization, and parameters held fixed. These examples illustrate how feedback about one expansion informs other opportunities through the shared posterior.

Section 4.3 complements these examples with gain-prediction results: BayesianAgent achieves Spearman 0.2800, compared with 0.2040 for LLM rating and −0.0283 for MCGS reward, with an advantage over MCGS reward on 14 of 15 tasks. The learned model therefore provides an informative ranking signal for the allocation decisions tested above.

3. Component contributions and overall performance

Table 2. Targeted GP component controls.

VariantOnline average rank ↓AUC (%) ↑NPV (%) ↑
BayesianAgent (joint TS)4.5463.3961.19
w/o root covariance4.6660.6161.02
Zero-mean, full kernel5.3063.9557.70
w/o LLM-potential terms4.8463.2357.85

Every component variant retains joint TS. Online ranks use the common nine-method comparison; the offline protocols are those of Table 1 above. Bold marks the best value among the variants shown. The complete controls, including Context-only and Root-only, are reported in Appendix E, Tables 11–13 and Appendix J, Table 21.

The controls identify different contributions to allocation. Removing root covariance primarily reduces replay progress. Removing the LLM prior mean, or both its mean and covariance contributions, reduces fixed-menu NPV. The zero-mean full-kernel variant retains rating covariance, while w/o LLM-potential terms removes both rating contributions and retains solution reviews.

Across the complete controls, Root-only reaches the highest replay AUC of 64.67% and the lowest NPV of 51.50%, illustrating the distinct information captured by the two evaluation criteria. Taken together, the full model combines the highest fixed-menu NPV with competitive replay AUC, supporting the combined use of root covariance, LLM potential, and joint sampling for both decision quality and search progress. It also has the best mean rank in the three-seed online ablation comparison.

The prediction analysis, recorded posterior updates, and alternative-action evaluations provide complementary evidence for the learning and allocation mechanisms. An independent posterior-coverage study is a future direction for assessing probabilistic calibration.

AuthorsOpenReview ↗

Response 3/3: Execution Settings and Repeated-Run Robustness

W3: Execution comparability and allocation evidence

The revised main comparisons and AlphaEvolve ablations use three seeds under the protocol in Appendix B. Both main benchmarks use GPT-6-luna through Codex with medium reasoning effort. AlphaEvolve runs receive a four-hour active-time budget, 21 CPUs across three search slots, at most 500 evolution steps, and a 1,000-second candidate execution cap. MLE-bench Lite runs receive a 12-hour allocation, one NVIDIA A100, 16 CPUs, and three search workers. These benchmark-specific resources and model access are shared across policies. AlphaEvolve CPU nodes are assigned individually on the same cluster.

Appendix F separately reports the single-seed AlphaEvolve results and describes their provider, CPU-allocation, and potential-assessment resource-description differences. The revised main results and component comparisons use the three-seed experiments.

On AlphaEvolve, the main MLEvolve baseline retains its native scheduling and memory prompts, while Greedy and Agent-rating use BayesianAgent's. Appendix E, Table 10 additionally evaluates MCGS/Top-k with BayesianAgent's scheduling and memory-prompt settings. In that three-policy comparison, BayesianAgent obtains average rank 1.68, ahead of MLEvolve at 1.94 and MCGS/Top-k at 2.38. This result strengthens the allocation comparison by evaluating both the native baseline and its parent- ion rule adapted to the BayesianAgent execution framework.

W4: Repeated-run results and the magnitude of the gains

Table 1. Three-seed main benchmark summary.

PolicyAlphaEvolve average rank ↓MLE-bench Lite gold rate (%) ↑MLE-bench Lite medal rate (%) ↑
BayesianAgent1.9643.94 (6.94)69.70 (2.62)
MLEvolve2.3339.39 (2.62)57.58 (2.62)
Greedy2.6640.91 (4.55)62.12 (5.25)
Agent-rating3.0640.91 (0.00)66.67 (2.62)

AlphaEvolve ranks average the four-policy comparisons over 15 tasks and three seeds. MLE-bench Lite entries are mean (sample SD) across three seeds, each covering all 22 competitions. Medals are determined from official held-out scores.

BayesianAgent leads the AlphaEvolve average rank in all three seeds, with ranks of 1.73, 1.87, and 2.27. It records 22 wins, 11 ties, and 12 losses against MLEvolve across the 45 paired task–seed comparisons. The comparison with Greedy and Agent-rating supports using predicted expansion gains together with current solution quality to allocate further experiments. Appendix E, Table 9 reports each task's signed paired score difference, sample SD, and W/T/L counts.

On MLE-bench Lite, BayesianAgent obtains 46 medals versus 38 for MLEvolve and 29 golds versus 26 across 66 task–seed runs. The gains of 12.12 percentage points in mean medal rate and 4.55 points in mean gold rate demonstrate benefits at both quality thresholds. The three-seed AutoLab experiments in Appendix H provide supplementary validation in an additional application setting.

The revised manuscript's Tables 1 and 2 contain the complete task-level means and sample SDs. Together, the repeated main benchmarks, paired task-level results, and supplementary AutoLab evaluation support the empirical value of learning expansion outcomes for experimental allocation.

Reviewer_hEXeOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper characterizes parent selection in agentic evolution as an expansion-gain prediction problem. It uses a Gaussian process that integrates parent performance, expansion history, common ancestors, and LLM potential scores to predict the gain and uncertainty of the next expansion. The method adds the predicted gain back to the parent score and selects the next expansion via joint Thompson sampling. Across 15 AlphaEvolve tasks and 22 MLE-bench Lite tasks, the paper demonstrates better average ranking and medal performance, and studies the mechanism through prediction analysis and component ablations.

Critical strengths
  1. The contextual kernel supports information sharing across branches, the common-root kernel captures correlation among candidates from the same origin, and the LLM score enters the model through an updatable shared coefficient, avoiding complete reliance on a fixed scoring rule.
  2. The method preserves posterior correlation among candidates and explicitly es uncertainty in the latent expected gain from observation noise, making the technical presentation fairly rigorous.
  3. Mathematical optimization and ML engineering jointly test the method, with MLE-bench using an independent test scorer to determine medals; meanwhile, the ablations provide some support for the common-root information and posterior sampling.
Critical weaknesses
  1. Appendix B.2 acknowledges that some AlphaEvolve runs changed the model provider or CPU configuration, and that BayesianAgent's potential evaluation used an earlier resource description. These differences can alter candidate generation quality, throughput, and scoring, so matching only the nominal time budget is insufficient to attribute the gains to the selection strategy.
  2. The GP only learns gains under valid expansions; failures affect only the attempt-count feature, with no explicit modeling of success probability, execution cost, or long-term gain. The method may therefore prefer expansions that perform well when they succeed but have high failure rates or long runtimes, and may not maximize the quality of the best solution at budget exhaustion.
  3. The prediction analysis covers only the valid expansions chosen by BayesianAgent itself, and does not verify predictive performance or uncertainty calibration on unchosen candidates. The existing ablation compares joint sampling against posterior-mean selection, which cannot alone demonstrate the necessity of preserving cross-candidate correlation; the additional benefit of the LLM potential term also remains unclear.
Questions for authors
  1. Can multiple independent runs be performed on the two main benchmarks, reporting paired performance differences, medal rates, and confidence intervals for average ranking, to determine whether the improvements exceed run-to-run randomness?
  2. What are the differences among methods in invalid-expansion rate and per-expansion runtime? Would adding success probability or per-unit-cost improvement fix the current objective that targets only the expected quality of valid expansions?
  3. Can the authors add controls such as independent Thompson sampling, or keeping the full kernel while removing the LLM potential term, and provide prediction-interval coverage or other calibration metrics, to separately validate the value of correlation modeling and uncertainty estimation?
5 comments
AuthorsOpenReview ↗

Response 1/3: Matched Settings and Repeated Runs (W1)

We thank the reviewer for recognizing the contextual and common-root covariance design, the updatable LLM assessment term, and the distinction between latent uncertainty and observation noise. The revised evaluation includes complete three-seed comparisons on both main benchmarks, an additional MCGS/Top-k allocation comparison, and paired task-level reporting.

1. Experimental comparability

All four main policies use GPT-6-luna through Codex with medium reasoning effort on every task with seeds 42–44. AlphaEvolve runs have a four-hour active-time budget, 21 CPUs across three search slots, at most 500 evolution steps, and a 1,000-second execution cap per candidate. MLE-bench Lite runs have a 12-hour allocation, one NVIDIA A100, 16 CPUs, and three search workers. Appendix B specifies the shared model access and benchmark-specific resources and protocols.

On AlphaEvolve, MLEvolve retains its native scheduling and memory prompts; Greedy and Agent-rating use those of BayesianAgent. Appendix E, Table 10 adds MCGS/Top-k under BayesianAgent's scheduling and memory-prompt settings. BayesianAgent leads this three-policy comparison with average rank 1.68, followed by MLEvolve at 1.94 and MCGS/Top-k at 2.38. This comparison strengthens the evidence for the allocation rule alongside the evaluation of the native MLEvolve baseline.

The single-seed AlphaEvolve results and their execution conditions are reported separately in Appendix F. The revised main comparisons and ablations use the three-seed protocol described in Appendix B.

2. Repeated-run performance

Table 1. Three-seed main benchmark summary.

PolicyAlphaEvolve average rank ↓MLE-bench Lite gold rate (%) ↑MLE-bench Lite medal rate (%) ↑
BayesianAgent1.9643.94 (6.94)69.70 (2.62)
MLEvolve2.3339.39 (2.62)57.58 (2.62)
Greedy2.6640.91 (4.55)62.12 (5.25)
Agent-rating3.0640.91 (0.00)66.67 (2.62)

AlphaEvolve ranks average the four-policy comparisons over all 45 task–seed pairs. MLE-bench Lite entries are mean (sample SD) across three seeds, with all 22 competitions included in every seed. Official held-out grading determines medals after submissions are frozen.

BayesianAgent leads the AlphaEvolve average rank in each seed and obtains 22 wins, 11 ties, and 12 losses against MLEvolve across paired task–seed comparisons. The advantage over Greedy and Agent-rating supports allocating experiments using learned expansion gains together with current parent quality.

On MLE-bench Lite, BayesianAgent earns 46 medals versus 38 and 29 golds versus 26 for MLEvolve across 66 task–seed runs. The resulting gains of 12.12 percentage points in medal rate and 4.55 points in gold rate show benefits in both the proportion of runs attaining a medal and the proportion reaching gold. The AutoLab experiments in Appendix H provide supplementary three-seed validation in an additional setting.

The revised manuscript's Tables 1 and 2 report complete task-level means and sample SDs. Appendix E, Table 9 provides paired AlphaEvolve score differences, sample SDs, and W/T/L counts in the original task units.

AuthorsOpenReview ↗

Response 2/3: Validity and Execution Costs (W2)

The revised experiments evaluate how BayesianAgent's learned expansion model translates into final quality and execution behavior under limited budgets. The acquisition targets expected child quality conditional on a valid outcome. Valid signed gains update the GP, and completed failures enter the attempt-history feature. The benchmark comparisons measure the final solutions reached while these execution outcomes consume experimental resources.

1. Failure and runtime measurements

Table 1. AlphaEvolve failures and candidate execution-time percentiles.

PolicyFailure rate (%)P50 (s)P90 (s)P95 (s)
BayesianAgent9.0332.7997.51,001.5
MLEvolve13.7547.91,012.51,032.1
Greedy11.0783.31,022.11,041.3
Agent-rating11.1749.41,023.01,041.9

The audit gives each task–seed run equal weight and covers timed ordinary expansions and debugging executions completed within the four-hour budget. Failure means a logged invalid execution. Runtime includes candidate execution and scoring; P50, P90, and P95 denote its percentiles.

BayesianAgent has the lowest observed failure rate and shortest median runtime among the four main policies. Appendix E.3 provides the execution-time distributions and complete audits for the main and ablation comparisons in Tables 14 and 15.

2. Allocation value within the experimental budget

Under the four-hour active-time budget, BayesianAgent achieves the best AlphaEvolve average rank. Under the 12-hour MLE-bench Lite allocation, it improves mean medal and gold rates over MLEvolve by 12.12 and 4.55 percentage points. These repeated-run results evaluate the implemented policy with execution costs incurred during search.

Section 4.4 and Appendix J additionally evaluate decisions in a replay simulator and at fixed menus. Each expansion attempt counts toward the proposal budget, and known failures receive zero incumbent-improvement reward in fixed menus. BayesianAgent attains the highest AUC and NPV among the external allocation policies, connecting its learned quality estimates to useful expansion choices.

These experiments support the practical allocation value of the current model. Incorporating failure probability, execution cost, and the remaining experimental budget into the surrogate and acquisition rule is a natural extension for modeling the resource allocation objective more directly.

AuthorsOpenReview ↗

Response 3/3: Acquisition, Information Sharing, and GP Controls (W3)

BayesianAgent's formulation and shared surrogate provide a learned posterior over expansion outcomes that can support several acquisition rules. We have added acquisition controls, targeted GP component controls, repeated evaluations of alternative actions, and recorded posterior-update examples. Together, these studies examine how the model learns from feedback and how its predictions guide allocation.

1. Acquisition rules using the shared modeling framework

Posterior-mean selection, Independent-TS, and GP-UCB retain the expansion-gain formulation, GP representation, and feedback updates. The acquisition controls use the same proposal construction, GP fitting procedure, and resource configuration. Independent-TS preserves posterior means and marginal variances while removing cross-candidate dependence in the draw. GP-UCB selects by posterior mean child quality plus one latent posterior standard deviation. At fixed menus, all acquisition rules use the same fitted GP and decision history; during online search and in the replay simulator, each rule subsequently learns from its own revealed outcomes.

Table 1. Acquisition controls.

Acquisition ruleOnline average rank ↓AUC (%) ↑NPV (%) ↑
BayesianAgent (joint TS)4.5463.3961.19
Posterior mean4.6463.4160.67
Independent-TS5.3062.7759.13
GP-UCB5.5262.9060.67

Online ranks come from the nine-method comparison over 15 tasks and three seeds. Offline AUC uses 48 proposals and 14 positive-reference tasks; NPV uses fixed menus and 12 positive-reference tasks. Bold marks the best value among the acquisition rules shown.

Joint TS has the best online rank and highest fixed-menu NPV among these rules. Relative to Independent-TS, its higher NPV under the same fitted GP supports using posterior correlations to compare expansion opportunities. Joint TS and posterior-mean selection have nearly identical replay AUC, while joint TS attains higher NPV. These comparisons show how alternative acquisition rules use the same learned expansion model; the formulation, information-sharing surrogate, and feedback-learning contributions are retained across the rules.

2. Decisions among alternative actions

The offline evaluation uses recorded expansion histories spanning 45 source graphs from all 15 AlphaEvolve tasks. The replay simulator reveals an outcome after its action is chosen, and each expansion attempt counts toward the proposal budget. Fixed-menu evaluation holds the recorded history and candidate set constant and uses repeated outcomes to evaluate alternative choices.

The fixed menus contain 122 decision snapshots and 1,365 state–action pairs, including 2,460 valid recorded expansion outcomes for alternative actions. Valid negative gains remain regression targets, and known failures contribute zero positive incumbent improvement. The resulting NPV measures how well each policy chooses among the same expansion opportunities. The full method attains 61.19% NPV, compared with 53.22% for Greedy. Section 4.4 presents the main decision results, and Appendix J provides the protocols and complete comparisons.

3. Contributions of root covariance and LLM potential

The component controls distinguish the information used by the surrogate. Zero-mean, full kernel removes the rating prior mean while retaining its covariance. w/o LLM-potential terms removes both rating mean and covariance while preserving the context and root kernels; solution reviews remain enabled.

The full method attains 61.19% NPV, compared with 57.70% for Zero-mean, full kernel and 57.85% for w/o LLM-potential terms. Removing root covariance reduces replay AUC from 63.39% to 60.61%. Thus, the rating terms primarily improve choices at fixed menus, while root covariance helps progress along the replay search. The full model combines the highest fixed-menu NPV with competitive replay AUC. Complete online component results are in Appendix E, Tables 11–13, and offline controls are in Appendix J, Table 21.

Reviewer_yWrEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BayesianAgent uses a Gaussian process to predict gains from candidate solutions. Its covariance combines expansion context, shared ancestry, and LLM assessments. Thompson sampling selects parents using predicted child quality. Evaluation covers AlphaEvolve, MLE-bench Lite, and AutoLab tasks, alongside prediction analyses and ablations.

Critical strengths
  1. Modeling expansion potential separately from current solution quality addresses an allocation problem. The GP construction provides a way to share information across related candidates.
  2. The evaluation covers diverse tasks and includes held-out MLE-bench scoring, online ablations, and an AutoLab comparison. The prediction analysis respects decision time information and acknowledges that selected expansions do not represent all available alternatives.
Critical weaknesses
  1. The strongest performance and ablation claims remain confounded. Appendix B.2 reports provider or CPU allocation changes in some AlphaEvolve runs, while Appendix C describes replacement runs used in the main score tables. These conditions can change throughput and solution quality independently of allocation. The other evidence is encouraging but modest, as the MLE-bench advantage over MLEvolve is one task, and the AutoLab mean is 0.6558 versus 0.6547.
  2. The acquisition target is incompletely aligned with the resource allocation objective. It models expected child quality conditional on a valid outcome, while invalid and slow attempts also consume the budget. Appendix A.4 acknowledges that failure probability and execution cost are not modeled. The paper does not show that these differences are negligible or that conditional quality rankings identify the most useful experiments under a time limit.
Questions for authors
  1. Which simulations share identical provider, hardware, and execution settings? Do the advantages hold on that subset and under repeated runs?
  2. What are the failure rates and execution time distributions across allocation policies? Are the conclusions intact when selection accounts for validity probability and time consumption?
6 comments
AuthorsOpenReview ↗

Response 1/2: Experimental Settings and Robustness (W1)

We thank the reviewer for recognizing the distinction between current solution quality and expansion potential, and the value of sharing evidence across related candidates. We have expanded both complete main benchmarks to three seeds and added a comparison that places MCGS/Top-k under BayesianAgent's scheduling and memory-prompt settings. These experiments strengthen the evidence for learning expansion value to guide experimental allocation.

1. Experimental settings and allocation comparisons

All four main policies use GPT-6-luna through Codex with medium reasoning effort. Within each benchmark, they share model access, compute allocations, and experimental budgets:

  • AlphaEvolve: a four-hour active-time budget, 21 CPUs across three search slots, at most 500 evolution steps, and a 1,000-second candidate execution cap.
  • MLE-bench Lite: a 12-hour allocation, one NVIDIA A100, 16 CPUs, and three search workers.

Each benchmark uses seeds 42, 43, and 44 on every task. AlphaEvolve CPU nodes are assigned individually on the same cluster. MLEvolve retains its native scheduling and memory prompts, while Greedy and Agent-rating use those of BayesianAgent. Appendix B documents these implementations and the evaluation protocol; Appendix F separately reports the single-seed AlphaEvolve results and their execution conditions.

The additional MCGS/Top-k comparison directly examines the allocation rule under BayesianAgent's scheduling and memory-prompt settings. In Appendix E, Table 10, BayesianAgent, MLEvolve, and MCGS/Top-k obtain average ranks of 1.68, 1.94, and 2.38, respectively. These ranks are computed within that three-policy comparison over all 15 tasks and three seeds. BayesianAgent leads both the native MLEvolve implementation and the adapted MCGS/Top-k policy, strengthening the evidence for the learned allocation approach.

2. Three-seed results and their implications

Table 1. Three-seed main benchmark summary.

PolicyAlphaEvolve average rank ↓MLE-bench Lite gold rate (%) ↑MLE-bench Lite medal rate (%) ↑
BayesianAgent1.9643.94 (6.94)69.70 (2.62)
MLEvolve2.3339.39 (2.62)57.58 (2.62)
Greedy2.6640.91 (4.55)62.12 (5.25)
Agent-rating3.0640.91 (0.00)66.67 (2.62)

AlphaEvolve ranks are computed among the four policies within each task and seed, then averaged over 45 task–seed pairs. MLE-bench Lite entries are mean (sample SD) of the three seed-level rates, each based on all 22 competitions. Medals use official held-out scores, with submissions frozen before private grading.

AlphaEvolve supports allocating experiments using learned expansion gains alongside parent quality. BayesianAgent has the best average rank in each seed and records 22 wins, 11 ties, and 12 losses against MLEvolve across the 45 paired comparisons. Its advantage over Greedy and Agent-rating shows the value of learning from experimental feedback beyond selection by current scores or LLM ratings alone.

MLE-bench Lite shows gains at both quality thresholds. BayesianAgent earns 46 medals versus 38 for MLEvolve and 29 golds versus 26 across 66 task–seed runs. These correspond to mean medal- and gold-rate gains of 12.12 and 4.55 percentage points: more runs earn a medal, and more reach gold. AutoLab provides supplementary three-seed validation in an additional setting, as reported in Appendix H.

The revised manuscript's Tables 1 and 2 provide all task-level means and sample SDs. Appendix E, Table 9 reports paired AlphaEvolve score differences and W/T/L counts in each task's original units.

AuthorsOpenReview ↗

Response 2/2: Failures, Runtime, and Decision Quality (W2)

BayesianAgent learns expected child quality conditional on a valid expansion and uses this model to allocate the next experiment. We evaluate its practical value through final solution quality under time budgets, execution measurements, and choices among alternative expansions. These studies connect the acquisition target to the outcomes of budgeted search.

1. Execution behavior

Table 1. AlphaEvolve failures and candidate execution time.

PolicyFailure rate (%)P50 (s)P90 (s)P95 (s)
BayesianAgent9.0332.7997.51,001.5
MLEvolve13.7547.91,012.51,032.1
Greedy11.0783.31,022.11,041.3
Agent-rating11.1749.41,023.01,041.9

The audit covers timed ordinary expansions and debugging executions completed within the four-hour search budget, with equal weight per task–seed run. Failure denotes a logged invalid execution, and execution time includes candidate execution and scoring. P50, P90, and P95 are execution-time percentiles.

Among the four main policies, BayesianAgent has the lowest observed failure rate and shortest median execution time. These measurements show favorable execution behavior alongside its final-quality gains. Appendix E.3, Table 14 and the accompanying figures provide the audit and execution-time distributions; Table 15 reports the component and acquisition variants.

2. Final quality within time budgets

The three-seed benchmark comparisons evaluate the best solutions reached while search incurs execution time and failed attempts. BayesianAgent achieves the best AlphaEvolve average rank under the four-hour active-time budget. On MLE-bench Lite, it improves mean medal and gold rates over MLEvolve by 12.12 and 4.55 percentage points under the same 12-hour allocation. Thus, the learned allocation policy improves observed final quality within the evaluated experimental budgets.

3. Decisions among alternative expansions

The replay simulator reveals a recorded expansion outcome after the corresponding action is chosen. Each expansion attempt counts toward the proposal budget. Fixed menus compare choices under the same recorded history and candidate set, using repeated action outcomes; known failures contribute zero incumbent improvement.

Table 2. Offline allocation comparison.

PolicyAUC (%) ↑NPV (%) ↑
BayesianAgent (joint TS)63.3961.19
Greedy61.4753.22
MCGS (offline adapter)57.8534.27
Random56.9749.45
Agent-rating58.3553.90

AUC summarizes normalized incumbent improvement over 48 proposals across 14 tasks with positive reference improvement. NPV evaluates fixed-menu choices across 12 tasks with positive reference value. MCGS is the offline adaptation of MLEvolve. Full metric definitions and the recorded expansion histories are provided in Appendix J.

BayesianAgent leads both metrics among these policies. The AUC advantage reflects stronger progress over a common proposal budget; the NPV advantage reflects choices with higher normalized expected improvement under common decision conditions. Together with the time-budgeted benchmark results, these studies support the practical value of learning expansion outcomes for allocation. Explicitly incorporating failure probability, execution cost, and the remaining budget into the model and acquisition rule is a natural extension of this framework.

The experiments based on three seeds and the scheduling MCGS comparison resolve my experimental confounding concern. The runtime audit and related results with a time budget also improve the empirical evidence. My remaining concern is that the offline evaluation uses proposal budgets, while the acquisition still omits validity probability and execution cost. Can you provide a comparison with a validity-based and cost-based selection rule, and clarify how unfinished attempts and generation or orchestration affect the budget?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers and Area Chair,

Thank you for your time and constructive feedback. We appreciate your recognition of the motivation for learning expansion potential separately from current solution quality. BayesianAgent formulates parent selection as a Bayesian allocation problem: a structured GP shares experimental feedback across related candidates, and the resulting joint posterior guides subsequent expansions.

Your comments motivated substantial additional experiments and analyses during the rebuttal and discussion. We summarize the main additions below.

  1. Repeated-run evaluation and execution controls — Reviewers [yWrE, hEXe, k7qB]. We expanded the evaluations on 15 AlphaEvolve tasks and 22 MLE-bench Lite competitions to three seeds, reporting task-level variability and paired comparisons. On AlphaEvolve, BayesianAgent achieves the best average rank of 1.96 versus 2.33 for MLEvolve, and leads in every seed. We additionally evaluated MCGS/Top-k under BayesianAgent’s scheduling and memory-prompt settings to better distinguish allocation effects from execution-framework differences.

  2. Direct decision evaluation and component controls — Reviewers [hEXe, k7qB]. We added offline replay and frozen decision menus with repeated alternative-action outcomes, moving beyond prediction diagnostics on selected expansions. BayesianAgent achieves a fixed-menu normalized policy value of 61.19%, compared with 53.22% for parent-score Greedy. We also added Independent-TS, posterior-mean selection, GP-UCB, and targeted root-covariance and LLM-potential ablations to distinguish the contributions of learning, information sharing, and acquisition design.

  3. Validity, execution costs, and budget accounting — Reviewers [yWrE, hEXe]. We audited expansion failures and execution-time distributions, then evaluated validity-aware, deadline-aware, and gain-rate variants in paired offline comparisons. We clarified how replay charges generation, execution, scoring, and repairs, and how unfinished attempts consume budget without revealing their outcomes. These experiments demonstrate how validity and cost can enter the same allocation framework.

  4. Uncertainty diagnostics and mechanism analysis — Reviewers [k7qB, hEXe]. We added paired uncertainty estimates, predictive-interval coverage, and decision-level analyses of ranking ambiguity. Joint TS departs from the mean-optimal choice in 56.17% of ambiguous states but only 8.50% of clear-ranking states; preserving correlations further reduces exploration loss relative to independent sampling in clear-ranking states. A repeated-execution Kissing number case illustrates how sampling preserves a valuable nearly tied alternative while also exposing the risk of failure.

The revised manuscript already includes the expanded benchmark evaluations, decision studies, and component controls. We will incorporate the latest resource-aware comparisons, uncertainty analyses, and case study, together with their assumptions and limitations. Thank you again for helping us strengthen both the empirical evidence and the explanation of BayesianAgent’s allocation mechanism.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.