acceptodds
Under review as a conference paper at ICLR 2027

SeaEvo: Strategy as State for LLM-Driven Evolution

Abstract

Large language model (LLM)-driven evolutionary search is increasingly used for automated algorithm discovery and program optimization. However, most existing methods represent search progress primarily through executable programs and scalar fitness, while language-level strategies remain transient mutation context or unstructured memory. This makes it difficult to organize search progress across related algorithmic directions or reason about which directions are promising, saturated, or underexplored. We introduce \model, which treats strategy as persistent population-level evolutionary state. Each candidate is associated with an explicit strategy description, and these candidate-aligned strategies form a semantic space that supports complementary experience retrieval and population-level search guidance. \model operates as a lightweight layer over existing evolutionary backbones without modifying their selection or evaluation logic. Across mathematical and systems optimization benchmarks, \model improves the mean performance of its matched evolutionary backbone in 23 of 24 comparisons across two mutation models. By retrieving compact strategy descriptions rather than full source programs, \model reduces average LLM token usage by 48.8%. These results suggest that persistent strategy representations provide a practical mechanism for improving the effectiveness and cost-efficiency of LLM-guided evolutionary search, pointing toward compound AI systems whose search capabilities benefit from the structured accumulation and reuse of algorithmic strategies.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-driven evolution, algorithm discovery, program optimization, strategy representation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_hfraOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces SeaEvo, a way to make LLM-guided evolutionary search remember the “idea” behind each program, not just the code and its score. It then uses these strategy descriptions to retrieve useful past attempts and guide the search away from directions that seem stuck.

Critical strengths
  • The main idea is simple and useful: instead of passing full programs around, SeaEvo keeps short strategy descriptions that are easier for the LLM to reuse. This also makes the search more interpretable and reduces token cost.

  • The method is modular and can be added to existing evolutionary search systems. The paper also reports some modest improvements in best scores, so the direction has some promise

Critical weaknesses
  • The empirical evidence is weak. The main results use only three runs over 9 general tasks, so the reported gains are not statistically convincing.
  • Many improvements are small, close to the standard deviation, or look like near-ties. This makes it hard to know whether SeaEvo is genuinely better or just benefiting from noise.
  • I also have significant concerns about some of the obtained results. Indeed, the Prism task in Figure 3 lists the Average as 43.6+-29.8, and the best result as 78. This differs substantially from other entries in the table and, as confirmed by authors in the Appendix, is not statistically significant. Therefore, this entry is misleading and should not have been added to the table.
  • The model choice is not well justified. The paper mainly uses two models, Gemini-3-Flash and MiMo-V2-Pro, but it is unclear why these are enough to support general claims.
Questions for authors

Could the authors clarify whether SeaEvo has been compared with newer frontier models or high-thinking configurations?

2 comments

Thank you for recognizing the simplicity, interpretability, and practical value of SeaEvo's strategy-level representation. Below we address your Weaknesses and Questions.


[W1 & W2]: Statistical strength of three-run evaluations, near-ties, and controlled evidence beyond noise

We agree that with three runs per configuration on the main benchmarks, several task-level margins are small relative to run-to-run variance, and that descriptive mean gains must be clearly distinguished from statistically significant improvements. We therefore clarify the statistical scope of our results and highlight complementary controlled analyses that provide additional evidence for SeaEvo's mechanism:

  1. Explicit qualification of aggregate mean comparisons. The original 23/24 summary (and 19/20 after excluding Prism and EPLB, as discussed below) counts configurations with a higher observed sample mean rather than statistically established per-task wins. We will explicitly clarify this distinction in the abstract and main text, and avoid presenting near-ties as decisive improvements.

  2. Matched 100-state retrieval control (Table 12). To examine whether semantic retrieval provides a genuine mutation-level benefit beyond stochastic variation across complete evolutionary runs, Table 12 conducts a controlled comparison across 100 matched search states on Circle Packing, sampled across three independent trajectories. Holding the parent program, SLN navigation guidance, model configuration, reference count (2), and inference call count (1) fixed, SER's semantic retrieval outperforms uniformly random strategy retrieval with 56 wins, 34 losses, and 10 ties (p=0.026p=0.026, sign test). This provides a paired, single-step counterfactual analysis of retrieval effectiveness under matched inference budgets, complementary to the full-run evaluations.

  3. Variance reduction in four-run ablations (Table 3) and association-disruption controls (Table 4). In Table 3, our audit confirmed that seven configurations aggregate n=4n=4 runs and SER+SLN aggregates n=3n=3 runs; the corresponding raw run-level results are provided in our response to Reviewer 7cVZ. Full SeaEvo achieves 2.6297±0.00772.6297 \pm 0.0077 across four runs, with individual run-level best scores of 2.6343, 2.6186, 2.6305, and 2.6353. Its sample standard deviation is approximately an order of magnitude smaller than those of all non-empty ablated variants (0.0724–0.0855), each of which has at least one run whose best-so-far score falls between 2.44 and 2.51. Moreover, Table 4 shows that disrupting program–strategy associations (SER-Permute) or strategy–fitness associations (SLN-Shuffle) on TXN reduces the mean score from 3886.1 to 3340.1 and 3538.5, respectively.

While we carefully qualify the statistical scope of the three-run benchmark averages, the matched retrieval comparison (Table 12), the substantially lower run-to-run variance in the Full configuration (Table 3), and the association-disruption controls (Table 4) provide complementary evidence supporting the mechanism-level contribution of structured strategy reuse beyond descriptive mean comparisons across a limited number of independent runs.

[W3]: High variance and withdrawal of the Prism result

We fully agree with your critique of the Prism result (43.6±29.843.6 \pm 29.8, Best 78): the variance is exceedingly large and, as noted in Appendix D.8, the difference is not statistically significant.

  1. Withdrawal of Prism and EPLB from performance and mechanism claims. Beyond the statistical uncertainty you highlighted, code-level inspection during review, together with independent findings in recent work, CORAL [1], identified unguarded feasibility and coverage paths in the upstream ADRS/SkyDiscover evaluators for Prism and EPLB. We therefore agree that these results should not be used to support our performance claims. In the revision, we will remove both Prism and EPLB from our performance, efficiency, and navigation claims, remove their entries from Table 2 and the corresponding EPLB comparison from Figure 4, and remove both associated case studies from Appendix E.

[1] Ao Qu et al., CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery, arXiv:2604.01658, 2026.

  1. Revised empirical scope. After excluding the four affected comparisons from the original 24 matched comparisons, SeaEvo achieves a higher reported mean in 19 of the remaining 20 comparisons. This remains a descriptive comparison of sample means rather than a claim of statistical significance. We will consistently reflect this narrower empirical scope in the abstract, main results, and conclusion. We will also recompute the aggregate token-saving statistic over the six retained tasks using the existing logs, rather than retaining the original eight-task aggregate without qualification.

[W4 & Q1]: Model selection rationale and evaluation with a stronger frontier model

To clarify our model selection and address whether SeaEvo remains effective with stronger frontier models, we explain the role of our primary mutation models and highlight an additional evaluation with Claude Opus 4.8 already included in Appendix D.3:

  1. Evaluation with a stronger frontier model: Claude Opus 4.8 (Appendix D.3, Table 14). Gemini-3-Flash and MiMo-V2-Pro were selected for the main experiments to provide broad multi-task and multi-backbone coverage within our experimental budget. We additionally evaluate Claude Opus 4.8 as a third mutation model on TXN, using identical initialization, evaluator, and 100-generation search budgets, with three independent runs. The results demonstrate improvements across both evolutionary backbones:
  • ShinkaEvolve: improves from 4085.79±160.174085.79 \pm 160.17 to 4196.65±72.774196.65 \pm 72.77 with SeaEvo (individual run scores: 4065.04, 4255.32, 3937.01 vs. 4255.32, 4219.41, 4115.23).

  • OpenEvolve: improves from 3925.39±151.423925.39 \pm 151.42 to 4202.27±60.654202.27 \pm 60.65 with SeaEvo (individual run scores: 3861.00, 4098.36, 3816.79 vs. 4237.29, 4237.29, 4132.23).

  1. Scope across model tiers and reasoning configurations. With Claude Opus 4.8, SeaEvo improves the mean score across both evolutionary backbones and reduces the observed run-to-run standard deviation by more than half in these three-run comparisons.

We will move the Claude Opus 4.8 results from Appendix D.3 into a more prominent position in the main text, providing additional evidence that persistent strategy-state organization remains effective when paired with a stronger frontier mutation model.


We thank the reviewer again for highlighting these important limitations. We will revise the manuscript to present a more careful account of statistical uncertainty, evaluator reliability, and model coverage, while making the evidence for SeaEvo's strategy-level contribution more explicit.

Reviewer_7cVZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

SeaEvo augments LLM-driven evolutionary search with persistent, candidate-aligned strategy descriptions. Experiments across mathematical optimization, systems optimization, and agent-scaffold design report improved mean performance and reduced token consumption.

Critical strengths
  1. The method addresses a relevant problem: organizing and reusing search experience while reducing expensive source-code context.
  2. Evaluation spans multiple task families, mutation models, and evolutionary backbones, supporting potential practical applicability.
Critical weaknesses
  1. The showcased programs contain correctness issues that require validation of the systems results.
  • In Figure 12 (p. 39, program lines 87–128), the Prism implementation removes models from curr_p in place. When repair is feasible but the proposal fails both acceptance conditions, the original state is not restored. Subsequent iterations can therefore record an incomplete placement as an improvement.
  • In Figure 13 (p. 40, program lines 21–41), selecting the globally largest weight/k values does not guarantee coverage of every positive-load expert. For example, weights [100, 1] with four physical replicas produce replica counts [4, 0]. These examples establish invalid-output paths in the presented programs; their effect on the reported benchmark scores remains to be established.
  1. Several Table 3 statistics appear internally inconsistent under the natural interpretation of the caption. For SA-only, Mean = 2.5043, SD = 0.0855, and Best = 2.6310 across three runs. If these summarize the same three run-level outcomes, even a population SD requires Best − Mean ≤ √2 × SD. Here, 0.1267 exceeds approximately 0.1209; a sample SD gives a stricter bound. Rounding cannot explain the discrepancy. Other rows exhibit related inconsistencies. Raw run-level results and precise metric definitions are needed.
  2. The novelty boundary and mechanism attribution need clarification. EoH already maintains natural-language descriptions, code, and fitness within an evolving population. SeaEvo’s clearer contribution is the semantic organization and its use for retrieval and navigation. A matched strategy archive without clustering, together with a budget-matched history-reflection baseline, would better isolate this contribution. Moreover, on TXN, SLN-only achieves 3881.1 ± 142.9 versus 3886.1 ± 517.9 for the full method, while costing approximately 24% less. This provides limited support for the claimed benefit of coupling all three modules on that task.
  3. Performance and efficiency claims require stronger qualification. The 23/24 result counts higher means rather than statistically established improvements, with only three runs for most configurations. Token savings also confound representation compression with reference count: OpenEvolve uses five context programs, whereas SER retrieves two inspirations. The accounting of additional post-hoc description calls should also be explicit.
Questions for authors
  1. Which evaluator commits were used, and do all reported best programs pass strict model-placement and expert-coverage checks? Please provide corrected evaluations where necessary.
  2. How does SeaEvo compare with an unclustered strategy archive and ordinary history reflection under matched reference counts and API budgets?
  3. Do the reported token and cost totals include initialization and fallback description calls, as well as mutation, navigation, embedding, and failed calls?
3 comments

Thank you for the careful code-level inspection and for identifying issues concerning statistical reporting, evaluator validation, and cost accounting. We audited the archived run-level results, evaluator snapshots, and token logs, and agree that several claims require correction or qualification. Below we address your concerns point by point.


[W1 & Q1]: Program correctness on Prism and EPLB, evaluator provenance, and scope of claims

We appreciate the reviewer for identifying the correctness issues in our representative programs. We inspected the archived candidate implementations and corresponding benchmark evaluator files and confirm both issues:

Confirmation of correctness issues and evaluator provenance. In Prism, models are removed from curr_p in place. When a repair is feasible but fails both acceptance criteria, the original placement is not restored, allowing subsequent iterations to proceed from an incomplete placement. In EPLB, selecting the globally largest weight/k terms does not guarantee at least one replica for every positive-load expert. Our inspection confirms that neither prism/evaluator.py (SHA-256: ee953fbf5f08f2313bdfdf2137ea6bded5ba0eb81b593e1a6c786849cb566d77) nor eplb/evaluator.py (SHA-256: b8a5c7dd53e13c0623d4a5c24dfb79ed6e776a886193db03d8ce6cf07d178c92) enforces post-hoc model-completeness or positive-load expert-coverage checks on these paths.

We also note that recent work, CORAL [1], reports and fixes several issues in the original ADRS/SkyDiscover evaluators, including silently skipped failed placements in PRISM, unpenalized zero-replica experts in EPLB. These findings independently highlight the importance of explicit feasibility and validity checks in these benchmark evaluators. We will clarify the provenance of our evaluator snapshots.

[1] CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery. Ao Qu et al. arxiv'26.

Withdrawal of affected benchmark claims. To maintain the soundness of our conclusions, we will exclude both Prism and EPLB from our performance, efficiency, and navigation claims, and remove both corresponding case studies from the revision. After excluding the four affected comparisons from the original 24 matched comparisons, SeaEvo achieves a higher reported mean in 19 of the remaining 20 comparisons.

Accordingly, we will revise the abstract, main results, and conclusion to ensure that our empirical claims are based on benchmarks not affected by the two identified evaluator issues.

[W2]: Internal consistency of Table 3 statistics, raw run-level results, and metric definitions

Your mathematical observation is correct under the assumption of three runs. Our audit of the archived run directories and summary scripts revealed that the Table 3 caption incorrectly stated three runs for all configurations: seven configurations aggregate four runs, while SER+SLN aggregates three runs.

  1. Corrected run counts and raw results. Using the actual sample sizes and sample standard deviation (ddof=1ddof=1), all eight reported Mean, SD, and Best triples in Table 3 reproduce to the printed four decimal places:
ConfigurationnRun-level best scoresMeanSample SDBestGen@Best
None42.502797, 2.506405, 2.509899, 2.4902342.50230.00862.509999
SA42.461278, 2.446018, 2.630957, 2.4787642.50430.08552.631067
SER42.476697, 2.481509, 2.500501, 2.6309572.52240.07312.631095
SLN42.627565, 2.503870, 2.635590, 2.4866512.56340.07912.635661
SA+SER42.630439, 2.493775, 2.455193, 2.4894902.51720.07742.630459
SA+SLN42.627803, 2.624813, 2.630831, 2.4646202.58700.08162.630893
SER+SLN32.627565, 2.506641, 2.6359832.59010.07242.636088
Full42.634292, 2.618592, 2.630475, 2.6353282.62970.00772.635344
  1. Precise metric definitions and corrected reporting. Avg and Std are computed over the per-run best-so-far fitness scores, with Std denoting the sample standard deviation (ddof=1ddof=1). Best denotes the maximum among these run-level best scores. Gen@Best denotes the generation at which this overall best candidate was discovered, rather than an average convergence generation across runs. For example, the four SA-only run outcomes yield a mean of 2.504254 and a sample SD of 0.085521, matching the reported values.

We will correct the Table 3 caption, report the sample size for each configuration, and include the raw run-level results and metric definitions in the revision. We will also correct the corresponding run-count descriptions in Section 4.1 and Tables 1 and 19, where the Full configuration uses four runs rather than three.

The corrected run-level results also clarify the consistency of the full configuration: all four Full runs achieve scores between 2.6186 and 2.6353, whereas every ablated configuration has at least one run with a best score between 2.44 and 2.51. The Full configuration's sample SD (0.0077) is approximately an order of magnitude smaller than those of the non-empty ablations (0.0724–0.0855), while the no-module baseline has an SD of 0.0086.

[W3 & Q2]: Novelty boundary relative to EoH, controlled archive comparisons, and module attribution on TXN

We agree that EoH already maintains natural-language thoughts alongside code and fitness within an evolving population. SeaEvo's contribution is not the introduction of thought–code pairs themselves, but the persistent organization of candidate-aligned strategies into a semantic space that supports complementary retrieval (SER) and history-grounded navigation (SLN).

  1. Controlled retrieval and alignment evidence under matched budgets (Table 12, Table 3, and Table 4). Table 12 provides a direct single-step comparison between SER's intra-/cross-family retrieval and uniformly random retrieval of two strategies from an unclustered archive across 100 matched Circle Packing states. The parent program, SLN reflection guidance, model configuration, reference count (2), and inference call count (1) are held fixed. Semantic retrieval achieves 56 wins, 34 losses, and 10 ties (p=0.026p=0.026). This isolates the retrieval benefit under matched reference counts and inference budgets.

Complementing this single-step control, Table 3 provides full-run component attribution: Full SeaEvo (2.6297±0.00772.6297 \pm 0.0077) outperforms SA+SLN (2.5870±0.08162.5870 \pm 0.0816), which retains strategy articulation and periodic reflection without SER. Table 4 further shows that disrupting program–strategy associations (SER-Permute) or strategy–fitness associations (SLN-Shuffle) on TXN reduces the mean score from 3886.1 to 3340.1 and 3538.5, respectively.

These experiments provide complementary evidence at the single-mutation and full-evolutionary-run levels. We will clarify the distinction between these controls rather than presenting them as a single end-to-end flat-archive baseline.

  1. Task-dependent module attribution on TXN. We also agree with your observation that on TXN, SLN-only (3881.1±142.93881.1 \pm 142.9) achieves a mean comparable to Full SeaEvo (3886.1±517.93886.1 \pm 517.9), with lower average API cost (0.56050.5605 versus 0.69660.6966). This indicates that the contribution of individual modules depends on the task. In contrast, on Circle Packing (Table 3), Full SeaEvo achieves a higher mean and substantially lower variance than SLN-only (2.6297±0.00772.6297 \pm 0.0077 versus 2.5634±0.07912.5634 \pm 0.0791).

These results suggest that SER and SLN offer complementary but task-dependent benefits: semantic retrieval can improve mutation-level experience selection, while population-level navigation can provide substantial gains on its own in some settings. We will revise our discussion to avoid implying that all three modules are universally necessary on every benchmark.

[W4 & Q3]: Qualification of aggregate performance, token-saving factors, and explicit cost accounting

We agree that the original presentation requires clearer qualification. We distinguish descriptive mean comparisons from statistical significance, separate representation compression from reference-count effects, and clarify our cost-accounting procedure.

  1. Qualification of aggregate performance claims. The original 23/24 figure, and the revised 19/20 figure after excluding Prism and EPLB, summarize the direction of observed sample means rather than statistically established per-task improvements. We will explicitly qualify this distinction in the abstract and main text.

  2. Disentangling representation compression from reference count. We agree that the end-to-end token comparison between OpenEvolve (five full-code context programs) and SeaEvo (two compact strategy inspirations) changes both the per-item representation length and the number of retrieved references. The observed token reduction therefore reflects their combined effect, rather than representation compression alone. We will make this explicit in the revised efficiency discussion. Table 12 provides a complementary matched two-reference comparison for retrieval effectiveness, although it does not isolate token compression.

Since Prism and EPLB will no longer support our efficiency claims, we will recompute the aggregate token-saving statistic over the six retained tasks using the existing logs and update the abstract accordingly.

Regarding fallback behavior, the actual controller embeds the source code directly when a candidate lacks an articulated strategy description (e.g., initialization or the ε-greedy fallback), rather than issuing a separate post-hoc LLM description call. We will correct Algorithm 1 to reflect this implementation. Finally, we will explicitly distinguish token-based LLM inference cost estimates from embedding API charges and unlogged provider overhead, which are not included in the above cost figures.

By correcting the statistical reporting, withdrawing claims dependent on the two affected evaluators, clarifying the incremental contribution of semantic strategy organization, and providing transparent token accounting, we will substantially narrow and strengthen the empirical claims of the revised manuscript.

Reviewer_NMivOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SeaEvo, a framework that maintains natural-language algorithmic strategies as persistent evolutionary state. It introduces three components: Strategy Articulation (SA) for explicit strategy generation, Stratified Experience Retrieval (SER) for retrieving complementary strategies through semantic clustering, and Strategic Landscape Navigation (SLN) for guiding exploration based on historical search progress.

Experiments across mathematical optimization, systems optimization, and agent-scaffold tasks demonstrate performance improvements over existing evolutionary backbones, alongside reduced token consumption.

Critical strengths

1. Practical strategy-level representation.

The paper provides an intuitive approach to organizing evolutionary search history beyond executable programs and fitness scores. Its modular design allows integration with existing evolutionary frameworks.

2. Consistent empirical improvements.

SeaEvo demonstrates encouraging improvements across multiple tasks, backbones, and mutation models. The ablation studies and additional validation provide useful supporting evidence, while the reduction in token consumption offers practical benefits.

Critical weaknesses

1. Limited methodological novelty and comparison with adaptive search methods

The proposed components mainly combine established techniques: structured prompting (SA), embedding-based retrieval (SER), and LLM-based reflection (SLN). The paper does not sufficiently isolate the benefits of strategy-state organization over simpler language-memory approaches. Although AdaEvolve is evaluated in the appendix, comparisons with stronger adaptive search methods remain limited.

2. Insufficient validation of strategy diversity and retrieval effectiveness.

SER partitions strategy embeddings into a fixed number of clusters and retrieves one intra-family and one cross-family inspiration. However, it is unclear whether these clusters capture genuinely distinct algorithmic strategies, especially when candidates may share similar approaches or combine multiple techniques. The paper provides limited quantitative evidence that this retrieval design improves strategy diversity or enables more effective exploration compared with simpler retrieval alternatives.

3. Limited validation of strategy-level exploration guidance

SLN identifies effective, saturated, and underexplored directions based on historical strategies and fitness scores. However, experiments are limited to 100 generations, with guidance updated every 10 generations. It remains unclear whether such limited exploration supports reliable saturation judgments, and the paper does not directly validate these decisions or their effectiveness over longer search horizons.

Questions for authors

Q1. Incremental contribution. Could the authors compare SeaEvo against a simpler baseline maintaining a flat strategy-memory archive with periodic LLM reflection? This would help isolate the benefits of the proposed strategy-space organization beyond existing thought–code representations and language-based reflection.

Q2. Strategy diversity and retrieval. How many genuinely distinct algorithmic strategies are discovered during evolution? Does the fixed intra-family/cross-family retrieval design improve strategy diversity and exploration effectiveness compared with simpler retrieval alternatives?

Q3. Reliability of exploration guidance. How reliable are SLN's judgments of saturated and underexplored strategy directions given the 100-generation search budget? Can the authors provide evidence that these judgments predict subsequent improvement opportunities?

Q4. Stronger baselines and long-horizon evaluation. Could the authors further compare or discuss SeaEvo against recent adaptive evolutionary frameworks such as AdaEvolve, PACEvolve, and EvoX? In particular, what additional benefits does strategy-state navigation provide beyond adaptive search control or explicitly separating high-level hypothesis generation from program implementation? How do these benefits change over longer search horizons?

2 comments

Thank you for recognizing the practical value, empirical consistency, and modularity of SeaEvo. Below we address your Weaknesses and Questions.


[W1 & Q1]: Incremental contribution beyond a flat strategy archive and periodic reflection

To isolate the benefit of organizing strategies into a semantic state space beyond maintaining thought–code pairs or periodic reflection, our submission provides three complementary controls across both single-step mutations and full evolutionary runs:

  1. Controlled single-step retrieval: Semantic vs. Flat Random Archive (Table 12). At 100 matched search states on Circle Packing, we compare SER's semantic retrieval against uniformly random retrieval of two strategy records from the flat archive, holding the parent program, SLN reflection guidance, model configuration, reference count (2), and inference call count (1) strictly identical. Semantic retrieval achieves 56 wins, 34 losses, and 10 ties (p=0.026p = 0.026, sign test). This directly isolates the value of structured semantic selection over flat archive sampling under matched reflection and budget conditions.
  2. Full-run ablation with strategy articulation + reflection (SA+SLN, Table 3). On Circle Packing (n=4n=4 runs; we identified and will correct the Table 3 caption's misstatement of three runs), retaining candidate strategy descriptions and periodic SLN reflection without SER (SA+SLN) yields 2.5870±0.08162.5870 \pm 0.0816, whereas incorporating SER (Full) reaches 2.6297±0.00772.6297 \pm 0.0077, improving both the mean and run-to-run stability.
  3. Association disruption controls (Table 4). On TXN, preserving the exact prompt structure and call budget while permuting program–strategy links (SER-Permute) or strategy–fitness links (SLN-Shuffle) drops the reported mean from 3886.13886.1 to 3340.13340.1 and 3538.53538.5, confirming that search gains depend on accurate semantic and fitness alignment rather than extra prompting context alone.

These controlled results directly establish the incremental contribution of structured strategy-state organization. Specifically, Table 12 demonstrates that semantic selection outperforms random archive sampling under identical reflection and call budgets; Table 3 shows that integrating SER into the evolutionary loop improves the mean performance and yields lower run-to-run variance in the evaluated setting. Table 4 verifies that these gains are fundamentally driven by the accurate alignment of historical strategy–fitness associations rather than prompt length alone. We will highlight this unified evidence for our core mechanism more prominently in the revision.

[W2 & Q2]: Strategy diversity, semantic cluster fidelity, and intra-/cross-family retrieval effectiveness

To address what distinct algorithmic strategies are discovered and whether partitioning them for intra- and cross-family retrieval is more effective than simpler retrieval alternatives, we examine both the semantic fidelity of the discovered clusters and their controlled impact on exploration:

  1. Discovered algorithmic strategies and cluster fidelity (Table 10 & Table 11). Across evolutionary runs, SeaEvo discovers diverse algorithmic paradigms rather than merely exploring surface-level variations of the same approach, specifically spanning constructive packing, physics-based repulsion relaxation, constrained nonlinear programming, and multi-start search on Circle Packing, and greedy scheduling, local search, simulated annealing, GRASP, and proxy-assisted insertion on TXN. To verify whether the embedding clusters accurately capture these paradigms even when candidates combine multiple techniques, Table 11 audits 20 cluster representatives against their executable code: 10 match the articulated strategy exactly and 10 match the primary algorithmic paradigm (with minor omissions in auxiliary details), with 0 paradigm mismatches (complemented by Table 10, where 1,116 program–strategy pairs across 12 runs show strong semantic alignment). Dynamic repartitioning as the archive grows allows clusters to track these primary paradigms without requiring strategies to be strictly mutually exclusive.
  2. Effectiveness of intra-/cross-family retrieval vs. simpler alternatives (Table 12 & Table 4). To directly test whether retrieving one intra-family and one cross-family strategy improves exploration over simpler unclustered retrieval, Table 12 compares our intra-/cross-family selection against uniformly random two-strategy retrieval across 100 matched Circle Packing states under identical parent, reflection, and call budgets, achieving 56 wins, 34 losses, and 10 ties (p=0.026p = 0.026). Furthermore, Table 4 shows that disrupting the semantic mapping in SER (SER-Permute) drops TXN performance from 3886.13886.1 to 3340.13340.1.

These results suggest that the induced clusters serve as useful coarse-grained semantic neighborhoods rather than a rigid taxonomy of mutually exclusive algorithms. While the cluster boundaries are not perfectly separated, the representative-level audit and strategy–implementation alignment analysis indicate that they preserve meaningful information about the primary algorithmic directions. This level of semantic organization is sufficient to support the complementary retrieval of intra-family refinement and cross-family inspiration, as evidenced by the controlled retrieval comparisons in Table 12.

[W3 & Q3]: Reliability of SLN exploration guidance within the 100-generation budget

To clarify how SLN operates within a 100-generation horizon, we distinguish between static global optimality certificates and history-grounded search scheduling, and verify the latter through counterfactual controls:

  1. Operational meaning of SLN labels. In SLN, "saturated" and "underexplored" are operational summaries of the observed search history, indicating directions with recent fitness stagnation versus sparse historical sampling, used to steer the next 10 generation window, rather than static global certificates that an algorithmic family has no remaining potential.
  2. Empirical validation within the evaluated budget (Table 4 & Table 3). Table 4 directly tests whether the fidelity of these historical judgments drives downstream improvement: when SLN-Shuffle randomizes the strategy–fitness pairing fed to SLN (corrupting which directions appear stagnant vs. promising while keeping reflection frequency and prompt format fixed), the TXN mean drops from 3886.13886.1 to 3538.53538.5. Consistently, in Table 3, adding SLN alone improves the Circle Packing mean from 2.50232.5023 (None) to 2.56342.5634.

In short, the accumulated strategy–fitness trajectory provides informative signals about the observed search history, which can be used to construct history-grounded exploration guidance. While our current experiments do not establish that these labels reliably predict which directions will yield future gains, the SLN-Shuffle results demonstrate that preserving the association between strategies and their observed fitness is important for effective search guidance. We will clarify this operational interpretation and avoid implying that SLN provides global saturation guarantees.

[W1 & Q4]: Comparison and integration with stronger adaptive search frameworks (AdaEvolve, PACEvolve, EvoX)

To clarify how SeaEvo relates to recent adaptive evolutionary frameworks, we evaluate both its empirical composability as a plug-in layer on an adaptive backbone and its conceptual distinction in the broader design space:

  1. Empirical plug-in gains on top of AdaEvolve (Table 13). SeaEvo is designed as an attachable strategy-state layer rather than acting as a standalone replacement for adaptive controllers. Appendix Table 13 directly evaluates integrating SeaEvo into AdaEvolve, which already features adaptive search control and stagnation-triggered meta-guidance. Adding SeaEvo improves AdaEvolve's reported mean from 2.62232.6223 to 2.62832.6283 on Circle Packing, from 0.03220.0322 to 0.03290.0329 on Heilbronn, and from 3622.03622.0 to 4238.04238.0 on TXN. This demonstrates that organizing candidate-aligned strategies for complementary retrieval provides additive value on top of adaptive resource allocation and stagnation-triggered reflection.
  2. Conceptual positioning relative to PACEvolve and EvoX. We will expand our related-work discussion to clearly position SeaEvo alongside concurrent adaptive frameworks:
  • AdaEvolve dynamically allocates search resources and triggers meta-guidance under stagnation;
  • PACEvolve structures hierarchical historical context, explicit hypothesis generation, and backtracking;
  • EvoX meta-evolves the evolutionary search operators and process itself;
  • SeaEvo maintains candidate-aligned strategy descriptions in a clustered semantic space to pair local refinement (intra-family) with cross-family inspiration (cross-family) at compact token cost.

Our conceptual analysis and the AdaEvolve integration experiments highlight a clear takeaway: adaptive search control (how resources, operators, or hypotheses are scheduled) and persistent strategy-state organization (how historical algorithmic ideas are indexed and retrieved) are complementary dimensions. Since SeaEvo addresses the latter as a lightweight state layer, it complements rather than duplicates advanced adaptive evolutionary frameworks. Our current evaluation is limited to 100 generations, and we do not yet establish how the benefits of SeaEvo scale over substantially longer search horizons. We will clarify this limitation in the revised manuscript.

Comments on the paper

1 comment

Dear Reviewers, the authors have submitted their response. Kindly review it and update your review where appropriate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.