acceptodds
Under review as a conference paper at ICLR 2027

WikiEvolve: Coupling Foundational and Experiential Knowledge in Self-Evolving Agents

Abstract

Self-evolving agents can improve solutions for various tasks by iteratively proposing them based on past experience (e.g., proposed algorithms and their utilities), evaluating their utilities, and using evolutionary search to select the best solution. However, as these agents might only make minor modifications and do finite exploration, useful solutions can remain undiscovered, limiting the effectiveness of experience-based evolution. To address this limitation, we introduce WikiEvolve, a novel framework that couples a fixed foundational wiki with an accumulating experiential memory. The foundational wiki provides established foundational knowledge from literature to enrich candidate solutions, while the experiential memory records past experience to guide future selection. Across distributed constraint optimization (DCOP) and Bayesian optimization (BO) tasks, our framework usually improves the final solution utility over experience-based evolution. Moreover, our experiments show how the amount of foundational knowledge influences the final solutions. These results demonstrate that combining a structured foundational wiki with an experiential memory enables self-evolving agents to break out of trial-and-error traps and discover solutions beyond what pure experience can reach.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, self-evolving agents, autonomous research

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_HbthOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WikiEvolve, an LLM-driven self-evolving framework for algorithm improvement that adds a fixed foundational wiki to experience-based evolution. The wiki is a graph of hyperlinked Markdown pages: source cards, synthesized concept pages, and an index node. A curator agent powered by Claude Opus 4.6 builds it offline from a task-specific literature corpus. In each iteration, the agent formulates a research agenda and research questions, retrieves relevant wiki knowledge and past experiences, and proposes local modifications to the current best algorithm. A surrogate model then selects promising candidate algorithms for implementation and evaluation. Experiments cover three DCOP tasks and two BO tasks.

Critical strengths
  1. The motivation is reasonable: experiential memory records how well past attempts performed, but carries no information about mechanisms the agent has not yet explored. The paper connects research planning, the structured foundational wiki, experience retrieval, and candidate selection into a single loop. Each candidate algorithm also cites the source cards it builds on, which links foundational knowledge to the experiential memory.

  2. The paper specifies the wiki schema and construction pipeline, the retrieval scores, the robust standardization and noise model of the GP surrogate, the candidate selection rule, and the utility normalization. The evaluation spans two quite different domains and two LLMs. It also provides several ablation studies and concrete examples of algorithms.

Critical weaknesses
  1. WikiEvolve runs T=50T=50 iterations with up to k=5k=5 implemented candidates each. The Single-Iteration baselines use only 50 evaluations. Could the authors report, for each method, the number of evaluations, token usage, and time? Ideally, all methods would also be compared under a matched evaluation budget.

  2. WikiEvolve greedily updates a single current best algorithm, which is closer to hill climbing than to population-based evolution. Could the authors evaluate an existing evolutionary framework (e.g., OpenEvolve) given access to the same foundational wiki through standard retrieval? Could the authors also compare against other recent open-source evolutionary frameworks?

Typos and grammatical errors: "sugggesting" (Sec. 4.2), "This design help to agent" (Sec. 3.2), "algorithms that specifically works" (Sec. 4.2), "forms a pool the top 16 … that having" (App. B.2), and "and and" (Fig. 8 caption).

Questions for authors

Please see the questions under Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Hbth (Part 1/3)

We thank the reviewer for the thorough review, for recognizing our motivation and method design, and for pointing out the typos. In summary, we (i) report the realized tokens and evaluations of all methods under comparable token budgets, and (ii) evaluate OpenEvolve with BM25 retrieval over the same foundational wiki, together with EoH and ReEvo. Additionally, we have revised the PDF version of the paper, with all revisions marked in blue. We apologize that, due to time constraints, we have not yet added the results for the DeepSeek model. We plan to include these results in a subsequent revision.

W1. WikiEvolve runs T = 50 iterations with up to implemented candidates each. The Single-Iteration baselines use only 50 evaluations. Could the authors report, for each method, the number of evaluations, token usage, and time? Ideally, all methods would also be compared under a matched evaluation budget.

W1. Evaluations and token usage

We thank the reviewer for the request. We do not report wall-clock time because the experiments were run in parallel on a shared server, where varying CPU/GPU load, resource contention, and process queuing introduce substantial variability. The table below reports, for each method, the improvement, the total tokens, the implemented candidates, and the successful evaluations of a self-evolution run on the three DCOP tasks (App. J, Table 12). In our experiments, we primarily use total token consumption as the criterion for matching computational budgets across methods. We acknowledge that the budgets are not fully matched in the single-iteration setting. To address this concern, we have conducted additional experiments under more comparable budgets, with the results also presented in the table below.

TaskMethodImprovement (%) ↑Total token (M)Number of Realized EvaluationsNumber of proposals = implementations = evaluator calls
Graph ColoringSingle-Iteration(No wiki) Best of 505.076.154650
Single-Iteration(No wiki) matched budget6.3011.908997
Single-Iteration(Given wiki) Best of 5010.4310.363750
Single-Iteration(Given wiki) matched budget10.4311.914358
Experience Only11.6213.96178187
ReAct + Partial History12.6111.16146211
OpenEvolve0.0010.10187207
WikiEvolve (Ours)18.0111.87204224
Max-DisCSPSingle-Iteration(No wiki) Best of 5011.994.343550
Single-Iteration(No wiki) matched budget11.9914.25111176
Single-Iteration(Given wiki) Best of 5015.058.063750
Single-Iteration(Given wiki) matched budget15.0514.406290
Experience Only22.8312.58211217
ReAct + Partial History12.1614.70143202
OpenEvolve18.6813.68150188
WikiEvolve (Ours)26.7414.39187206
Weighted DCOPSingle-Iteration(No wiki) Best of 503.366.643850
Single-Iteration(No wiki) matched budget3.365.643142
Single-Iteration(Given wiki) Best of 503.209.764050
Single-Iteration(Given wiki) matched budget3.205.612428
Experience Only5.687.24199228
ReAct + Partial History4.505.40146187
OpenEvolve5.065.07160226
WikiEvolve (Ours)6.795.58182201

We have added this table to App. J (Table 12), which additionally reports OpenEvolve (Given wiki), EoH (Given sources), and ReEvo (Given sources). We also report the token efficiency analysis (Section 4.6, Figure 3). These comparisons show that WikiEvolve achieves higher utility improvement under comparable total-token budgets on all three DCOP tasks (App. J, Table 12), including against the Single-Iteration baselines run with comparable token budgets. Its advantage therefore does not simply result from spending more tokens. The trajectories in Sec. 4.6 (Fig. 3) show substantial improvements with relatively few tokens on Graph Coloring and Weighted DCOP, and continued improvement as more tokens are consumed on Max-DisCSP. WikiEvolve also outperforms the baselines given access to foundational knowledge, showing that access alone does not account for its gains. Together, these results support the effectiveness of the structured wiki, research planning, and evaluation feedback in achieving greater utility improvement within comparable token budgets.

AuthorsOpenReview ↗

Response to Reviewer Hbth (Part 2/3)

W2. WikiEvolve greedily updates a single current best algorithm, which is closer to hill climbing than to population-based evolution. Could the authors evaluate an existing evolutionary framework (e.g., OpenEvolve) given access to the same foundational wiki through standard retrieval? Could the authors also compare against other recent open-source evolutionary frameworks?

W2. Population-based evolution and existing evolutionary frameworks with access to foundational knowledge

We thank the reviewer for this question. We agree that WikiEvolve updates a single current best algorithm greedily, which resembles hill climbing with several candidates per step. However, WikiEvolve explicitly balances exploitation and exploration when it generates and selects candidates. The candidates are proposed from the foundational knowledge retrieved from the wiki, so a candidate can introduce a mechanism from a method that is dissimilar to the current algorithm. On Graph Coloring, for example, the lineage of the best-found algorithm cites DPOP and Bounded Max-Sum, which rank 16 and 11 in similarity to the best-found algorithm (App. I, Fig. 10). Candidate selection then balances exploitation of candidates with higher predicted utility and exploration of candidates with uncertain utility under the surrogate model (Sec. 3.3, App. D).

We provide additional results below comparing the performance of OpenEvolve with and without access to our wiki knowledge base, alongside the results of EoH and ReEvo with similar token budget. And for EoH, ReEvo and OpenEvolve with wiki, we use BM25 to provide 16 documents for the agent to use per iteration, which is the same setting as our method.

TaskMethodImprovement (%) ↑Total token (M)
Graph ColoringEoH (Given sources)8.2511.98
ReEvo (Given sources)2.9911.96
OpenEvolve (Given wiki)4.8111.40
OpenEvolve0.0010.10
WikiEvolve (Ours)18.0111.87
Max-DisCSPEoH (Given sources)22.1114.48
ReEvo (Given sources)17.8114.48
OpenEvolve (Given wiki)20.0814.94
OpenEvolve18.6813.68
WikiEvolve (Ours)26.7414.39
Weighted DCOPEoH (Given sources)1.825.63
ReEvo (Given sources)1.925.65
OpenEvolve (Given wiki)5.115.53
OpenEvolve5.065.07
WikiEvolve (Ours)6.795.58

Coupling WikiEvolve with population-based evolution is a promising direction for future work, and we have added it to Sec. 6.

AuthorsOpenReview ↗

Response to Reviewer Hbth (Part 3/3)

Typos. "sugggesting" (Sec. 4.2), "This design help to agent" (Sec. 3.2), "algorithms that specifically works" (Sec. 4.2), "forms a pool the top 16 … that having" (App. B.2), and "and and" (Fig. 8 caption).

Typos

We thank the reviewer and have corrected all of them in the revised paper.

We hope that our clarifications address your concerns and improve your opinion of our work. We would be happy to answer any further questions.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer Hbth,

As the discussion period is coming to an end, we would like to follow up on our response. Please let us know if you have any remaining concerns, and we will be glad to address them before the discussion closes.

If our clarifications and additional experiments have resolved your concerns, we would be grateful if you could consider updating your score accordingly.

Thank you once again for your time and constructive feedback.

Best regards,

The Authors

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

WikiEvolve combines a fixed wiki of domain literature with a growing archive of proposed algorithms, implementations, and evaluation results. Each iteration updates a research agenda, retrieves relevant experience and wiki pages, generates candidate modifications, and uses a Gaussian-process surrogate to select candidates for implementation. The evaluation covers three distributed constraint optimization tasks and two Bayesian optimization tasks with two language models. The paper reports gains over experience-only evolution on all six DCOP task-model combinations and better mean BO results when algorithms tuned at 60 evaluations are deployed for 200 evaluations. Ablations examine the amount and organization of retrieved knowledge, research planning, and surrogate selection.

Critical strengths
  1. Experience Only keeps the proposed search framework and removes the wiki, making it the most informative baseline for the paper's main claim. WikiEvolve improves the mean utility in all six DCOP task-model combinations in Table 1. The BO results also include a meaningful change in deployment budget: the selected algorithms were tuned at 60 evaluations and are then tested at 200. The paper reports both settings, including the mixed results at the tuning budget.

  2. The mechanism is described concretely enough to inspect. Appendix A explains source-card construction and graph links, Appendices B–D specify retrieval and candidate selection, and Appendices F–G trace modifications back to source methods and show code. The result that 20 retrieved cards outperform both no wiki and all 40 cards on Graph Coloring is also useful, even though it currently concerns one task and model. These details make the work more informative than a comparison of final scores alone.

Critical weaknesses
  1. The external validation of the discovered algorithms is limited. The DCOP search evaluates candidates on 18 frozen instances per task, and the reported utility is defined over those instances; I could not find a separate final evaluation on unseen instances or fresh evaluation seeds (Appendix E.1–E.3). The BO tasks have three HPO instances and one DM instance. Increasing the horizon from 60 to 200 is a useful test, and the paper correctly describes it as transfer to a larger budget. It leaves transfer to new problems untested. The DM results at 200 evaluations also have large between-run variation: 2.88 ± 61.75 for Qwen and 21.17 ± 36.52 for DeepSeek. These use a first-order transformation of log-regret variability across three evolution runs; the error terms are not confidence intervals. The results support a promising horizon-transfer observation, but give limited evidence about how reliably the discovered algorithms help beyond the fixed objectives used for search.

  2. The comparisons do not fully isolate the benefit of structure at a comparable resource budget. WikiEvolve uses 50 iterations with up to five implementations per iteration, while the single-iteration baseline evaluates 50 proposals in total. Realized implementation, evaluator, and token counts for the other methods are not reported. The internal Experience Only comparison is stronger, but the structured-versus-flat ablation also needs clarification: the structured system has synthesized concept pages as well as source cards, so removing structure may change the available semantic content and context length. Without holding those factors fixed, Figure 2b supports the complete wiki design more directly than graph connectivity itself. The surrogate ablation is encouraging, but does not supply an end-to-end resource comparison.

Questions for authors
  1. Were the final DCOP algorithms reevaluated on any instances or evaluation seeds that were not used during evolution? If so, please report those results separately. For BO, can the existing runs provide per-run or paired log-regret differences at 200 evaluations? This would establish the scope and consistency of the reported transfer result.

  2. What were the realized numbers of proposals, implementations, evaluator calls, and model tokens for WikiEvolve and its baselines? Do the existing logs permit a comparison against Experience Only at a common evaluation or inference budget? Please distinguish the reported iteration limits from actual resource use.

  3. In the flat-source-card ablation, are the synthesized concept-page text, total retrieved context, and planning stages preserved? Please specify exactly what changes relative to the structured wiki. This determines whether Figure 2b isolates graph structure or evaluates a broader change in accessible information.

6 comments
AuthorsOpenReview ↗

Response to Reviewer jctU (Part 1/5)

We thank the reviewer for the careful and detailed review, and for noting that Experience Only is the most informative baseline for our main claim and that the mechanism is described concretely enough to inspect. In summary, we (i) re-evaluate the best-found DCOP algorithms on unseen instances and report the paired log-regret differences of every BO self-evolution run at 200 BO evaluations, (ii) report the realized tokens, implemented candidates, and evaluations of all methods, together with Single-Iteration runs under a matched token budget, and (iii) specify exactly what Flat Source Cards changes relative to WikiEvolve. Additionally, we have revised the PDF version of the paper, with all revisions marked in blue. We apologize that, due to time constraints, we have not yet added the results for the DeepSeek model. We plan to include these results in a subsequent revision.

W1. The external validation of the discovered algorithms is limited. The DCOP search evaluates candidates on 18 frozen instances per task, and the reported utility is defined over those instances; I could not find a separate final evaluation on unseen instances or fresh evaluation seeds (Appendix E.1–E.3). The BO tasks have three HPO instances and one DM instance. Increasing the horizon from 60 to 200 is a useful test, and the paper correctly describes it as transfer to a larger budget. It leaves transfer to new problems untested. The DM results at 200 evaluations also have large between-run variation: 2.88 ± 61.75 for Qwen and 21.17 ± 36.52 for DeepSeek. These use a first-order transformation of log-regret variability across three evolution runs; the error terms are not confidence intervals. The results support a promising horizon-transfer observation, but give limited evidence about how reliably the discovered algorithms help beyond the fixed objectives used for search.

Q1. Were the final DCOP algorithms reevaluated on any instances or evaluation seeds that were not used during evolution? If so, please report those results separately. For BO, can the existing runs provide per-run or paired log-regret differences at 200 evaluations? This would establish the scope and consistency of the reported transfer result.

W1 + Q1. Validation beyond the instances used during self-evolution

To demonstrate the reliability of the discovered algorithms, we report the DCOP results on held-out instances and the paired log-regret differences for BO below.

DCOP, unseen instances. We re-evaluate the best-found DCOP algorithms of Table 1 on 18 held-out instances per task, which are constructed in the same way as the 18 instances used during self-evolution but are not seen during self-evolution. The evolved algorithms exhibit some performance degradation on the held-out instances. Nevertheless, WikiEvolve consistently outperforms the baselines across all three tasks, demonstrating its ability to generalize effectively to unseen instances. We have added these results to App. E.6 (Table 6).

TaskMethodOriginal casesHeldout cases
Graph ColoringReAct + PartialHistory12.616.85
Graph ColoringExperience Only11.62-0.04
Graph ColoringOpenEvolve0.000.00
Graph ColoringWikiEvolve18.0110.77
Max-DisCSPReAct + PartialHistory12.168.22
Max-DisCSPExperience Only22.8320.32
Max-DisCSPOpenEvolve18.6816.91
Max-DisCSPWikiEvolve26.7424.56
Weighted DCOPReAct + PartialHistory4.503.57
Weighted DCOPExperience Only5.683.96
Weighted DCOPOpenEvolve5.064.78
Weighted DCOPWikiEvolve6.795.56
AuthorsOpenReview ↗

Response to Reviewer jctU (Part 2/5)

W1 + Q1. Validation beyond the instances used during self-evolution (continue)

BO, paired log-regret differences at 200 evaluations. The table below reports the paired log-regret differences between the algorithm returned by WikiEvolve and the one returned by each other method for the same task and self-evolution run with seeds 1101, 1102 and 1103. A negative value means that WikiEvolve has the lower regret. WikiEvolve has the lower mean log regret at 200 evaluations against every method for both LLMs. Each column is one self-evolution run (task and seed). Each value is averaged over evaluation seeds 1, 2 and 3.

LLMWikiEvolve minusHPO 1101HPO 1102HPO 1103DM 1101DM 1102DM 1103Mean
Qwen3.8-27BOpenEvolve−1.38−1.22−0.69−1.70−0.63+0.73−0.82
ReAct + Partial History−1.38−1.22−0.25−0.600.00+0.11−0.56
Single-Iteration (Given wiki)−1.37−1.22−0.88−0.78−0.35−0.15−0.79
Single-Iteration (No wiki)−1.34−1.48−0.15−0.94+0.29+0.55−0.51
Experience Only−0.59−0.78+0.14−1.51−0.31+0.17−0.48
DeepSeek-V4-FlashOpenEvolve−0.59−1.25−0.66−1.29+0.03−0.05−0.63
ReAct + Partial History−0.59−1.25−0.66−0.99+0.03−0.55−0.67
Single-Iteration (Given wiki)−0.20−1.23−0.22−1.26−0.56−0.22−0.61
Single-Iteration (No wiki)−0.31−0.96+0.03−1.80−0.07−0.08−0.53
Experience Only+0.20−0.36−0.59−1.06−0.55−0.34−0.45
AuthorsOpenReview ↗

Response to Reviewer jctU (Part 3/5)

W2. The comparisons do not fully isolate the benefit of structure at a comparable resource budget. WikiEvolve uses 50 iterations with up to five implementations per iteration, while the single-iteration baseline evaluates 50 proposals in total. Realized implementation, evaluator, and token counts for the other methods are not reported. The internal Experience Only comparison is stronger, but the structured-versus-flat ablation also needs clarification: the structured system has synthesized concept pages as well as source cards, so removing structure may change the available semantic content and context length. Without holding those factors fixed, Figure 2b supports the complete wiki design more directly than graph connectivity itself. The surrogate ablation is encouraging, but does not supply an end-to-end resource comparison.

Q2. What were the realized numbers of proposals, implementations, evaluator calls, and model tokens for WikiEvolve and its baselines? Do the existing logs permit a comparison against Experience Only at a common evaluation or inference budget? Please distinguish the reported iteration limits from actual resource use.

W2 + Q2. Resources and the Flat Source Cards ablation

We thank the reviewer for pointing out these ambiguities and we will address them below.

Flat Source Cards ablation. Flat Source Cards places all source cards in an unstructured pool without connections between them. Its source cards are ranked by lexical relevance (Eq. 1) without graph proximity (Eq. 2) during retrieval. In WikiEvolve, the concept pages serve as traversal seeds for graph proximity (App. B.2), while final candidate implementation uses only the source cards. The agent's context for implementation therefore never includes concept pages, in both WikiEvolve and Flat Source Cards. Fig. 2b thus evaluates the graph structure of the foundational wiki, i.e., the concept pages together with their edges to the source cards. We explicitly clarified this setting in the revision.

Reviewer_AfX5OpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WikiEvolve, a self-evolving agent framework that combines foundational knowledge from existing literature with experiential knowledge accumulated through iterative algorithm design and evaluation. The framework leverages structured knowledge retrieval, historical experiences, and surrogate-assisted candidate selection to guide algorithm generation and code refinement. Experiments on algorithm design tasks, including Bayesian optimization (BO) and distributed constraint optimization (DCOP), demonstrate promising performance improvements over several baselines.

Critical strengths

The proposed WikiEvolve framework demonstrates promising performance improvements across multiple benchmark tasks, including Bayesian optimization (BO) and distributed constraint optimization (DCOP). The experimental results provide encouraging evidence of the effectiveness of the proposed approach.

Critical weaknesses
  1. Combining external knowledge retrieval with experiential learning and iterative exploration is not uncommon in existing agentic systems. Although the authors discuss related approaches, the distinctions presented in the Related Work section do not sufficiently establish the technical novelty of WikiEvolve. In particular, it remains unclear what fundamentally new mechanisms are introduced beyond integrating established components.
  2. The experimental evaluation lacks comparisons with several representative approaches in automated algorithm design. Although OpenEvolve is included, the limited baseline coverage makes it difficult to assess the competitiveness of WikiEvolve against recent state-of-the-art methods.
Questions for authors
  1. Could the authors provide additional comparisons with recent state-of-the-art automated algorithm design approaches, such as EoH or ReEvo, where applicable? Such comparisons would help establish the competitiveness and practical advantages of WikiEvolve.
  2. Alternatively, could the authors demonstrate how WikiEvolve can be naturally integrated into existing automated algorithm design frameworks and whether such integration leads to substantial performance improvements? This would provide stronger evidence of the general applicability and effectiveness of the proposed approach, even if its individual components are not entirely novel.
  3. What are the key technical challenges that distinguish the setting considered in this paper from existing knowledge-augmented agents and automated algorithm design approaches? Currently, the proposed framework appears to be a relatively incremental integration of existing techniques. It would be helpful if the authors could more clearly identify the unique technical difficulties arising from the considered setting and explain how WikiEvolve addresses them beyond existing methods.
9 comments
AuthorsOpenReview ↗

Response to Reviewer AfX5 (Part 1/4)

We thank the reviewer for the helpful comments and for acknowledging the promising performance improvements of WikiEvolve on both BO and DCOP. In summary, we (i) clarify the mechanisms that distinguish WikiEvolve from existing methods, (ii) compare WikiEvolve with EoH and ReEvo equipped with retrieval under a matched token budget, and (iii) identify the technical challenges of our setting, supported by an analysis of which sources each method uses (App. I). Additionally, we have revised the PDF version of the paper, with all revisions marked in blue. We apologize that, due to time constraints, we have not yet added the results for the DeepSeek model. We plan to include these results in a subsequent revision.

W1. Combining external knowledge retrieval with experiential learning and iterative exploration is not uncommon in existing agentic systems. Although the authors discuss related approaches, the distinctions presented in the Related Work section do not sufficiently establish the technical novelty of WikiEvolve. In particular, it remains unclear what fundamentally new mechanisms are introduced beyond integrating established components.

W2. The experimental evaluation lacks comparisons with several representative approaches in automated algorithm design. Although OpenEvolve is included, the limited baseline coverage makes it difficult to assess the competitiveness of WikiEvolve against recent state-of-the-art methods.

Q1. Could the authors provide additional comparisons with recent state-of-the-art automated algorithm design approaches, such as EoH or ReEvo, where applicable? Such comparisons would help establish the competitiveness and practical advantages of WikiEvolve.

Q2. Alternatively, could the authors demonstrate how WikiEvolve can be naturally integrated into existing automated algorithm design frameworks and whether such integration leads to substantial performance improvements? This would provide stronger evidence of the general applicability and effectiveness of the proposed approach, even if its individual components are not entirely novel.


W1 + W2 + Q1 + Q2. Novelty and comparison with EoH and ReEvo

Novelty. We agree that retrieval, memory, and iterative exploration are not new on their own. WikiEvolve starts from a different perspective than existing automated algorithm design methods. Those methods evolve algorithms from experience, i.e., from the candidates evaluated so far, whereas WikiEvolve aims to learn from the foundational knowledge of the domain and to couple it with experience. Its self-evolution loop is therefore designed around the foundational wiki, rather than obtained by adding retrieval to an existing framework, and it is independent of those frameworks. This design is realized through three mechanisms (Sec. 3.1–3.3, App. B.2).

  1. Structured foundational wiki. The foundational wiki is a graph of source cards and concept pages joined by hyperlinks. The links encode how methods relate, which lexical similarity alone does not capture.
  2. Enforced exploration of foundational knowledge. The relevance of a source card is a weighted sum of its lexical similarity to the research questions and a graph bonus for source cards close to the top-ranked concept pages (Eq. 3). Recently retrieved source cards are penalized, which encourages exploration of the wiki.
  3. Surrogate model and experience guide the next retrieval. The iteratively learned surrogate model selects candidates for evaluation. Their implementations and utilities update the experiential memory, which shapes the next research agenda and research questions, i.e., the retrieval queries.

We have revised the Related Work (Sec. 5) to state these distinctions explicitly, and the Introduction now presents the three challenges (Ch1–Ch3) that these mechanisms address.

AuthorsOpenReview ↗

Response to Reviewer AfX5 (Part 2/4)

W1 + W2 + Q1 + Q2. Novelty and comparison with EoH and ReEvo (continue)

Comparison with EoH and ReEvo. We therefore report the performance of EoH and ReEvo when provided with access to the same foundational knowledge as WikiEvolve under comparable computational budgets. For each iteration, we use BM25 to provide 16 papers for the agent to use, which is the same setting as our method. The results show that WikiEvolve consistently outperforms both baselines across all three DCOP tasks.

Additionally, we demonstrate the important contribution of structured foundational wiki and surrogate model by removing them in WikiEvolve (Flat source cards) and WikiEvolve (No surrogate) respectively, which is presented in the table as well.

TaskMethodImprovement (%) ↑Total token (M)
Graph ColoringEoH (Given sources)8.2511.98
ReEvo (Given sources)2.9911.96
OpenEvolve (Given wiki)4.8111.40
OpenEvolve0.0010.10
WikiEvolve (Ours)18.0111.87
Max-DisCSPEoH (Given sources)22.1114.48
ReEvo (Given sources)17.8114.48
OpenEvolve (Given wiki)20.0814.94
OpenEvolve18.6813.68
WikiEvolve (Ours)26.7414.39
Weighted DCOPEoH (Given sources)1.825.63
ReEvo (Given sources)1.925.65
OpenEvolve (Given wiki)5.115.53
OpenEvolve5.065.07
WikiEvolve (Ours)6.795.58

Integration into existing frameworks. We can provide existing automated algorithm design frameworks with source cards or raw papers from our curated corpus to improve utility. With additional engineering, these frameworks could also exploit the wiki structure by replacing BM25 with WikiEvolve’s relevance score (Eqs. 1–4, App. B.2). In our experiments, giving OpenEvolve access to its source cards through BM25 retrieval (OpenEvolve (Given wiki); Table 1 and App. J, Table 12) improves the utility on Graph Coloring (from 0.00% to 4.81%) and Max-DisCSP (from 18.68% to 20.08%). The relevance score of WikiEvolve (Eq. 1–4, App. B.2) could also replace BM25 with additional engineering. The complete pipeline, however, is not plug-and-play, since research planning and surrogate-guided candidate selection have no counterpart in existing automated algorithm design frameworks like EoH, ReEvo, or OpenEvolve, which evaluate every program they generate. Accordingly, OpenEvolve (Given wiki) remains below WikiEvolve on all three tasks (18.01%, 26.74%, and 6.79%), which indicates that access to foundational knowledge accounts for only a part of the improvement.

AuthorsOpenReview ↗

Response to Reviewer AfX5 (Part 3/4)

Q3. What are the key technical challenges that distinguish the setting considered in this paper from existing knowledge-augmented agents and automated algorithm design approaches? Currently, the proposed framework appears to be a relatively incremental integration of existing techniques. It would be helpful if the authors could more clearly identify the unique technical difficulties arising from the considered setting and explain how WikiEvolve addresses them beyond existing methods.

Q3. Challenges of self-evolving from a foundational wiki

We have revised the Introduction (Sec. 1, Ch1–Ch3) to clarify the technical challenges of learning from a foundational wiki during self-evolution and explain how WikiEvolve addresses them. Existing automated algorithm design methods such as EoH, ReEvo, and OpenEvolve evolve algorithms from experience, and knowledge-augmented agents provide knowledge through the prompt or through similarity-based retrieval. Our setting asks the agent to learn from a foundational wiki during self-evolution, which raises three technical challenges that these methods do not address.

  1. Which foundational knowledge to provide. More knowledge is not better. On Graph Coloring, the utility improvement is 11.62% without wiki retrieval, 21.93% with 20 retrieved source cards, and 6.76% when the entire wiki is provided (Sec. 4.3, Fig. 2a). Placing the knowledge in the prompt is therefore not sufficient. WikiEvolve retrieves a targeted set of source cards with the research questions derived from its research agenda, and removing the research agenda and research questions reduces the utility on all three DCOP tasks (Sec. 4.5, Fig. 2c).
  2. How to reach useful knowledge beyond the neighbourhood of the current algorithm. Under similarity-based retrieval, the agent keeps using the sources closest to the current algorithm. WikiEvolve organizes the foundational knowledge as a graph of source cards and concept pages, adds a graph bonus to the relevance of a source card, and penalizes recently retrieved source cards (Eq. 4). We compare it with Flat Source Cards, which places all source cards in an unstructured pool without graph structure (Sec. 4.4). The source cards are therefore ranked by lexical relevance (Eq. 1) only, without graph proximity (Eq. 2). The structured wiki achieves a higher utility than Flat Source Cards on all three DCOP tasks, by 3.36 to 7.14 points (Sec. 4.4, Fig. 2b, note that Fig. 2b use 90 iterations for ablations, we use 50 iterations here to here to maintain a computational budget comparable to that of WikiEvolve in the main results.).
  3. Which candidates to evaluate. Only a few candidates can be implemented and evaluated per iteration. WikiEvolve does not leave this choice up to the LLM's reasoning capability nor control whether to exploit or to explore the foundational knowledge. Instead, WikiEvolve uses its learned surrogate model to predict the utility of candidates and use a BO-inspired acquisition function to balance between candidates with high utility and candidates with high uncertainty. Candidates with uncertain utility are likely generated from previously unexplored source cards (which may be boosted by the repetition penalty on explored source cards). Removing the surrogate model (and selection based on contextual relevance alone) leads to a lower final utility on Max-DisCSP and Weighted DCOP and to a less steady improvement on Graph Coloring (App. E.4, Fig. 7).

Evidence from the comparison with EoH and ReEvo (App. I). The comparison above illustrates the second challenge. We randomly sample 20 sources of the corpus and record, for each method, how often each source is used over the first 6M tokens of a self-evolution run (Fig. 11) and the best improvement among the candidates that use it (Fig. 12). EoH and ReEvo retrieve from the same sources, yet they keep returning to the local search variants close to the initial algorithm (ALS / DSA variants and MGM / MGM-2) and use only 1 to 3 of the 20 sources (Table 11). WikiEvolve uses 15 to 17 of them.

TaskReEvoEoHWikiEvolve
Graph Coloring1115
Max-DisCSP3316
Weighted DCOP1217

Several sources that EoH and ReEvo never use are in the lineage of the best-found algorithm of WikiEvolve (stars in Fig. 12). On Graph Coloring, these are Bounded Max-Sum, DPOP, Deep Coordination Graphs, NCBB, ConcFB, and PC-SyncBB. On Max-DisCSP, these are the DCOP survey, Bounded Max-Sum, VL-DCOP, and D-Bay. This wider use of the foundational knowledge comes with a higher improvement. WikiEvolve achieves 18.01%, 26.74%, and 6.79% on the three tasks, against at most 8.25%, 22.11%, and 1.92% for EoH and ReEvo under a comparable token budget.

Reviewer_z4NNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WikiEvolve, a self-evolving agent framework that combines foundational knowledge from the scientific literature with experiential knowledge accumulated through algorithm evaluation. The foundational knowledge is organized as a fixed, graph-structured wiki consisting of source cards, concept pages, and an index. At each evolution iteration, the agent updates a research agenda and research questions, retrieves relevant wiki pages and past experiences, generates incremental algorithm modifications, and uses a Gaussian-process surrogate model to select candidates for implementation and evaluation.

The method is evaluated on three distributed constraint optimization (DCOP) tasks and two Bayesian optimization (BO) tasks using two language models. The results show consistent improvements over experience-only evolution on DCOP tasks and better performance under a larger deployment evaluation budget on BO tasks. Additional ablations examine wiki structure, retrieval quantity, and research planning.

Critical strengths
  1. The paper addresses an important limitation of experience-only evolutionary agents by incorporating established domain knowledge into iterative algorithm improvement. The structured wiki, research planning, and experiential memory form a coherent framework for guiding algorithm discovery beyond previously evaluated candidates. This approach is practically meaningful for knowledge-intensive optimization tasks.

  2. WikiEvolve consistently outperforms the experience-only baseline across the three DCOP tasks and shows promising generalization to larger evaluation budgets on BO tasks. The comparisons between raw documents, flat source cards, and structured wiki retrieval provide evidence for the value of knowledge organization. The ablations on research planning and retrieval quantity further offer useful insights into how foundational knowledge affects evolutionary search.

Critical weaknesses
  1. The overall workflow of retrieving literature and past experiences, planning research directions, generating candidate solutions, and refining them based on evaluation feedback closely resembles existing retrieval-augmented research agents. Although WikiEvolve introduces a structured wiki and surrogate-guided candidate selection, the paper does not sufficiently establish the benefits of its proposed knowledge–experience coupling over a conventional agent equipped with retrieval, persistent memory, and iterative planning. A direct comparison under matched resources is needed to clarify the distinct methodological contribution.
  2. WikiEvolve requires a foundational wiki to be constructed offline for each target domain using a separate curator agent and predefined knowledge-organization procedures. Although multiple tasks can share the same domain wiki, the construction process introduces additional preparation costs and depends on the availability and quality of relevant literature. The paper does not sufficiently investigate the scalability of wiki construction or its robustness to incomplete or lower-quality knowledge bases.
Questions for authors

Q1. What distinguishes WikiEvolve from a conventional retrieval-augmented research agent?

Existing research agents can already retrieve literature, maintain experimental memories, formulate research plans, and iteratively refine candidate solutions. Could the authors compare WikiEvolve against such an agent under matched retrieval and evaluation budgets? In particular, it would be useful to isolate the contribution of explicitly associating foundational knowledge with accumulated experiences.

Q2. How scalable and generalizable is the construction of the foundational wiki?

Could the authors report the computational cost and human intervention involved in constructing the domain-specific wikis? How does WikiEvolve perform when the wiki is automatically constructed for a new domain or contains incomplete or noisy information? Such experiments would help clarify the practical scalability and robustness of the proposed framework.

Q3. How sensitive are the BO generalization results to the deployment evaluation budget?

WikiEvolve and Experience Only achieve comparable performance at the tuning budget of 60 evaluations, while WikiEvolve performs better at 200 evaluations. Could the authors evaluate the evolved algorithms across a wider range of budgets, such as 100, 300, and 500 evaluations?

6 comments
AuthorsOpenReview ↗

Response to Reviewer z4NN (Part 1/5)

We thank the reviewer for finding the paper clear and easy to follow, for recognizing that WikiEvolve addresses an important limitation of experience-only evolution with a coherent framework, and for the insightful questions. In summary, we (i) clarify how WikiEvolve differs from a conventional retrieval-augmented research agent and compare it with EoH, ReEvo, and OpenEvolve equipped with retrieval under a matched token budget, (ii) report the cost of constructing the foundational wiki and the performance of WikiEvolve with fewer or unrelated source cards, and (iii) evaluate the best-found BO algorithms at 100, 300, and 500 BO evaluations. Detailed answers are provided below. Additionally, we have revised the PDF version of the paper, with all revisions marked in blue. We apologize that, due to time constraints, we have not yet added the results for the DeepSeek model. We plan to include these results in a subsequent revision.

W1. The overall workflow of retrieving literature and past experiences, planning research directions, generating candidate solutions, and refining them based on evaluation feedback closely resembles existing retrieval-augmented research agents. Although WikiEvolve introduces a structured wiki and surrogate-guided candidate selection, the paper does not sufficiently establish the benefits of its proposed knowledge–experience coupling over a conventional agent equipped with retrieval, persistent memory, and iterative planning. A direct comparison under matched resources is needed to clarify the distinct methodological contribution.

Q1. What distinguishes WikiEvolve from a conventional retrieval-augmented research agent? Existing research agents can already retrieve literature, maintain experimental memories, formulate research plans, and iteratively refine candidate solutions. Could the authors compare WikiEvolve against such an agent under matched retrieval and evaluation budgets? In particular, it would be useful to isolate the contribution of explicitly associating foundational knowledge with accumulated experiences.

W1 + Q1. Difference from a conventional retrieval-augmented research agent

Answer W1 + Q1. Thank you for insightfully pointing out that existing retrieval-augmented research agents can already retrieve literature, maintain experimental memories and iteratively refine research plans and candidate solutions. To better distinguish WikiEvolve from these agents, we highlight the key differences, report results under matched retrieval and token budgets, and examine why WikiEvolve outperforms these agents. We have revised the Introduction, Sec. 4.1–4.2 (Table 1), and the Related Work (Sec. 5) accordingly.

WikiEvolve differs in four key ways (Sec. 3.1–3.3, App. B.2):

  1. Associating foundational knowledge with accumulated experiences. Each experience records the algorithm, code and the utility. In WikiEvolve, each algorithm additionally cites the source cards used for its generation. This is important for exploiting the structured foundational wiki for better retrieval of foundational knowledge.
  2. Structured foundational wiki. The foundational wiki is a graph of source cards and concept pages joined by hyperlinks as the edges. The edges encode prior knowledge about how methods relate (e.g., one extends the other), which lexical similarity alone may not capture. WikiEvolve decides which source card to retrieve by its relevance score. A source card’s relevance is the sum of two terms, namely its lexical similarity to the research questions, and a graph bonus that is larger when the card has a shorter path to the top-ranked concept pages (Eq. 3, App. B.2). This structure helps WikiEvolve to discover other useful methods related to a method that contributed the current best algorithm. Existing research agents can only use raw papers or flat source cards instead.
  3. Enforced exploration of foundational knowledge. Recently retrieved source cards are penalized to reduce repeated use of the same source cards and encourage exploration of the wiki. Existing research agents may not penalize repeated use.
  4. Surrogate model and experience guide the next selection. WikiEvolve selects promising candidates for evaluation based on contextual relevance to the research questions, expected improvement in utility, and uncertainty in the utility estimates produced by the surrogate model. Unlike existing research agents, WikiEvolve is explicitly designed to balance exploitation of candidates with high predicted utility and exploration of candidates with high uncertainty.
AuthorsOpenReview ↗

Response to Reviewer z4NN (Part 2/5)

W1 + Q1. Difference from a conventional retrieval-augmented research agent (continue)

Comparison under matched budget (App. J, Table 12). To instantiate the conventional research agents, we use existing automated algorithm design frameworks (EoH, ReEvo and OpenEvolve) and equip each with retrieval from the same pool of raw papers/source cards from which the foundational wiki was built. In each iteration, the agent uses BM25 to retrieve 16 papers/source cards, which matches WikiEvolve's retrieval budget. We keep each framework's original mechanism for carrying information across iterations, which plays the role of the persistent memory in the question above: EoH [1] evolves a population of algorithms together with their natural-language thoughts; ReEvo [2] keeps a population together with short-term and long-term reflections; OpenEvolve [3] keeps a program database. All methods use the same LLM and the same token budget per self-evolution run.

Additionally, the table includes WikiEvolve (Flat Source Cards) and WikiEvolve (No surrogate), which assess the contributions of (2) the structured foundational wiki and (4) the surrogate model, respectively. (These runs use 50 iterations, compared with 90 in Sec. 4.4 and App. E.4, hence the different values.)

TaskMethodImprovement (%) ↑Total token (M)
Graph ColoringEoH (Given sources)8.2511.98
ReEvo (Given sources)2.9911.96
OpenEvolve (Given wiki)4.8111.40
WikiEvolve with flat-source cards0.0011.90
WikiEvolve without surrogate model6.5712.01
WikiEvolve (Ours)18.0111.87
Max-DisCSPEoH (Given sources)22.1114.48
ReEvo (Given sources)17.8114.48
OpenEvolve (Given wiki)20.0814.94
WikiEvolve with flat-source cards21.8414.44
WikiEvolve without surrogate model24.1714.89
WikiEvolve (Ours)26.7414.39
Weighted DCOPEoH (Given sources)1.825.63
ReEvo (Given sources)1.925.65
OpenEvolve (Given wiki)5.115.53
WikiEvolve with flat-source cards4.005.62
WikiEvolve without surrogate model6.725.83
WikiEvolve (Ours)6.795.58

Furthermore, WikiEvolve explores more of the foundational knowledge (App. I.2, Table 11 and Fig. 11). We randomly sample 20 sources of the corpus and count how many of them each method uses at least once within the first 6M tokens of a self-evolution run. EoH and ReEvo retrieve from the same sources, yet similarity-based retrieval keeps returning the sources closest to the initial algorithm (ALS / DSA variants and MGM / MGM-2), and they use only 1 to 3 of the 20 sources. WikiEvolve uses 15 to 17 of them.

TaskReEvo (Given sources)EoH (Given sources)WikiEvolve
Graph Coloring1115
Max-DisCSP3316
Weighted DCOP1217

Useful foundational knowledge is often not the most lexically similar (App. I.1, Fig. 10). On Graph Coloring, the lineage of the best-found algorithm of WikiEvolve cites the source cards of Bounded Max-Sum, DPOP, ConcFB, NCBB, PC-SyncBB, and Deep Coordination Graphs, and EoH and ReEvo use none of the corresponding papers. Four of the seven source cards cited in this lineage rank outside the six most similar to the best-found algorithm (by TF–IDF cosine similarity), including DPOP at rank 16 and Deep Coordination Graphs at rank 19 among 20 sources. On Max-DisCSP, all five cited source cards rank between 10 and 18. Useful source cards therefore need not be among those most similar to the best-found algorithm. In WikiEvolve, the graph bonus and the repetition penalty bring these source cards to the agent, and the experiential memory records which of them improved the utility. This is the contribution of coupling foundational knowledge with experiences.

[1] Liu, F., Tong, X., Yuan, M., Lin, X., Luo, F., Wang, Z., Lu, Z., and Zhang, Q. Evolution of heuristics: Towards efficient automatic algorithm design using large language model. ICML, 2024.

[2] Ye, H., Wang, J., Cao, Z., Berto, F., Hua, C., Kim, H., Park, J., and Song, G. ReEvo: Large language models as hyper-heuristics with reflective evolution. NeurIPS, 2024.

[3] Sharma, A. OpenEvolve: An open-source evolutionary coding agent. GitHub, 2025.

AuthorsOpenReview ↗

Response to Reviewer z4NN (Part 3/5)

W2. WikiEvolve requires a foundational wiki to be constructed offline for each target domain using a separate curator agent and predefined knowledge-organization procedures. Although multiple tasks can share the same domain wiki, the construction process introduces additional preparation costs and depends on the availability and quality of relevant literature. The paper does not sufficiently investigate the scalability of wiki construction or its robustness to incomplete or lower-quality knowledge bases.

Q2. How scalable and generalizable is the construction of the foundational wiki? Could the authors report the computational cost and human intervention involved in constructing the domain-specific wikis? How does WikiEvolve perform when the wiki is automatically constructed for a new domain or contains incomplete or noisy information? Such experiments would help clarify the practical scalability and robustness of the proposed framework.

W2 + Q2. Cost and robustness of the foundational wiki

We thank the reviewer for these questions.

Computational cost and human intervention (App. A, Table 2). Human intervention is limited to specifying the sources from which the corpus is assembled (App. A). For DCOP, the corpus is directly curated to cover the taxonomy of the survey of Fioretto et al. [1]. For BO, a person chooses one survey [2] and two community-curated lists of BO papers, namely awesome-bo [3] and awesome-bayesian-optimization [4] on GitHub, and the curator agent selects the sources from the works they cite or list by citation snowballing [5]. Following a human-written skill, the curator agent then writes, cross-references, and verifies all source cards and concept pages. The cost below is paid once per domain, since the wiki is built once and shared by all tasks of that domain. Surveys and community-curated lists of papers are available for many other research topics, so the corpus for a new domain can be constructed in the same way.

BO wikiDCOP wiki
Source cards6340
Concept pages1831
Input tokens (M)43.6632.34
Output tokens (M)0.200.15

This corresponds to about 0.7M (BO) and 0.8M (DCOP) input tokens per source. We have added the construction cost to App. A (Table 2), together with a paragraph on the scalability of the wiki.

Automatic construction for a new domain. Both the DCOP and BO wikis are constructed by the curator agent using the same reusable skill. The skill is domain-agnostic and can be applied to any new domain once a human specifies the sources. The specified sources can be the complete corpus, as in the direct curation for DCOP, or a seed from which the curator agent expands the corpus, as in the citation snowball sampling for BO. The token cost of constructing the wiki grows with the number of selected sources.

[1] Ferdinando Fioretto, Enrico Pontelli, and William Yeoh. Distributed constraint optimization problems and applications: A survey. Journal of Artificial Intelligence Research, 61:623–698, 2018.

[2] Xilu Wang, Yaochu Jin, Sebastian Schmitt, and Markus Olhofer. Recent advances in Bayesian optimization. ACM Computing Surveys, 55(13s), 2023.

[3] https://github.com/richardcsuwandi/awesome-bo

[4] https://github.com/yiminzme/awesome-bayesian-optimization

[5] Claes Wohlin. Guidelines for snowballing in systematic literature studies and a replication in software engineering. In Proc. EASE, pp. 1–10, 2014.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

Dear ACs,

We appreciate the time, effort, and valuable feedback from the reviewers and ACs. We summarize the discussions with the reviewers:

I. Contributions and Significance

  • WikiEvolve is a meaningful contribution to automated algorithm design (AfX5):
    • We proposed how to construct a structured foundational wiki and how to retrieve from it, so that the agent reaches useful knowledge beyond similarity. This differs from existing wiki-based agents, which target the knowledge most relevant to a query.
    • We proposed research planning informed by experiential memory and the selection of promising candidates before implementation. This differs from existing algorithm design frameworks, which generate the code of every candidate.
    • The coupling of foundational and experiential knowledge is non-trivial, as removing any of these components lowers the improvement, and its designs are motivated by the characteristics of algorithm design.
  • We addressed an important limitation of experience-only self-evolving agents with a coherent framework that connects research planning, the foundational wiki, experience retrieval, and candidate selection in a single loop (z4NN, Hbth).
  • The mechanism is described concretely enough to inspect (jctU, Hbth).
  • Experiments across two quite different domains and two LLMs show consistent improvements over experience-only evolution (z4NN, jctU, Hbth, AfX5), and the ablations offer useful insights into how foundational knowledge affects evolutionary search (z4NN, jctU).
  • The reviewers recognized the clarity of our writing (z4NN, jctU, Hbth).

II. Key Reviewer Concerns/Questions Addressed

Reviewer AfX5:

  • As suggested, we compared WikiEvolve with EoH and ReEvo, together with OpenEvolve, when they are given the same foundational knowledge under comparable budgets. WikiEvolve still achieves the highest improvement, and the reviewer confirmed that the concerns about the experimental evaluation were sufficiently addressed.
  • We clarified the technical novelty of WikiEvolve when compared with existing wiki-based agents and with existing algorithm design frameworks, and why their combination is non-trivial. We also explained what makes algorithm design distinctive and how this motivates our designs. The reviewer acknowledged our contribution and raised the rating.

Reviewer z4NN:

  • We clarified how WikiEvolve differs from a conventional retrieval-augmented research agent, and showed that existing frameworks equipped with retrieval from the same sources remain below WikiEvolve and explore much less of the foundational knowledge.
  • We reported the cost of constructing the wiki, which is paid once per domain and needs only a human to specify the sources. We also showed that WikiEvolve remains effective with incomplete wikis and with irrelevant source cards.
  • As suggested, we evaluated the BO algorithms at more evaluation budgets, where WikiEvolve generally remains the best.

Reviewer jctU:

  • As suggested, we re-evaluated the best-found DCOP algorithms on held-out instances and reported paired comparisons for BO, which show that the gains of WikiEvolve hold beyond the instances used during self-evolution.
  • We reported the realized resources of all methods and added baselines under matched budgets, which show that the gains of WikiEvolve do not come from using more resources.
  • We clarified that the Flat Source Cards ablation removes only the graph structure from retrieval and keeps everything else unchanged.

Reviewer Hbth:

  • We reported the number of evaluations and token usage of all methods under comparable budgets.
  • As suggested, we evaluated OpenEvolve given the same foundational wiki, together with EoH and ReEvo. Access to the wiki helps OpenEvolve, but it remains below WikiEvolve.
  • We explained how WikiEvolve differs from hill climbing, added population-based evolution as future work, and corrected all the typos.

Concluding our discussions, reviewer AfX5 indicated satisfaction with our responses and raised the rating; reviewers z4NN, jctU, and Hbth have not indicated further questions.

In our revised paper (edits in blue), we clarified the technical challenges and novel contributions in the Introduction and reorganized the Related Work in Sec. 5 (AfX5, z4NN), added the new baselines in Sec. 4.1 and Table 1 (AfX5, z4NN, Hbth), the token efficiency analysis in Sec. 4.6 and the realized resources in App. J (jctU, Hbth), the clarification of Flat Source Cards in Sec. 4.4 (jctU), the held-out DCOP results in App. E.6 (jctU), more BO budgets in App. E.7 (z4NN), the wiki construction cost in App. A (z4NN), the analysis of source use in App. I (AfX5, z4NN), the robustness to irrelevant source cards and incomplete wikis in App. K and L (z4NN), and population-based evolution as future work in Sec. 6 (Hbth).

We thank the ACs and all reviewers for evaluating our submission.

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.