acceptodds
Under review as a conference paper at ICLR 2027

HER: Hierarchical Evolutionary Reasoning for LLM-driven Algorithmic Optimization

Abstract

Large language model (LLM)-driven evolutionary agents hold strong promise for algorithmic optimization, yet most existing paradigms still rely on flat population structures and uniform compute allocation. This fundamentally limits their ability to coordinate broad exploration with sustained refinement. To address this, we present HER (ierarchical volutionary easoning). By organizing candidate programs into interacting layers, the framework couples cross-layer population management with bellwether-guided transitions and layer-wise prompt-context selection. Under limited test-time budgets, such a structure enables adaptive compute allocation and more structured search dynamics, allowing promising candidates to guide higher-level refinement while preserving diversity in lower layers. Extensive evaluation across diverse benchmarks, spanning both mathematical and system tasks, shows that HER consistently outperforms state-of-the-art baselines in solution quality, achieving an average relative improvement of 6.67% on mathematical benchmarks and 12.72% on system benchmarks. These results suggest that hierarchical organization is not merely a structural modification, but a promising strategy for improving LLM-driven algorithmic optimization.

Primary area:
optimization
Keywords:
large language models, self-evolving, hierarchical search, automated algorithm discovery, test-time reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TmZ8OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

HER is an LLM-driven program optimization framework that organizes candidates into hierarchical populations with promotion, demotion, and elimination. Experiments cover six mathematical and five systems benchmarks, with multiple baselines, ablations, and an approximate token-matched comparison. The inconsistencies in task validity and algorithm specification undermine confidence in the claims.

Critical strengths
  1. Concrete and well-motivated design. Returning stagnant candidates to exploratory layers provides a plausible mechanism for combining sustained refinement with renewed exploration.
  2. Useful ablations. Tables 4–6 examine hierarchy, rewriting, initialization, inertia, and reference selection. Transferring HER’s initialization to OpenEvolve also addresses an important alternative explanation.
Critical weaknesses
  1. The signal-processing examples appear inconsistent with the stated causal constraints. Appendix C.1 describes real-time, causal, low-latency filtering. However, Appendix P includes backward RTS smoothing and forward–backward filtering, with zero-phase processing enabled by default. If the evaluator supplies complete sequences and scores the resulting outputs directly, these examples do not establish performance under the stated online constraints. The evaluation protocol and its relationship to the reported results require clarification. This concern applies specifically to this task and its case studies.
  2. Reference-selection probabilities are specified inconsistently. Equations (9)–(10) define reference inclusion probabilities as ϕ\phi and ψ\psi, whereas Appendix D states that implementation conditions r>ϕr>\phi and r>ψr>\psi yield probabilities 1−ϕ1-\phi and 1−ψ1-\psi. These define different algorithms. Under the latter interpretation, the additional factor of 0.5 in Equation (16) increases global-reference inclusion, contrary to its stated purpose of maintaining layer-level dominance. This ambiguity affects a central mechanism and reproducibility.
  3. The novelty claim and attribution to migration need sharper support. The introduction’s characterization of prior approaches as using uniform resource allocation is too broad: AdaEvolve already dynamically allocates resources across populations. HER’s more distinctive contribution is its ordered candidate hierarchy and bidirectional migration. Although the hierarchy ablations are informative, the description does not establish whether layer-specific operators, prompts, and sampling policies remain controlled when hierarchy is removed. Consequently, the independent contribution of migration remains unclear.
Questions for authors
  1. How does the signal-processing evaluator enforce causality, and which programs produced the reported scores? Would those programs pass a test that changes future inputs while holding the observed prefix fixed?
  2. Which reference-inclusion rule was used in the experiments? How should Equations (9)–(10) and the explanation of Equation (16) be corrected?
  3. With layer-specific operators and prompting retained, how much performance is lost when migration or stagnation-triggered demotion is disabled?
3 comments

Thank you for your thoughtful comments and for the time devoted to reviewing our manuscript. We address each point in detail below.

Comment 1 (W1 / Q1): “The signal-processing examples appear inconsistent with the stated causal constraints. Appendix C.1 describes real-time, causal, low-latency filtering. However, Appendix P includes backward RTS smoothing and forward–backward filtering, with zero-phase processing enabled by default. If the evaluator supplies complete sequences and scores the resulting outputs directly, these examples do not establish performance under the stated online constraints. The evaluation protocol and its relationship to the reported results require clarification. This concern applies specifically to this task and its case studies.”

Response: We are grateful for these precise questions. They led us to re-examine the signal-processing evaluator and uncovered a real inconsistency between the task description and the evaluation code that we had missed. We answer the questions in order, and the short version is: the evaluator does not enforce causality, the showcased programs would not pass the test the reviewer describes, and we have re-run the comparison with that test enforced.

  • How does the evaluator enforce causality? We have to admit that it does not. The evaluator hands the complete noisy sequence to the evolved program and scores the returned filtered sequence directly. There is no check on whether a given output sample depends on later input samples. The task, including both the evaluator and the description, was taken as is from the SkyDiscover suite [1]. Its description speaks of real-time, causal filtering, but its evaluator does not enforce this, and the causal / low-latency wording in Appendix C.1 was carried over from that description without our noticing the mismatch. All five methods in Table 2 were evaluated under this same full-sequence protocol, so the comparison there is fair within the protocol, but it is not evidence on a causal task. We should have stated this clearly and we will.

  • Which programs produced the reported scores? The Signal Processing entry in Table 2 is the mean over 3 seeds of the final best program of each HER run. The two programs in Appendix P are the iteration-42 diff and the iteration-156 full rewrite on the lineage of the final best program of one of these runs. They are intermediate ancestors rather than the scored program itself, but the final best programs of all three runs likewise rely on RTS smoothing or forward-backward filtering.

  • Would they pass a prefix-fixed test? Unfortunately they would not. We implemented exactly this test as a corrected evaluator. It requires that the output has one sample per complete window (input length minus window size plus one), and that running the program on input prefixes of 50%, 70% and 85% of the sequence reproduces the corresponding prefix of the full-sequence output within 1e-6. Any violation scores 0. The final best programs of all three HER runs behind Table 2 fail this test, two because the prefix output changes with future samples and one because of a wrong output length, and so do the Appendix P programs.

With this evaluator we re-ran all five methods with 3 seeds each. The model, iteration budget and all other settings are as in the paper, and the causal requirement is also stated in the task prompt for this experiment:

MethodMean ± stdBest
OpenEvolve0.5407 ± 0.00880.5489
ShinkaEvolve0.5470 ± 0.00670.5520
ContextEvolve0.5422 ± 0.00510.5473
AdaEvolve0.5445 ± 0.00830.5493
HER (ours)0.5691 ± 0.02530.5968

All best programs here pass the causal check, and HER's best program under this protocol is a purely forward adaptive Kalman / trend filter with no backward pass. Mean scores drop for every method relative to Table 2, as expected. HER keeps the highest mean and two of its three seeds (0.5635, 0.5968) exceed every baseline run (best 0.5520).

In the revision we will (1) rewrite Appendix C.1 to describe the task as actually evaluated and drop the causal / low-latency claims from the benchmark description, the case study and the original task prompt, (2) add the causal-enforced experiment and evaluator description as a new appendix next to Table 2, with a footnote on the Signal Processing row pointing to it, (3) mark the Appendix P programs as non-causal and add the best causal HER program as a case study, and (4) release the corrected evaluator with our code. We agree this concern is specific to this task and its case study, and we thank the reviewer for pinpointing it.


[1] Liu et al. "SkyDiscover: A Flexible, Adaptive Framework for AI-Driven Scientific and Algorithmic Discovery."

Comment 2 (W2 / Q2): “Reference-selection probabilities are specified inconsistently. Equations (9)–(10) define reference inclusion probabilities as (ϕ) and (ψ), whereas Appendix D states that implementation conditions (r>ϕ) and (r>ψ) yield probabilities (1-ϕ) and (1-ψ). These define different algorithms. Under the latter interpretation, the additional factor of 0.5 in Equation (16) increases global-reference inclusion, contrary to its stated purpose of maintaining layer-level dominance. This ambiguity affects a central mechanism and reproducibility.”

Response: We thank the reviewers for raising this point. Equations (9)–(10) as written do not accurately describe the implementation, and we will correct them in the revision. We respond to each point below.

  • Which rule was used. All reported experiments used the rule described in Appendix D: a value r∼U[0,1]r \sim U[0,1] is drawn, and a reference is included when rr exceeds the mapping value, giving effective inclusion probabilities 1−ϕ(⋅)1-\phi(\cdot) (personal best), 1−ψℓ(⋅)1-\psi_\ell(\cdot) (layer best), and 1−ψglob(⋅)1-\psi_{\text{glob}}(\cdot) (global best). A larger fitness gap thus makes a reference less likely to be included. In the revision we will (i) correct Equations (9)–(10) to 1−ϕ(⋅)1-\phi(\cdot) and 1−ψ(⋅)1-\psi(\cdot), (ii) refer to these mappings as inclusion thresholds throughout the paper, and (iii) add a clarifying sentence in Section 3 so the main text matches Appendix D and the code that produced the reported results.

  • The factor 0.5 in Equation (16). Under the implemented rule, this factor halves the threshold for the global best and thereby increases its inclusion probability, exactly as the reviewer notes. For example, in L2 the layer best is included with probability 0.6–0.8 and the global best with 0.8–0.9; the two typically appear together in the prompt. Our statement that the factor maintains "layer-level guidance dominant" was written under the mistaken reading of ϕ\phi and ψ\psi as probabilities and will be replaced with an accurate description of the implemented behaviour (L1 never sees the global best; L2 and L3 reference it frequently, with the layer best still included at a high rate). The ablations in Tables 4–6 used this same rule, so their conclusions are unaffected. We are additionally running a control with the 0.5 factor removed (ψglob=ψℓ\psi_{\text{glob}} = \psi_\ell) and will report it in the revision.

Comment 3 (W3 / Q3): “The novelty claim and attribution to migration need sharper support. The introduction’s characterization of prior approaches as using uniform resource allocation is too broad: AdaEvolve already dynamically allocates resources across populations. HER’s more distinctive contribution is its ordered candidate hierarchy and bidirectional migration. Although the hierarchy ablations are informative, the description does not establish whether layer-specific operators, prompts, and sampling policies remain controlled when hierarchy is removed. Consequently, the independent contribution of migration remains unclear.”

Response: Thank you for raising this point. We agree that the characterization of prior approaches as using uniform resource allocation is inaccurate, since AdaEvolve already performs dynamic allocation across populations. We will revise the Introduction accordingly. To isolate the independent contribution of migration, we conducted two controlled ablations in which all layer-specific operators, prompts, sampling policies, and computational budgets were held fixed. One disables only stagnation-triggered demotion, while the other disables all cross-layer migration.

MethodCircle P. (Sq.) ↑\uparrowPrism ↑\uparrowTXN ↑\uparrow
HER2.6354 ±\pm 0.001026.50 ±\pm 0.344256.7 ±\pm 349.4
w/o stagnation demotion2.6337 ±\pm 0.002626.44 ±\pm 0.344162.9 ±\pm 112.5
w/o migration2.6314 ±\pm 0.000526.37 ±\pm 0.104121.1 ±\pm 188.0

Disabling stagnation-triggered demotion reduces performance on all three tasks, by 0.0017 on Circle Packing, 0.06 on Prism, and 93.8 on TXN. Removing migration entirely causes larger drops of 0.0040, 0.13, and 135.6, respectively. These controlled results show that migration provides gains beyond the layer-specific search policies, while stagnation-triggered demotion makes a complementary contribution by returning stalled candidates to more exploratory layers. We will add these results and clarify in the revision that HER's distinctive contribution is the ordered candidate hierarchy with bidirectional migration, rather than dynamic resource allocation alone.

AuthorsOpenReview ↗

Reminder: Discussion Period Ending in a Few Days

Dear Reviewer TmZ8,

Just a brief reminder that the author–reviewer discussion period concludes in a few days.

Thank you for your insightful questions, which pushed us to verify several details of our setup. Our response keeps the signal-processing question to that task and its case study. Under the prefix-fixed causal check, HER scores 0.5691 ± 0.0253 (best 0.5968), and the best program is a forward-only filter. We would also like to clarify that the reported runs already follow Appendix D: a reference is included when a uniform draw exceeds the threshold, and Equations (9)–(10) in the revised manuscript use the probabilities 1−ϕ and 1−ψ. Under that same rule, the factor 0.5 in Equation (16) increases global-best inclusion, and Tables 4–6 were run with it, so their conclusions stand. As for the ablation, with layer-specific operators, prompts, and budgets held fixed, full HER scores 2.6354, 26.50, and 4256.7 on Circle Packing (Square), Prism, and TXN. Removing stagnation-triggered demotion lowers these to 2.6337, 26.44, and 4162.9, and removing migration lowers them further to 2.6314, 26.37, and 4121.1. Taken together, we believe these results support framing the contribution as an ordered hierarchy with bidirectional migration, and the revised manuscript now describes it in these terms. These updates are marked in blue.

We hope this addresses your concerns, but if any point remains open, we would be glad to discuss it further before the discussion ends.

Thank you again for the time and effort you put into the review.

Best regards,

Authors

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

The paper introduces HER (Hierarchical Evolutionary Reasoning) for algorithmic optimization motivated by the research question: can we design an LLM-driven hierarchical evolutionary framework that dynamically allocates computational resources and coordinates the population, thereby yielding superior optimization performance? HER combines the idea of Particle Swarm Optimization with a hierarchical population structure. In this hierarchical structure different layers are assigned different search roles and mutation operators, so that lower layers encourage broader exploration while upper layers focus on refinement of promising solutions. The base layer (L1, ca. 60% of the population) prioritizes exploration; The intermediate layer (L2, ca 30% of population) bridges the gap between exploration and exploitation; and the top layer (L3 ca 10% of the population) is dedicated to exploitation. During evolution, a batch of candidates is updated via bellwetherguided cross-layer transitions and PSO-inspired prompt-context selection. They report state of the art results on six mathematical tasks and on five system benchmarks with gpt-oss-120b, as also providing ablation studies within the HER method.

Critical strengths

• HER beats four strong baselines on all 11 tasks with gpt-oss-120b (Table 2 and 3), and still on 10 of 11 under a token-matched budget (Table 12). • Ablations, a sensitivity analysis and the layer-dynamics analysis show effort to understand the method

Critical weaknesses

• Limited novelty. Fitness-layered populations with promotion are an old idea in evolutionary computation (Hierarchical Fair Competition, ALPS), and neither is cited. PSO “in language” was already done in LMPSO, which is cited. HER mostly combines these with the OpenEvolve setup, which limits the conceptual contribution, although the empirical results are consistent.

Questions for authors

None

2 comments

Thank you very much for your valuable comments and the time you devoted to reviewing our manuscript. We address each of your comments in detail below.

Comment (W1): “Limited novelty. Fitness-layered populations with promotion are an old idea in evolutionary computation (Hierarchical Fair Competition, ALPS), and neither is cited. PSO “in language” was already done in LMPSO, which is cited. HER mostly combines these with the OpenEvolve setup, which limits the conceptual contribution, although the empirical results are consistent.”

Response: Thank you for this valuable suggestion. We think these methods provide complementary conceptual support and useful context for HER’s design. HFC[1] uses fitness-stratified subpopulations with one-way upward migration, while ALPS[2] separates individuals by age rather than fitness and periodically refreshes its youngest layer. LMPSO[3] demonstrates that PSO updates can be expressed through language prompts. Together, these works establish population stratification, migration, and language-mediated optimization as useful but separate design principles.

The key distinction is that HFC, ALPS, and LMPSO instantiate these ideas separately and with different search dynamics. HFC permits only one-way upward migration, ALPS organizes layers by age rather than fitness, and LMPSO uses a language prompt to simulate a fixed PSO update. By contrast, HER couples behaviour-specialized, fitness-ordered layers with bidirectional, state-dependent migration for LLM-driven program evolution. Each layer jointly configures distinct mutation operators, prompts, sampling policies, and reference schedules, while candidates are promoted by fitness and demoted upon stagnation, allowing migration to coordinate multiple search behaviours. OpenEvolve, by comparison, maintains a single flat population. Thus, HER’s conceptual novelty lies not in any component in isolation, but in coupling them into a unified search architecture. We will expand the related work section to clarify both the conceptual connections and the methodological distinctions.


[1] Hu et al. "The Hierarchical Fair Competition (HFC) Framework for Sustainable Evolutionary Algorithms."

[2] Hornby, G. S. "ALPS: The Age-Layered Population Structure for Reducing the Problem of Premature Convergence."

[3] Shinohara et al. "Large Language Models as Particle Swarm Optimizers."

AuthorsOpenReview ↗

Reminder: Discussion Period Ending in a Few Days

Dear Reviewer QwBN,

Just a brief reminder that the author–reviewer discussion period concludes in a few days.

Thank you for raising the question about how HER relates to prior work. In the revised manuscript, we now cite HFC, ALPS, and LMPSO in the related-work section and explain, one by one, where HER differs from each. To briefly recap: HFC only promotes a candidate upward, while ALPS never uses fitness to define its layers. Candidates there are grouped by age, and the youngest layer is refreshed periodically. LMPSO, on the other hand, is not a population scheme at all. The prompt simply carries out one fixed PSO step. HER is different in that it keeps a fitness ordering, gives each layer its own operators, prompts, sampling, and references, and moves a candidate back down when it stops improving. To check that this cross-layer movement matters beyond those per-layer choices, we turned migration off and left the per-layer settings fixed. The score dropped on Circle Packing, Prism, and TXN, most notably on TXN. The added discussion is marked in blue, and the rest of the experimental results remain unchanged.

We hope this helps clarify the distinction, but if anything is still unclear, we would be more than happy to discuss it further within the remaining discussion period.

Thank you again for your time, and for your kind note that the empirical results are consistent.

Best regards,

Authors

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

The paper proposes an LLM-driven framework for algorithmic optimization. HER organizes candidate programs into three layers with different exploration and refinement roles, using fitness and stagnation to govern promotion, demotion, and elimination. Its two-stage update generates a natural-language improvement direction from search history and reference programs, then revises the candidate through a full rewrite or edits. Experiments compare HER with existing evolutionary coding methods on six mathematical and five systems benchmarks.

Critical strengths

The paper covers 11 tasks with three runs each.

Critical weaknesses

These issues affect the central performance and efficiency claims:

  1. Appendix C.1 requires causal filtering, but both signal-processing programs use backward RTS smoothing, which depends on future samples (Appendix P). Their scores do not demonstrate improvements on the stated causal task. Presenting these programs as advances without acknowledging this violation undermines confidence in the evaluation.
  2. HER’s largest gains occur on EPLB, LLM-SQL, and Cloudcast, where reproduced baselines differ substantially from the published results used for comparison. ShinkaEvolve’s best EPLB score is 0.2238 versus 0.1272, and AdaEvolve’s mean Cloudcast score is 0.0160 versus 0.0016, comparing the paper’s gpt-oss-120b runs with published Gemini-3-Pro results (Tables 3 and 8). Appendix F.2 acknowledges differing settings, but the comparability of the evaluators and task constraints remains unresolved.
  3. Equations (9) and (10) specify reference-inclusion probabilities ϕ and ψ, whereas Appendix D describes an implementation using 1−ϕ and 1−ψ. This discrepancy changes the reference selection mechanism.
  4. The paper presents “bellwether-guided” transitions as a central component but does not provide an explicit operational definition or explain how this component differs from the transition rules.
  5. Averaging relative improvements over each task’s best baseline mean yields approximately 3.8% for mathematics and 12.3% for systems, rather than the advertised 6.67% and 12.72%. The aggregation rule of the reported percentages is not specified.
  6. Nominally identical HER configurations have different Circle Packing and TXN statistics across Tables 2–6 and 12.
  7. ShinkaEvolve’s Circle Packing mean increases from 2.5842 to 2.5947 despite already using 2.1× HER’s tokens, while all four LLM-SQL baseline scores change although none qualifies for additional updates under the stated protocol (Appendix M).
Questions for authors
  1. Does the signal-processing evaluator enforce the causal constraint in Appendix C.1? How are the backward RTS smoothers in the showcased programs (Appendix P) compatible with this requirement?
  2. Which evaluator versions, configurations, and task constraints were used for EPLB, LLM-SQL, and Cloudcast? What explains the substantial discrepancies between the reproduced baselines and the published results in Table 8?
  3. Which reference inclusion probabilities were used in the reported experiments: ϕ and ψ, as specified in Equations (9) and (10), or 1−ϕ and 1−ψ, as described in Appendix D?
  4. What precisely constitutes a “bellwether-guided” transition, and how does this mechanism differ from the transition rules?
  5. What baseline, per-task statistic, and aggregation formula produce the improvements of 6.67% and 12.72%? How do these reconcile with approximately 3.8% and 12.3% when averaging relative improvements over each task’s best baseline mean?
  6. Were the HER results in Tables 2–6 and 12 computed from the same runs and configurations? What explains the differing Circle Packing and TXN statistics, and which values should be treated as definitive?
  7. How were the token-matched results obtained for baselines that already met or exceeded HER’s token budget? Why do ShinkaEvolve’s Circle Packing score and all four LLM-SQL baseline scores change in Appendix M?
7 comments

We thank the reviewer very much for the constructive and in-depth suggestions and appreciation of our work. We provide responses to the reviewer’s specific questions and concerns below.

Comment 1 (W1 / Q1): “Appendix C.1 requires causal filtering, but both signal-processing programs ... this violation undermines confidence in the evaluation.

Response: Thank you for pointing this out. It helped us catch a mismatch between the task description and the evaluation code that we had overlooked. We agree with the reviewer's point: the evaluator does not enforce the causal constraint, and the backward RTS smoothers in Appendix P are not compatible with the description in Appendix C.1. We should have caught this before submission. Below we explain how it happened, what Table 2 actually compares, and report new results with the constraint enforced.

  • After reading the review we went back and checked the signal processing evaluator. It passes the whole sequence to the program and scores the returned output directly. Nothing in it checks whether the output uses future samples. The causality / low-latency wording in Appendix C.1 was adapted from the task description of the SkyDiscover suite [1], and we ran our experiments with the SkyDiscover evaluator as is (this is also the evaluator used by AdaEvolve [2] and EvoX [3], the two search methods built into SkyDiscover). The description and the code simply disagree, and we did not notice. So the programs in Appendix P are valid solutions to the task that was actually evaluated (offline filtering of a fully observed sequence), but not to the task Appendix C.1 describes, and we should not have presented them that way. In the revision we will rewrite Appendix C.1 to describe the task as it is really evaluated, remove the causality and low-latency claims from the benchmark description, the case studies and the original task system prompt, and add a note in Appendix P stating that the showcased programs are non-causal.

  • All five methods in Table 2 were run under this same full-sequence protocol, and none of them was ever restricted to causal solutions. So the lack of causality is a property of the shared protocol, not something specific to HER, and the ranking in Table 2 is still a fair comparison within that protocol. We will say this explicitly in the revision so that readers do not read Table 2 as evidence on a causal task.

  • To check whether HER's advantage survives when causality is actually enforced, we wrote a corrected evaluator and re-ran all five methods with it. The corrected evaluator requires that (1) the output length equals len(input) − window_size + 1; (2) for input prefixes at 50%, 70% and 85% of the sequence, running the program on the prefix reproduces the corresponding prefix of the full-sequence output within 1e-6; (3) any violation scores 0. Condition (2) is essentially the "fix the observed prefix, change the future" test, and it rules out zero-phase filtering, backward smoothing and centered windows; the programs showcased in Appendix P all fail it and score 0 under this evaluator. This constraint is also explicitly stated in the task system prompt. All remaining settings are identical to the protocol described in the paper. We report the results in the following table.

    MethodMean ± stdBest
    OpenEvolve0.5407 ± 0.00880.5489
    ShinkaEvolve0.5470 ± 0.00670.5520
    ContextEvolve0.5422 ± 0.00510.5473
    AdaEvolve0.5445 ± 0.00830.5493
    HER (ours)0.5691 ± 0.02530.5968

All the best programs reported here pass the causal check, so these numbers do reflect performance on the causal task. Mean scores drop for every method compared with Table 2, which is expected once future samples are unavailable. HER still has the highest mean and the highest single run, and two of its three seeds (0.5635, 0.5968) are above every baseline run (best 0.5520). We will add this table and the evaluator description as a new appendix experiment next to Table 2 (which we keep because its protocol matches prior work), and release the corrected evaluator together with our code.

We hope this addresses the concern. We sincerely thank the reviewer for the meticulous review, It helped us find a mistake we had missed.


[1] Liu et al. "SkyDiscover: A Flexible, Adaptive Framework for AI-Driven Scientific and Algorithmic Discovery."

[2] Cemri et al. "AdaEvolve: Adaptive LLM Driven Zeroth-Order Optimization."

[3] Li et al. "EvoX: Evolving Large Language Models with Large Language Models."

Comment 2 (W2 / Q2): “HER’s largest gains occur on EPLB, LLM-SQL, and Cloudcast, where reproduced baselines differ substantially ... Gemini-3-Pro results (Tables 3 and 8). Appendix F.2 acknowledges ... of the evaluators and task constraints remains unresolved.”

Response: Upon re-inspection of the programs behind the anomalous scores and the corresponding evaluators, we found that reward hacking occurred to varying degrees during the search: the evaluators did not sufficiently encode the task constraints, allowing constraint-violating programs to obtain high rewards, which the search then exploited without reaching a valid solution. We have accordingly patched the evaluators for EPLB, LLM-SQL, and Cloudcast, and re-run all experiments with the corrected versions. Details and updated results are provided below.

  • LLM-SQL. By inspecting the evaluator and the high-scoring results (e.g., HER = 0.9780), we found that the original evaluator scored any table returned by the program: dropping columns and computing the prefix-match metric on the remaining shorter text inflated the score, which is a typical case of reward hacking. We therefore revised the evaluator to accept only results that are a reordering of the same table: the row count must be unchanged, columns outside the merged groups must be preserved, a merged group may only disappear as a whole, and any other deleted column receives a score of 0. The corrected results are shown below.

    MethodMean ± stdBest
    OpenEvolve0.7002 ± 0.01370.7085
    ShinkaEvolve0.7130 ± 0.00250.7145
    ContextEvolve0.7119 ± 0.00640.7159
    AdaEvolve0.7111 ± 0.00370.7149
    HER (ours)0.7236 ± 0.00790.7327
  • EPLB. The issue lies in the load accumulation step of the evaluator: experts with positive load but zero replicas are skipped, so their traffic enters neither the average nor the maximum load, overstating the balance. This is another case of reward hacking. To prevent this, the revised evaluator requires that every expert with positive load has at least one replica, that exactly 288 replicas are used per layer, and that the physical slots are exactly the integers from 0 to 287, with any violation scored as 0. All methods have been re-run with the revised evaluator, and the scores of the legitimate programs are shown below.

    MethodMean ± stdBest
    OpenEvolve0.1427 ± 0.00170.1446
    ShinkaEvolve0.1432 ± 0.00170.1451
    ContextEvolve0.1436 ± 0.00190.1448
    AdaEvolve0.1456 ± 0.00010.1457
    HER (ours)0.1560 ± 0.00930.1621
  • Cloudcast. We found that this evaluator also admits reward hacking: it awards transition scores even when some targets are unvisited, so the ignored traffic never enters the cost. The revised evaluator now accepts a result only if all five fixed configurations succeed, requiring each partition to form a connected path from the source to every target, with cost and bandwidth taken from the network. Any violation scores 0. We re-ran all five methods under this evaluator, and the results are shown below.

    MethodMean ± stdBest
    OpenEvolve0.0011 ± 0.00010.0011
    ShinkaEvolve0.0014 ± 0.00020.0015
    ContextEvolve0.0010 ± 0.00010.0011
    AdaEvolve0.0015 ± 0.00000.0015
    HER (ours)0.0015 ± 0.00010.0016

Meanwhile, we have also checked the evaluators and results of all other tasks. Except for the signal processing task mentioned in Comment 1, for which we have made revisions and additional experiments, all other tasks yield valid results. Thank you again for pointing out this issue, which helped us improve the soundness of our experiments. We will update these results in the revised manuscript.

Comment 3 (W3 / Q3): “Equations (9) and (10) specify reference-inclusion probabilities ϕ and ψ, whereas Appendix D describes an implementation using 1−ϕ and 1−ψ. This discrepancy changes the reference selection mechanism. ”

Response: We thank the reviewer for catching this typo in the description of the reference-selection rule. All reported experiments used the rule described in Appendix D, i.e., inclusion probabilities 1−ϕ(⋅)1-\phi(\cdot), 1−ψℓ(⋅)1-\psi_\ell(\cdot), and 1−ψglob(⋅)1-\psi_{\text{glob}}(\cdot). The phrases "with prob. ϕ\phi" and "with prob. ψ\psi" in Equations (9) and (10) are a writing error on our part: ϕ\phi and ψ\psi are in fact inclusion thresholds, and a reference is included when a uniform draw r∼U[0,1]r \sim U[0,1] exceeds the corresponding threshold. Consequently, a larger fitness gap leads to a lower inclusion probability, which is precisely the behaviour described in the last paragraph of Appendix D and implemented in the reported experiments. In the revision, we will rewrite Equations (9) and (10) accordingly and consistently refer to these mappings as "inclusion thresholds" throughout the paper, ensuring that the main text specifies exactly the algorithm that produced the reported results.

Comment 4 (W4 / Q4): “The paper presents “bellwether-guided” transitions as a central component but does not provide an explicit operational definition or explain how this component differs from the transition rules.”

Response: Thank you for raising this question. “Bellwether” was intended as a metaphor for a leading, high-fitness individual, rather than as a separate transition mechanism. A “bellwether-guided” transition is therefore a transition governed by the promotion, demotion, and replacement rules already described in the paper: promising candidates are promoted according to their fitness relative to the leading and boundary candidates in higher layers, while stagnant candidates are returned to lower, more exploratory layers. Once promoted, these strong candidates undergo further refinement and may serve as layer-best or global-best references, allowing their successful solution patterns to guide the subsequent evolution of other candidates. Thus, “bellwether-guided transitions” and the transition rules are not two different sets of rules. The former was intended to describe, metaphorically, the search dynamics produced by the latter. We agree that the manuscript did not make this relationship clear and that the phrase could be read as introducing an additional algorithmic component. In the revision, we will either replace this metaphor with more precise terminology, such as “fitness- and stagnation-controlled cross-layer migration,” or define it explicitly where it first appears, and we will revise the overview, method description, and figure caption consistently to remove this ambiguity.

Comment 5 (W5 / Q5): “Averaging relative improvements over each task’s best baseline mean yields approximately 3.8% for mathematics and 12.3% for systems, rather than the advertised 6.67% and 12.72%. The aggregation rule of the reported percentages is not specified. ”

Response: We thank the reviewer for checking this aggregation. The numbers 6.67% and 12.72% were intended to represent, for each task, the relative improvement of HER’s three-run average over the average of the best baseline for that task. Upon re-checking, we found that two tasks were compared against incorrect baselines, and it is precisely these two substitutions that caused the reported percentages.

  • On Heilbronn (Convex) we used AdaEvolve’s mean of 0.0230, which gives (0.0287−0.0230)/0.0230=+24.78%(0.0287-0.0230)/0.0230=+24.78\%. The best baseline mean is ShinkaEvolve at 0.0267, so the correct figure is (0.0287−0.0267)/0.0267=+7.49%(0.0287-0.0267)/0.0267=+7.49\%.
  • On Prism we used ContextEvolve’s mean of 25.7013, which gives (26.5030−25.7013)/25.7013=+3.12%(26.5030-25.7013)/25.7013=+3.12\%. The best baseline mean is AdaEvolve at 26.2560, so the correct figure is (26.5030−26.2560)/26.2560=+0.94%(26.5030-26.2560)/26.2560=+0.94\%.

Every other task already used that task’s best baseline mean. Averaging the six mathematical ratios with the incorrect Heilbronn (Convex) entry yields 6.67%, and averaging the five system ratios with the incorrect Prism entry yields 12.72%. Replacing only those two entries gives 3.78% and 12.29%, which match the reviewer’s figures of approximately 3.8% and 12.3%.

Taking together the corrections to EPLB, LLM-SQL, and Cloudcast in Comment 2, the recomputed result is that HER improves the mathematics tasks by 3.78% and the systems tasks by 2.64%. We have included this in the revised paper.

Comments on the paper

1 comment

Dear Reviewers,

The rebuttal deadline is approaching. I am leaving this friendly reminder to encourage active discussion. Thank you for your valuable time and engagement!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.