acceptodds
Under review as a conference paper at ICLR 2027

Evade: Evolvability-guided Algorithmic Discovery

10 views

Abstract

Large language models (LLMs) are increasingly used as mutation operators in evolutionary search, enabling the discovery of novel algorithms for problems in mathematics, computer systems, and other domains. Yet existing LLM-based frameworks largely inherit the classical evolutionary paradigm of selecting candidates for mutation based on quality and diversity. These criteria capture the current state of individual candidates—such as their fitness (quality) or how they differ from other selected candidates (diversity)—but do not capture their potential for further evolution. We introduce evolvability as a search signal: a candidate's capacity to generate offspring that improve fitness or exhibit behaviors that differ from those of its ancestors. We then propose Evade, which uses the estimated evolvability score to guide candidate selection, exploration–exploitation, and mutation context. Across mathematical and systems optimization problems, Evade reduces redundant search and achieves the best mean and best solution on most tasks. Case studies demonstrate Evade's ability to escape from poor solutions and discover an algorithm that surpasses the human state of the art on a real-world cloud data transfer task. Our code is available for reproducibility.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, algorithmic discovery, scientific discovery, evolution, evolvability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QHt1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EVADE, an LLM-guided evolutionary framework that estimates whether a behavioral region can still generate improved or new offspring. The estimated evolvability controls search modes, parent selection, inspiration programs, and experience retention. Experiments on 11 tasks with two LLM backbones show strong performance.

Critical strengths
  1. Evolvability shows a meaningful alternative to selecting programs only by fitness.

  2. EVADE shows strong performance across 11 tasks.

Critical weaknesses
  1. The proposed signal is not convincingly validated as evolvability. Since the main experiments use scalar fitness as the behavior descriptor, it largely captures score transitions and recent mutation success.
  2. EVADE is considerably over-designed. Its performance depends on numerous manually specified modes, weights, clustering rules, lineage operations, complexity preferences, and memory parameters, while the limited ablations change multiple factors simultaneously.
  3. The experimental comparison is not fair or reproducible. LLM calls for experience distillation are excluded from the budget.
Questions for authors
  1. Does estimated evolvability predict future offspring improvement after controlling for parent fitness, visit count, mutation mode, and prompt context?

  2. Why is scalar fitness an adequate behavioral descriptor?

  3. Can you provide an ablation that changes only the evolvability term while keeping cells, prompts, parent complexity, inspiration programs, and memory rules fixed?

  4. What principles determined the search modes, weights, clustering threshold, smoothing parameters, and experience-retention rules, and were any tasks used to tune them?

  5. How many LLM calls, tokens, wall-clock hours, and monetary cost does each method use?

8 comments
AuthorsOpenReview ↗

Response to Reviewer QHt1 [Part 1/4]

We sincerely thank Reviewer QHt1 for the thoughtful review. We are glad that the reviewer found that evolvability "shows a meaningful alternative to selecting programs only by fitness" and that EVADE "shows strong performance across 11 tasks". Below we respond to the reviewer's comments.

[W1] The proposed signal is not convincingly validated as evolvability.

In this paper, we define evolvability as a candidate's capacity to generate offspring that either improve upon the best fitness found so far or exhibit behaviors distinct from their ancestors, following prior work that operationalizes evolvability through offspring outcomes (Wagner & Altenberg, 1996; Mengistu et al., 2016; Gajewski et al., 2019).

While this behavioral criterion can be viewed as capturing lineage-relative novelty, we treat it as evidence of evolvability because it reflects a parent's capacity to unlock previously inaccessible regions of the behavioral space. The intuition aligns with biological evolution: flight evolved independently in both birds and mammals (bats). Although flight was not globally novel when bats emerged, acquiring it from a non-flying mammalian ancestor required a lineage capable of generating viable, functionally distinct phenotypes. We verify empirically below that this estimate predicts future improvement, and isolating the effect of evolvability is presented in our answer to Q1 below.

[Q1] Does estimated evolvability predict future offspring improvement after controlling for parent fitness, visit count, mutation mode, and prompt context?

Thank you for this insightful question. To verify whether estimated evolvability predicts future improvement, we compare it with other selection signals on the same search states.

  • Search states. We take one EVADE run per task with DeepSeek-V4-Flash and freeze its archive at iterations t=25t = 25, 5050, and 7575. Each signal uses only information available up to tt.
  • Selection signals. At each iteration, each signal selects 3 cells: evolvability (et(Cb)e_t(C_b)); fitness (mean score of the cell); fitness + diversity, with equal weights on the normalized mean score and the code diversity of AdaEvolve (Cemri et al., 2026), i.e., the mean code distance of a program to its 5 nearest neighbors, where the distance combines token and structural-feature Jaccard distances and relative code length; and random. Within a selected cell, the parent is the program with the lowest cyclomatic complexity for every signal.
  • Evaluation. Each parent is given three mutations in two settings. In Parallel, the parent is mutated three times independently, and the setting succeeds if the best of the three children scores higher than the parent. In Sequential, the three mutations are applied one after another, each to the child of the previous one, and the setting succeeds if any program in the chain scores higher than the parent.

Mutation mode and prompt context are held fixed: all mutations use the same prompt, containing only the parent and its metrics, with no inspiration programs or experience. Parent fitness and visit count are not fixed, but selecting by fitness yields the lowest success rate, so the advantage of evolvability is not explained by choosing fitter parents.

Each entry is the number of successful parents, out of 3 per iteration and 9 in total.

Prism

Signalt = 25t = 50t = 75Total
Parallel
Evolvability2/32/32/36/9
Fitness0/30/31/31/9
Fitness + diversity0/31/32/33/9
Random1/31/31/33/9
Sequential
Evolvability2/32/32/36/9
Fitness1/31/30/32/9
Fitness + diversity2/32/32/36/9
Random2/31/31/34/9

Signal Processing

Signalt = 25t = 50t = 75Total
Parallel
Evolvability2/33/32/37/9
Fitness0/30/31/31/9
Fitness + diversity1/31/31/33/9
Random2/32/32/36/9
Sequential
Evolvability2/32/33/37/9
Fitness0/31/30/31/9
Fitness + diversity1/30/31/32/9
Random2/32/31/35/9

Evolvability attains the highest total in every case, tied only with fitness + diversity under Sequential on Prism. Pooled over tasks and settings, its selected parents succeed in 26 of 36 cases (72%), against 5/36 for fitness, 14/36 for fitness + diversity, and 18/36 for random selection. Its success rate does not decline at later iterations, when the global best is higher, so the estimate remains predictive as the search progresses.

AuthorsOpenReview ↗

Response to Reviewer QHt1 [Part 2/4]

[W1] Since the main experiments use scalar fitness as the behavior descriptor, it largely captures score transitions and recent mutation success.

[Q2] Why is scalar fitness an adequate behavioral descriptor?

A behavior descriptor for estimating evolvability should (i) separate programs that behave differently, (ii) be aligned with the objective, so that reaching a new behavior is a change that matters for the task, and (iii) be cheap to compute and available for every task. We use the combined score because it satisfies (ii) and (iii) on all 11 tasks without task-specific engineering (§4); it satisfies (i) only coarsely, which we acknowledge. The combined score is one choice of descriptor: the framework does not depend on it, and the descriptor can be replaced by a richer one without modifying the method (§4).

To illustrate, we compare two richer descriptors built from the sub-scores that make up the combined score (Prism: 1/max_kvpr on each test instance; Signal Processing: slope changes, lag error, average error, false reversals, correlation, and noise reduction). The sub-score vector of Table 5 uses the sub-scores directly. The two-level descriptor first groups programs by combined score as in the default method, and then splits a group with at least two programs by clustering on the sub-scores, with the same 25th-percentile threshold.

Descriptor (DeepSeek-V4-Flash, 3 runs)Prism ↑Signal Processing ↑
Combined score26.50 ± 0.160.713 ± 0.029
Sub-score vector (Table 5)26.33 ± 0.100.707 ± 0.075
Two-level26.57 ± 0.100.733 ± 0.018

All three descriptors give comparable results, and the two-level descriptor attains the highest mean fitness on both tasks.

[W2] EVADE is considerably over-designed. Its performance depends on numerous manually specified modes, weights, clustering rules, lineage operations, complexity preferences, and memory parameters,

EVADE centers on two main contributions: (1) behavioral-clustering-based evolvability, and (2) its integration into the general LLM-EC framework. Mode selection, parent and inspiration selection, and experience are standard components of existing LLM-EC frameworks such as AdaEvolve and ShinkaEvolve; EVADE does not add new components, but drives all of them with a single quantity, the evolvability estimate. In our initial ablation study, we replaced evolvability with alternative strategies in each of these components (parent selection, exploration–exploitation, and experience). Nevertheless, we agree that more ablation studies would better show the value of evolvability as the control signal; we present such new evidence below.

[W2] while the limited ablations change multiple factors simultaneously.

[Q3] Can you provide an ablation that changes only the evolvability term while keeping cells, prompts, parent complexity, inspiration programs, and memory rules fixed?

Thank you for this suggestion. In each variant, we replace or remove only the evolvability estimate, keeping cells, prompts, parent complexity, inspiration programs, and memory rules fixed (DeepSeek-V4-Flash, 3 runs):

  • Cell selection. Evolvability is removed from cell selection (Eq. 7): cells are selected uniformly, or by fitness only (we=0w_e = 0).
  • Signal. The signal rtr_t (Eq. 1) is replaced, with the cell estimator unchanged: global behavioral novelty, where a child is credited for a new behavior only if the behavior is new to the whole archive rather than to its parent's lineage; historical improvement, rt=1r_t = 1 if the child surpasses the best fitness and rt=δ~tr_t = \tilde\delta_t (Eq. 2) otherwise; and mutation success rate, rt=1[f(p′)>f(p)]r_t = \mathbb{1}[f(p') > f(p)].
VariantPrism ↑Signal Processing ↑
EVADE26.50 ± 0.160.713 ± 0.029
Cell selection
Uniform26.26 ± 0.000.670 ± 0.039
Fitness only26.26 ± 0.000.661 ± 0.035
Signal
Global behavioral novelty26.33 ± 0.100.656 ± 0.023
Historical improvement26.34 ± 0.070.698 ± 0.027
Mutation success rate26.28 ± 0.030.657 ± 0.009

Every variant scores lower than EVADE on both tasks. If the gains came from the surrounding design, these variants, which keep all of it, would retain them; they do not.

AuthorsOpenReview ↗

Response to Reviewer QHt1 [Part 3/4]

[Q4] What principles determined the search modes, weights, clustering threshold, smoothing parameters, and experience-retention rules, and were any tasks used to tune them?

The mode anchors are fixed and evenly spaced, and the mode weights are fixed across all tasks and backbones (Appendix A.2). The clustering threshold is not a fixed constant: it is set adaptively to the 25th percentile of pairwise distances (§3.5). The smoothing parameters ρ\rho and κ\kappa and the weight α\alpha were chosen from the sensitivity analysis in Table 5, run on Prism and Signal Processing, and the selected values were applied unchanged to all 11 tasks and both backbones.

We further report the sensitivity to the clustering percentile, with all other components fixed (DeepSeek-V4-Flash, 3 runs).

PercentilePrism ↑Signal Processing ↑
1026.32 ± 0.110.682 ± 0.007
1526.38 ± 0.140.673 ± 0.040
25 (default)26.50 ± 0.160.713 ± 0.029
4026.27 ± 0.020.703 ± 0.008
5026.27 ± 0.030.704 ± 0.034

The default percentile attains the highest mean on both tasks, and performance varies moderately across the other settings.

[W3] The experimental comparison is not fair or reproducible. LLM calls for experience distillation are excluded from the budget.

We follow the budget convention of prior work on this benchmark suite (Cemri et al., 2026; Liu et al., 2026a): all methods share a budget of T=100T = 100 iterations (§4), where each iteration consists of one generation call, which produces a candidate program, and one evaluation (§3.4). LLM calls made for other purposes, which we refer to as auxiliary calls, are not counted, for EVADE and the baselines alike. Such calls are common in existing frameworks: AdaEvolve uses them to generate meta-guidance when the search stagnates, EvoX to generate new search strategies when progress stalls, and ShinkaEvolve to query a novelty judge and generate meta-recommendations. GEPA, as implemented in SkyDiscover (Liu et al., 2026b), combines reflection and generation in a single prompt and makes no auxiliary calls.

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

This paper proposes EVADE, an LLM-based evolutionary search framework that introduces evolvability as an additional signal beyond candidate fitness and diversity. Evolvability is estimated from historical mutation outcomes, rewarding offspring that improve the global best fitness or reach behavioral regions not previously visited by their ancestral lineage. The estimated signal guides parent selection, exploration–exploitation strategies, and mutation experience filtering.

Experiments on 11 mathematical and systems optimization tasks demonstrate improvements over GEPA, ShinkaEvolve, AdaEvolve, and EvoX, with reduced behavioral redundancy and better optimization performance on most tasks.

Critical strengths

1. Interesting motivation and search perspective.

The paper identifies an important limitation of fitness- and diversity-based evolutionary search: high-quality candidates are not necessarily the most promising candidates for further evolution. Introducing evolvability as a search signal is a meaningful direction.

2. Coherent framework design.

EVADE integrates evolvability into multiple search decisions, including parent selection, adaptive mode selection, and mutation experience management. The signal is estimated from existing mutation outcomes without requiring additional fitness evaluations.

3. Promising empirical results.

The method demonstrates consistent improvements across multiple tasks and two backbone LLMs. The ablation studies and search trajectory analyses provide evidence for the effectiveness of the overall framework.

Critical weaknesses

1. Insufficient validation of evolvability as a predictor of future improvement

EVADE estimates evolvability from historical mutation outcomes, but past success does not necessarily indicate remaining improvement potential. As the global best fitness increases, previously productive regions may become exhausted while retaining relatively high evolvability estimates. Although discounted updates partially address this issue, the paper does not directly evaluate whether estimated evolvability predicts future mutation success or how quickly it adapts to changes in search potential. Such analysis is important given that evolvability is the central motivation of the proposed method.

2. Insufficient isolation of evolvability from other search mechanisms

EVADE introduces several design choices simultaneously, including behavioral clustering, complexity-aware parent selection, adaptive exploration, and experience filtering. While the component ablations demonstrate improvements over their respective replacements, they do not fully isolate the contribution of the evolvability signal itself.

In particular, behavioral clustering may independently improve search diversity and reduce redundant exploration. The parent-selection ablation also changes multiple selection criteria rather than replacing only the evolvability signal. Comparisons with uniform or fitness-based cell selection under the same clustering scheme, as well as simpler historical improvement or stagnation-based allocation strategies, would better establish the incremental benefit of evolvability.

3. Limited validation of behavioral representations and redundancy

The main experiments use the combined fitness score as the behavioral descriptor for clustering. However, candidates with similar fitness may implement fundamentally different algorithms, while candidates with different fitness may share similar underlying strategies. It is therefore unclear whether the resulting cells capture meaningful behavioral distinctions.

Furthermore, the behavioral redundancy metric is closely aligned with the proposed evolvability definition, making reductions in redundancy partly expected by construction. Although richer descriptors are evaluated on two tasks, more analysis is needed to establish the robustness of the approach to different behavioral representations and clustering granularities.

Questions for authors
  1. Predictive validity: How strongly does estimated evolvability correlate with subsequent improvement rates or the probability of discovering a new global best? How does this relationship change as the search progresses?

  2. Clustering contribution: What happens if the same behavioral clustering is retained, but cells are selected uniformly or using fitness/novelty alone, without evolvability estimation?

  3. Alternative search signals: How does evolvability-guided allocation compare with simpler historical mutation success rates, improvement statistics, or stagnation-aware exploration strategies?

  4. Experience management: Would fixed-size, recency-based experience filtering achieve similar improvements to evolvability-guided filtering under a comparable context budget?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zBct [Part 1/4]

We sincerely thank Reviewer zBct for the thorough and constructive review. We are glad that the reviewer found introducing evolvability as a search signal "a meaningful direction", our framework design "coherent", and our empirical results "promising", with "consistent improvements across multiple tasks and two backbone LLMs". Below we respond to the reviewer's comments.

[W1] EVADE estimates evolvability from historical mutation outcomes, but past success does not necessarily indicate remaining improvement potential. As the global best fitness increases, previously productive regions may become exhausted while retaining relatively high evolvability estimates. Although discounted updates partially address this issue, the paper does not directly evaluate whether estimated evolvability predicts future mutation success or how quickly it adapts to changes in search potential. Such analysis is important given that evolvability is the central motivation of the proposed method.

[Q1] Predictive validity: How strongly does estimated evolvability correlate with subsequent improvement rates or the probability of discovering a new global best? How does this relationship change as the search progresses?

Thank you for this insightful question, which concerns the central motivation of our method. To further verify whether estimated evolvability predicts future improvement, we compare it with other selection signals on the same search states.

  • Search states. We take one EVADE run per task with DeepSeek-V4-Flash and freeze its archive at iterations t=25t = 25, 5050, and 7575. Each signal uses only information available up to tt.
  • Selection signals. At each iteration, each signal selects 3 cells: evolvability (et(Cb)e_t(C_b)); fitness (mean score of the cell); fitness + diversity, with equal weights on the normalized mean score and the code diversity of AdaEvolve (Cemri et al., 2026), i.e., the mean code distance of a program to its 5 nearest neighbors, where the distance combines token and structural-feature Jaccard distances and relative code length; and random. Within a selected cell, the parent is the program with the lowest cyclomatic complexity for every signal.
  • Evaluation. Each parent is given three mutations in two settings. In Parallel, the parent is mutated three times independently, and the setting succeeds if the best of the three children scores higher than the parent. In Sequential, the three mutations are applied one after another, each to the child of the previous one, and the setting succeeds if any program in the chain scores higher than the parent. All mutations use the same prompt, containing only the parent and its metrics, so that the outcome depends only on the selected parent.

Each entry is the number of successful parents, out of 3 per iteration and 9 in total.

Prism

Signalt = 25t = 50t = 75Total
Parallel
Evolvability2/32/32/36/9
Fitness0/30/31/31/9
Fitness + diversity0/31/32/33/9
Random1/31/31/33/9
Sequential
Evolvability2/32/32/36/9
Fitness1/31/30/32/9
Fitness + diversity2/32/32/36/9
Random2/31/31/34/9

Signal Processing

Signalt = 25t = 50t = 75Total
Parallel
Evolvability2/33/32/37/9
Fitness0/30/31/31/9
Fitness + diversity1/31/31/33/9
Random2/32/32/36/9
Sequential
Evolvability2/32/33/37/9
Fitness0/31/30/31/9
Fitness + diversity1/30/31/32/9
Random2/32/31/35/9

Evolvability attains the highest total in every case, tied only with fitness + diversity under Sequential on Prism. Pooled over tasks and settings, its selected parents succeed in 26 of 36 cases (72%), against 5/36 for fitness, 14/36 for fitness + diversity, and 18/36 for random selection.

AuthorsOpenReview ↗

Response to Reviewer zBct [Part 2/4]

[W2] EVADE introduces several design choices simultaneously, including behavioral clustering, complexity-aware parent selection, adaptive exploration, and experience filtering. While the component ablations demonstrate improvements over their respective replacements, they do not fully isolate the contribution of the evolvability signal itself.

EVADE centers on two main contributions: (1) behavioral-clustering-based evolvability, and (2) its integration into the general LLM-EC framework. In our initial ablation study, we replaced evolvability with alternative strategies to isolate its impact on core LLM-EC components (parent selection, exploration–exploitation, and experience). However, we agree that a more granular analysis would strengthen these findings, which we provide below. Each experiment changes a single element while keeping the behavioral clustering and all other components fixed; all runs use DeepSeek-V4-Flash (3 runs per variant).

[W2] In particular, behavioral clustering may independently improve search diversity and reduce redundant exploration. The parent-selection ablation also changes multiple selection criteria rather than replacing only the evolvability signal.

[Q2] Clustering contribution: What happens if the same behavioral clustering is retained, but cells are selected uniformly or using fitness/novelty alone, without evolvability estimation?

We keep the same clustering and change only how cells are selected:

  • Uniform. Cells are selected uniformly at random.
  • Fitness only. Cells are selected by their mean fitness alone (we=0w_e = 0 in Eq. 7).
  • Global behavioral novelty. A child is credited for reaching a new behavior only if the behavior is new to the whole archive, rather than new to its parent's lineage.
Cell selectionPrism ↑Signal Processing ↑
Evolvability (EVADE)26.50 ± 0.160.713 ± 0.029
Uniform26.26 ± 0.000.670 ± 0.039
Fitness only26.26 ± 0.000.661 ± 0.035
Global behavioral novelty26.33 ± 0.100.656 ± 0.023

All three variants score lower on both tasks; on Prism, every run of uniform and fitness-only selection stalls at 26.26. The clustering alone therefore does not account for the gains.

[W2] Comparisons with […] simpler historical improvement or stagnation-based allocation strategies would better establish the incremental benefit of evolvability.

[Q3] Alternative search signals: How does evolvability-guided allocation compare with simpler historical mutation success rates, improvement statistics, or stagnation-aware exploration strategies?

Thank you for suggesting these baselines.

  • Mutation success rate. rt=1[f(p′)>f(p)]r_t = \mathbb{1}[f(p') > f(p)], with the cell estimator unchanged.
  • Historical improvement. rt=1r_t = 1 if the child surpasses the best fitness, and rt=δ~tr_t = \tilde\delta_t (Eq. 2) otherwise, with the cell estimator unchanged.
  • Stagnation-aware. Mode selection is replaced by that of AdaEvolve (Cemri et al., 2026): modes follow its exploration intensity, and breakthrough is triggered when the search stagnates.
SignalPrism ↑Signal Processing ↑
Evolvability (EVADE)26.50 ± 0.160.713 ± 0.029
Mutation success rate26.28 ± 0.030.657 ± 0.009
Historical improvement26.34 ± 0.070.698 ± 0.027
Stagnation-aware26.32 ± 0.330.636 ± 0.040

All three alternatives score lower than EVADE on both tasks.

[Q4] Experience management: Would fixed-size, recency-based experience filtering achieve similar improvements to evolvability-guided filtering under a comparable context budget?

We replace evolvability-guided retention (Eq. 8) by keeping only the KK most recent entries, with K=3K = 3 for experience and K=1K = 1 for strategies, the mean numbers of entries EVADE includes in each prompt, so the context budget is matched.

ExperiencePrism ↑Signal Processing ↑
Evolvability-guided (EVADE)26.50 ± 0.160.713 ± 0.029
Most recent entries26.43 ± 0.170.676 ± 0.043

Keeping the most recent entries under a matched budget is close on Prism and lower on Signal Processing.

AuthorsOpenReview ↗

Response to Reviewer zBct [Part 3/4]

[W3] The main experiments use the combined fitness score as the behavioral descriptor for clustering. However, candidates with similar fitness may implement fundamentally different algorithms, while candidates with different fitness may share similar underlying strategies. It is therefore unclear whether the resulting cells capture meaningful behavioral distinctions.

A practical behavior descriptor should (i) separate distinct behaviors, (ii) align with the task objective, and (iii) be cheap to compute across tasks. The combined score satisfies (ii) and (iii) on all 11 tasks without task-specific engineering (§4), while capturing (i) coarsely. Note that we expect "behavior distinction", rather than different code/algorithms. Grouping algorithms with different mechanisms into the same cell can be advantageous: each cell captures different lineages reaching similar behavior, preserving a diverse gene pool for mutation/crossover. Specifically, selecting a parent alongside inspirations from the same cell effectively cross-breeds different algorithmic families to potentially generate novel solutions. This cross-family recombination is only possible because cells link phenotypes with diverse underlying genotypes. A cell is sampled based on its evolvability, thus encouraging meaningful development as well.

It should also be noted that code difference is not an effective criterion in practice. ShinkaEvolve rejects programs whose code is too similar to previous ones and scores below EVADE (26.20 vs. 26.50 on Prism, 0.659 vs. 0.713 on Signal Processing; Tables 1–2). To test this further, we form cells by code difference instead of behavior, using the code-difference measure of AdaEvolve as implemented in SkyDiscover (Cemri et al., 2026; Liu et al., 2026b); all other components are unchanged (DeepSeek-V4-Flash, 3 runs).

CellsPrism ↑Signal Processing ↑
Behavior (EVADE)26.50 ± 0.160.713 ± 0.029
Code difference26.30 ± 0.090.645 ± 0.082

Separating programs by code lowers fitness on both tasks.

[W3] Furthermore, the behavioral redundancy metric is closely aligned with the proposed evolvability definition, making reductions in redundancy partly expected by construction.

Although BRR is related to evolvability, it is not guaranteed by EVADE's design. EVADE uses evolvability, estimated from the outcomes of past mutations, to guide cell selection, balance exploration and exploitation, and determine which experiences to include in the prompt for subsequent mutations. Once a mutation is executed, its actual outcome is used to update BRR, which measures the realized effectiveness of the mutation. In other words, evolvability is estimated before mutations without simulation/future look-ahead, while BRR is measured after the mutations. For a fair comparison, BRR is measured across all methods using the same set of cells; for baseline methods, these cells are only observed and do not influence the search process (§4). We will update the writing for clarity.

Reviewer_ByFoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes EVADE, an LLM-driven evolutionary search framework that uses evolvability as an additional signal for algorithm discovery. Rather than selecting candidates solely based on their current fitness or diversity, EVADE estimates evolvability from offspring outcomes. An offspring receives a maximal signal if it surpasses the global best; otherwise, its signal depends on whether it reaches a behavioral cell outside its parent's ancestral lineage and on the estimated evolvability of that cell and its fitness improvement. The resulting signal guides exploration–exploitation mode selection, parent selection, and the retention of mutation experience. Experimental results show that EVADE achieves competitive or superior performance on most tasks, together with lower behavioral redundancy on the reported benchmarks. Ablation studies and case studies are provided to investigate the contributions of the individual components and illustrate how EVADE can escape stagnation and discover improved algorithms.

Critical strengths
  1. The paper identifies a meaningful limitation of LLM-driven evolutionary search: a candidate's current fitness or diversity does not necessarily indicate whether further mutations from that candidate will be productive. Using outcomes from previous mutations to guide subsequent search is a reasonable and potentially valuable direction.
  2. The evaluation covers 11 optimization tasks and two LLM backbones, and includes component ablations and case studies of search trajectories. The reported performance improvements on multiple tasks, together with examples of escaping stagnation, provide encouraging evidence that the overall framework can be effective.
Critical weaknesses
  1. The operational definition of evolvability is not sufficiently validated. The proposed signal rewards an offspring for reaching a behavioral cell outside its parent's ancestral lineage. However, such a cell may already have been explored by other lineages, so the signal does not necessarily measure globally novel behavior. More importantly, most experiments use a combined fitness score as the behavior descriptor. Algorithms with substantially different mechanisms may therefore be grouped together if their scores are similar, while small score changes may create apparently novel behaviors. This raises concerns about whether the signal captures meaningful algorithmic novelty and whether the reported behavioral redundancy rate measures the functional redundancy claimed by the paper. The limited evaluation of richer behavior descriptors does not resolve this concern.
  2. The experiments do not sufficiently isolate the contribution of evolvability or establish cost fairness. The experiments do not directly compare the proposed signal against simpler alternatives, such as historical fitness improvement, mutation success rate, or global behavioral novelty. Since EVADE rewards offspring that reach behavioral cells outside their parents' ancestral lineages and uses this signal to guide exploration, the observed gains may stem primarily from lineage-aware behavioral exploration rather than from evolvability as a distinct signal of future search potential. The current ablations do not disentangle these explanations. Moreover, EVADE periodically uses additional LLM calls to summarize mutation experience, but the paper does not clearly report whether these calls and their associated token costs are included in the budget comparison. Equal numbers of evolutionary iterations do not necessarily imply equal total LLM costs. These issues weaken the evidence for the paper's central claim that evolvability itself accounts for the performance gains.
Questions for authors
  1. Clarification and evidence concerning the evolvability signal. The signal is positive when an offspring reaches a cell outside its parent's ancestral lineage, even if that cell has already been visited by other lineages. What is the rationale for treating this event as evidence of evolvability rather than lineage-relative novelty? Can the authors provide a direct comparison with global behavioral novelty and historical mutation-improvement signals? Such comparisons would help establish the specific value of the proposed definition.
  2. Clarification of behavioral descriptors and redundancy measurement. The main experiments primarily use the combined fitness score as the behavior descriptor. How sensitive are the reported performance gains and behavioral redundancy rates to the choice of descriptor? In particular, can the authors provide evidence that the reduction in redundancy persists when behavior descriptors distinguish algorithms with different behaviors but similar fitness values? Furthermore, Table 5 reports lower mean performance with the vector descriptor than with the combined score on both Prism and Signal Processing. Given that the vector descriptor captures richer behavioral information, why does it not yield a clear improvement? Could the authors clarify whether this is due to the descriptor design, distance metric, or other factors, and whether richer descriptors provide measurable benefits beyond final fitness?
  3. Clarification of the computational budget. Are the additional LLM calls used for periodic experience summarization included in the reported budget of 100 iterations? Can the authors report the total LLM calls and token consumption for EVADE and the baselines, and clarify whether the performance advantage persists under a matched total LLM-call or token budget? Equal numbers of evolutionary iterations do not necessarily imply equal computational costs.
6 comments
AuthorsOpenReview ↗

Response to Reviewer ByFo [Part 1/5]

We sincerely thank Reviewer ByFo for the thorough evaluation of our manuscript. We are glad that the reviewer found that our paper "identifies a meaningful limitation of LLM-driven evolutionary search", considered using outcomes from previous mutations to guide subsequent search "a reasonable and potentially valuable direction", and found that our results "provide encouraging evidence that the overall framework can be effective". Below we respond to the reviewer's comments.

[W1] The operational definition of evolvability is not sufficiently validated.

[Q1] What is the rationale for treating this event as evidence of evolvability rather than lineage-relative novelty?

Thank you for this important question. In this paper, we define evolvability as a candidate's capacity to generate offspring that either improve upon the best fitness found so far or exhibit behaviors distinct from their ancestors, following prior work that operationalizes evolvability through offspring outcomes (Wagner & Altenberg, 1996; Mengistu et al., 2016; Gajewski et al., 2019).

While this behavioral criterion can be viewed as capturing lineage-relative novelty, we treat it as evidence of evolvability because it reflects a parent's capacity to unlock previously inaccessible regions of the behavioral space. The intuition aligns with biological evolution: flight evolved independently in both birds and mammals (bats). Although flight was not globally novel when bats emerged, acquiring it from a non-flying mammalian ancestor required a lineage capable of generating viable, functionally distinct phenotypes. Similarly, in our framework, producing offspring with behaviors novel to their specific lineage demonstrates a candidate's capacity to drive meaningful evolutionary exploration. We will revise our paper to make this point clear, thank you.

[W1] The proposed signal rewards an offspring for reaching a behavioral cell outside its parent's ancestral lineage. However, such a cell may already have been explored by other lineages, so the signal does not necessarily measure globally novel behavior.

[Q1] The signal is positive when an offspring reaches a cell outside its parent's ancestral lineage, even if that cell has already been visited by other lineages.

When an offspring of lineage AA reaches a new behavioral cell, it demonstrates the evolvability of lineage AA, regardless of whether lineage BB reached that cell first. This mirrors how acquiring flight demonstrates the evolutionary capacity of bats, even though birds evolved flight earlier. Ultimately, our definition evaluates evolvability by linking genotype (lineage) to phenotype (behavior), rather than judging candidates solely on global phenotypic novelty.

AuthorsOpenReview ↗

Response to Reviewer ByFo [Part 2/5]

[W2] The experiments do not directly compare the proposed signal against simpler alternatives, such as historical fitness improvement, mutation success rate, or global behavioral novelty. Since EVADE rewards offspring that reach behavioral cells outside their parents' ancestral lineages and uses this signal to guide exploration, the observed gains may stem primarily from lineage-aware behavioral exploration rather than from evolvability as a distinct signal of future search potential. The current ablations do not disentangle these explanations.

[Q1] Can the authors provide a direct comparison with global behavioral novelty and historical mutation-improvement signals? Such comparisons would help establish the specific value of the proposed definition.

Thank you for this helpful suggestion. We replace only the signal rtr_t in Eq. 1, keeping the cell estimator and all other components unchanged; all runs use DeepSeek-V4-Flash.

  • Global behavioral novelty. A child is credited for reaching a new behavior only if the behavior is new to the whole archive, rather than new to its parent's lineage. A child that surpasses the best fitness still receives full credit.
  • Historical improvement. The parent is credited by how much its child improves on it, regardless of the behavior the child reaches: rt=1r_t = 1 if the child surpasses the best fitness, and rt=δ~tr_t = \tilde\delta_t (Eq. 2) otherwise.
  • Mutation success rate. The parent is credited whenever its child improves on it: rt=1[f(p′)>f(p)]r_t = \mathbb{1}[f(p') > f(p)].
Signal (3 runs)Prism ↑Signal Processing ↑
EVADE26.50 ± 0.160.713 ± 0.029
Global behavioral novelty26.33 ± 0.100.656 ± 0.023
Historical improvement26.34 ± 0.070.698 ± 0.027
Mutation success rate26.28 ± 0.030.657 ± 0.009

EVADE attains the highest mean fitness on both tasks. Each variant removes one part of our definition: historical improvement and mutation success rate drop the requirement of new behavior, and global behavioral novelty drops the lineage reference. All three score lower than EVADE.

[W1] More importantly, most experiments use a combined fitness score as the behavior descriptor. Algorithms with substantially different mechanisms may therefore be grouped together if their scores are similar, while small score changes may create apparently novel behaviors.

A practical behavior descriptor should (i) separate distinct behaviors, (ii) align with the task objective, and (iii) be cheap to compute across tasks. The combined score satisfies (ii) and (iii) on all 11 tasks without task-specific engineering (§4), while capturing (i) coarsely. Note that we expect "behavior distinction", rather than "mechanism distinction" (given that mechanism refers to genotype/algorithms). Grouping algorithms with different mechanisms into the same cell can be advantageous: each cell captures different lineages reaching similar behavior, preserving a diverse gene pool for mutation/crossover.

In addition, it should be noted that small score changes do not create new behaviors: a new program joins the nearest cell unless its distance exceeds ϵ\epsilon, which is set adaptively to the 25th percentile of pairwise distances (§3.5), so a new cell requires a change larger than the typical gap between programs.

AuthorsOpenReview ↗

Response to Reviewer ByFo [Part 3/5]

[W1] This raises concerns about whether the signal captures meaningful algorithmic novelty.

Our framework encourages meaningful algorithmic novelty. As noted above, each cell groups diverse genotypes that achieve a similar phenotype; consequently, selecting a parent alongside inspirations from the same cell effectively cross-breeds different algorithmic families to potentially generate novel solutions. This cross-family recombination is only possible because cells link phenotypes with diverse underlying genotypes. A cell is sampled based on its evolvability, thus encouraging meaningful development as well.

It should be noted that judging novelty based solely on genotype (code difference) is not always effective: ShinkaEvolve rejects programs whose code is too similar to previous ones and scores below EVADE (26.20 vs. 26.50 on Prism, 0.659 vs. 0.713 on Signal Processing; Tables 1–2). To test this further, we form cells by code difference instead of behavior, using the code-difference measure of AdaEvolve as implemented in SkyDiscover (Cemri et al., 2026; Liu et al., 2026b), a weighted combination of token and structural-feature Jaccard distances and relative code length; all other components are unchanged (DeepSeek-V4-Flash, 3 runs).

Cells (3 runs)Prism ↑Signal Processing ↑
Behavior (EVADE)26.50 ± 0.160.713 ± 0.029
Code difference26.30 ± 0.090.645 ± 0.082

Separating programs by code lowers fitness on both tasks.

[W1] […] and whether the reported behavioral redundancy rate measures the functional redundancy claimed by the paper.

[Q2] In particular, can the authors provide evidence that the reduction in redundancy persists when behavior descriptors distinguish algorithms with different behaviors but similar fitness values?

A mutation is redundant if its child p′p' lands in a cell already visited along the parent's ancestral lineage without improvement. Like our evolvability metric, this definition links genotype to phenotype: redundancy captures the stalled development of a single family—that is, revisiting behavioral states that the lineage has already explored without making progress. Although BRR is related to evolvability, it is not guaranteed by EVADE's design: evolvability is estimated before mutations without simulation or future look-ahead, while BRR is measured after the mutations.

To check that the reduction in redundancy does not depend on the score descriptor, we re-measure BRR on the existing runs of EVADE and the baselines (DeepSeek-V4-Flash, 3 runs per method) without rerunning any search. The programs, their parent pointers, and their fitness values are kept as generated; only the cells used to measure BRR change. We form these cells from three descriptors:

  • Combined score: the descriptor used in the paper (§4).
  • Sub-score vector: the scores that make up the combined score, as in Table 5 (Prism: 1/max_kvpr on each test instance; Signal Processing: slope changes, lag error, average error, false reversals, correlation, and noise reduction).
  • Two-level: programs are first grouped by combined score, and a group with at least two programs is then split by clustering on its sub-scores, with the same 25th-percentile threshold.

The results (BRR, %) are shown below.

Prism

MethodCombined scoreSub-score vectorTwo-level
EVADE5.3 ± 3.812.9 ± 4.23.8 ± 2.3
AdaEvolve55.1 ± 4.230.8 ± 5.824.5 ± 4.7
EvoX44.4 ± 4.339.9 ± 2.119.1 ± 2.9
ShinkaEvolve41.7 ± 2.536.4 ± 2.727.2 ± 5.3
GEPA53.7 ± 2.145.2 ± 8.339.1 ± 6.2

Signal Processing

MethodCombined scoreSub-score vectorTwo-level
EVADE8.3 ± 2.114.3 ± 1.41.8 ± 1.6
AdaEvolve60.8 ± 8.235.4 ± 7.330.7 ± 4.9
EvoX37.2 ± 12.431.8 ± 5.524.1 ± 6.3
ShinkaEvolve37.3 ± 1.223.5 ± 6.018.3 ± 6.4
GEPA68.7 ± 3.563.8 ± 2.355.2 ± 5.6

EVADE has the lowest BRR under all three descriptors on both tasks.

Comments on the paper

1 comment

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.