acceptodds
Under review as a conference paper at ICLR 2027

BeNav: Behavior-Space Navigation for Automatic Heuristic Design

Abstract

Large language models (LLMs) have recently enabled automatic heuristic design by generating and evolving optimization algorithms. However, existing methods mainly search directly in the algorithm space, leaving the direction of exploration largely implicit. This often leads to inefficient search and premature stagnation. In this work, we propose BeNav: a behavior-space navigation approach for automatic heuristic design. We represent each algorithm in an interpretable behavior space characterized by search style, information usage, uncertainty, and adaptivity. The behavior space is then used to explicitly navigate algorithm exploration: behavioral composition guides the generation of algorithms toward intermediate behaviors between two parent behaviors, while behavioral diversion drives the search toward under-explored behavioral regions. These mechanisms are integrated with local exploitation of the current best-performing algorithm to balance exploitation and exploration. Extensive experiments demonstrate that BeNav enables more effective search and consistently outperforms state-of-the-art LLM-driven automatic heuristic design methods.

Primary area:
optimization
Keywords:
Large Language Model Driven Automatic Heuristic Design, Automatic Heuristic Design, Evolutionary Search, Behavioral Representation, Behavior-Guided Search

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tSnWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BeNav, a behavior-space navigation framework for LLM-driven automatic heuristic design. BeNav maps each candidate algorithm to a 4D interpretable behavioral coordinate, covering search style, information usage, uncertainty, and adaptivity, and uses this coordinate to steer algorithm generation via three mechanisms: behavioral composition, behavioral diversion, and local exploitation. Experiments across four AHD settings show consistent improvements over recent LLM-AHD baselines, with ablation studies supporting the contributions of the proposed components.

Critical strengths

The paper presents a clear, timely, and intuitive perspective on algorithm search. Moving from treating algorithm behavior as a descriptive property to explicitly using it as a search signal is an interesting and potentially useful idea.

Critical weaknesses
  1. The paper does not sufficiently establish that the proposed behavior space is actually controllable for algorithm generation. The key premise of BeNav is that a behavioral representation can serve not only as a description of an existing algorithm, but also as a target for generating new algorithms. However, the paper mainly demonstrates the forward mapping from an algorithm to its behavioral coordinate. The reverse direction is much less clear: given a target behavior, can the LLM reliably generate an algorithm that exhibits that behavior? For example, Behavioral Composition defines the offspring behavior as an intended interpolation between two parents, but the paper does not directly measure how closely the generated offspring reaches the intended behavioral target. The reported performance improvements and behavioral clustering results therefore show that BeNav is effective, but do not clearly establish that the proposed behavior space is controllable or that the intended behavioral transitions actually occur.

  2. The experimental comparison does not fully resolve search-effort fairness, making the role of behavioral navigation difficult to isolate. All methods are constrained by the same 500 function evaluations, but BeNav additionally performs behavior extraction and behavioral navigation operations. The paper does not clearly report the number of generated candidate algorithms, LLM generation calls, behavior-extraction calls, reflection calls, or total LLM token/computational cost under this budget. Consequently, it is unclear whether BeNav achieves better results through more effective navigation or partly through a different amount of algorithm-generation/search effort. Reporting these quantities would be important for establishing a fair comparison.

  3. The paper's distinction from selection-based search is not entirely convincing. The paper criticizes existing approaches for primarily using performance or other signals to select promising parents, arguing that this only indirectly determines the direction of search. However, Behavioral Composition in BeNav also first selects two parents based on fitness and then uses their behavioral representations to guide offspring generation. Thus, the proposed method still relies on parent selection as an important mechanism for determining the search direction. It would be useful to clarify what fundamentally distinguishes this process from existing selection-guided evolutionary search, beyond additionally providing behavioral information to the LLM.

Questions for authors
  1. Can the authors provide direct evidence that a specified behavioral target leads to a generated algorithm whose measured behavior moves toward that target? More importantly, what is the fundamental advantage of using a behavioral representation as a target over simply using performance as the target when influencing the algorithm search space? Does the behavioral representation actually provide stronger or more controllable guidance for generation, and can this improvement in controllability be quantitatively measured?

  2. Under the 500-FE budget, how many candidate algorithms and LLM calls are made by BeNav versus each baseline? In particular, how many calls are used for behavior extraction, reflection, and algorithm generation, and what is the corresponding total token/computational cost?

2 comments
AuthorsOpenReview ↗

Response: Navigation Evidence, Search Effort, and Parent Selection

The reviewer's thoughtful assessment and constructive suggestions are greatly appreciated. In response to the concerns about controllability, search effort, and parent selection, we have clarified the method and provided additional empirical evidence.

1. Direct evidence of behavioral guidance and its advantage over a performance target

Appendix B.1, Table 7 reports a supplementary CVRP-ACO run recording operator inputs and re-extracted offspring coordinates. Its Diversion trigger threshold is set to zero to increase observations specifically for this diagnostic run.

Among successfully executed offspring with evaluable coordinates, explicit post-hoc criteria yield:

OperatorCriterionAdherent / validRate
CompositionParental coordinate intervals expanded by 0.10 per side103/14670.5%
Local ExploitationMaximum absolute coordinate change at most 0.2055/7474.3%
DiversionPositive displacement along the target direction36/5170.6%

For Diversion, the criterion is (b_child - b_ref) · (b_target - b_ref) > 0, where the dot denotes the vector inner product. The case study supports a directional tendency and directly measures behavioral combination, local preservation, and directional guidance, complementing performance and clustering results. These measurements characterize adherence to behavioral regions and directions rather than exact target-coordinate matching.

Appendix B.1, Table 5 also reports ICC(2,1) values of 0.765-0.910 across 160 algorithms with five extractions each. This supports extraction repeatability under the evaluated prompt and model configuration. Table 6 connects coordinates to execution measurements for 500 CVRP-ACO and 495 TSP-GLS algorithms; UN concordance reaches 73.9% and 74.8% at minimum coordinate separation 0.4.

Performance specifies how candidate quality is evaluated. Behavioral guidance additionally specifies characteristics of the modification to pursue: combining existing mechanisms, preserving a local behavioral pattern, or exploring another behavioral region. BeNav uses performance to guide quality-based selection and behavioral information to condition generation.

To examine numerical-coordinate contributions, Section 4.3 and Table 4 add BeNav-Random. It retains strategy extraction and semantic descriptions but replaces coordinates used for storage and downstream navigation with uniformly sampled vectors in the four-dimensional unit cube, fixed for each distinct implementation within a run.

Three-run MKP-ACO objectives (mean +/- standard deviation) are:

Methodn=100 (higher is better)n=300n=500
BeNav23.229 +/- 0.0262.067 +/- 0.20103.082 +/- 0.48
BeNav-Random23.107 +/- 0.1159.282 +/- 3.6595.922 +/- 9.71

Random replacement lowers mean performance at all three sizes and substantially increases run-to-run variability. This control supports the usefulness of numerical behavioral information within the generation pipeline, while the offspring analysis measures adherence to the intended operations.

2. Search-effort fairness and LLM overhead

The shared backbone and nominal 500-FE budget in Section 4.1 control model choice and algorithm-evaluation effort. We agree that equal evaluation budgets do not imply equal total inference cost. Generated candidates, generation requests, extraction requests, reflection requests, and token or timing costs are distinct quantities.

BeNav introduces additional overhead through behavioral extraction and behavioral context in generation and reflection prompts. Each extraction request jointly produces the semantic description and all four coordinates; coordinate operations, retrieval, and pruning are local computations. We have not completed a systematic inference-cost comparison covering all baselines. Our performance claims concern solution quality under a common algorithm-evaluation budget, rather than superiority under equal token or wall-clock budgets. The random-coordinate experiment provides a complementary control for coordinate information while retaining extraction and semantic descriptions.

3. Distinction from selection-based search

We agree that parent selection remains important. As described in Section 3.3, Composition provides parent implementations and behavioral representations to guide a transition between their behavioral patterns. Diversion constructs a target behavioral region and provides a feasible reference implementation. Selection identifies starting points; behavioral navigation adds a specification for subsequent generation. These operations are complementary, and the contribution lies in the additional behavioral guidance.

I believe the central controllability concern is only partially addressed. The reported adherence criteria measure relatively coarse behavioral constraints rather than how closely offspring reach their intended targets. For example, satisfying the expanded coordinate intervals in Composition does not necessarily indicate proximity to the intended interpolated behavior, while a positive directional projection in Diversion does not establish meaningful progress toward the target. These results support a tendency toward the intended behavioral regions or directions, but do not yet establish reliable target-conditioned generation.

A more direct experiment would be to take two parent algorithms with behavioral vectors A and B, specify an expected offspring behavior C (e.g., an interpolated point between A and B), and measure the deviation between the generated offspring's actual behavioral vector and this target. Repeating this experiment across multiple parent pairs (e.g., 10 pairs), with multiple offspring generated for each pair, and reporting the average deviation would provide a more direct and intuitive assessment of the controllability of the mapping from behavior space to algorithm space.

Regarding search-effort fairness, I appreciate the authors' transparency. Nevertheless, equal FE budgets only establish fairness in algorithm-evaluation effort, not in total LLM inference resources. Without the corresponding generation, extraction, and reflection call counts or token costs, the resource efficiency of BeNav remains unclear.

My overall assessment remains unchanged, and I maintain my original rating.

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

BeNav introduces behaviour-space navigation for large language model (LLM)-based automatic heuristic design (AHD). Rather than implicitly searching in code space, it maps candidate algorithms into a 4D interpretable space: search style, information use, uncertainty, and adaptivity. The space explicitly steers generation via three complementary mechanisms: behavioural composition (interpolating behaviours between two parents), behavioural diversion (expanding from elite centroids into under-explored sparse regions during stagnation), and local exploitation (refining the best-performing algorithm). Across several NP-hard combinatorial optimization benchmarks, BeNav consistently outperforms state-of-the-art LLM-AHD baselines.

Critical strengths

BeNav addresses the problem of search becoming trapped in redundant implementations that share identical behavioural patterns. By defining a 4D semantic coordinate system based on Decision (information usage, adaptivity) and Movement (search style, uncertainty), it converts behaviour from a passive analytical metric into an active coordinate system for explicit search steering. The interplay among behavioural composition, diversion, and local exploitation systematically balances exploration and exploitation, decoupling where to search from how to implement the algorithm. Crucially, the behaviour coordinate is extracted directly from the code prior to execution, avoiding costly trajectory rollouts.

Critical weaknesses

The framework relies on subjective, uncalibrated LLM self-evaluation for behaviour extraction. Relying on prompt instructions to assign normalized continuous coordinates in [0,1][0, 1] directly from static code introduces prompt sensitivity and potential hallucinations, since static analysis cannot reliably assess runtime stochasticity, iteration-dependent adaptivity, or dynamic search style without empirical trajectory execution. Furthermore, while continuous vector operations (e.g., Euclidean centroids, directional unit vectors, and interpolation) are computed geometrically, they must be translated back into qualitative natural-language prompts for code generation, making the actual adherence of generated offspring to the target coordinates uncertain. The empirical evaluation is also limited: it uses a small population size of 10, a budget of only 500 evaluations, and averages results over just three random seeds, leaving the statistical significance of narrow margins open to question. Finally, it remains unclear why behavioural diversity improves the LLM-based AHD (see Question 6 for the authors below).

Questions for authors
  1. When the LLM generates an offspring guided by an interpolated target or a diversion vector BtargetB_{\text{target}}, what is the empirical Euclidean distance between the requested target coordinate and the re-extracted coordinate of the resulting code? How frequently does the model fail to shift in the intended behavioural direction?
  2. Heuristic behaviours like adaptivity and uncertainty often manifest dynamically at runtime depending on the problem instance. Did you validate the LLM's static [0,1][0, 1] ratings against empirical search trajectory metrics (e.g., measuring actual step sizes or entropy across iterations), and how consistent are the ratings across repeated inferences?
  3. How is search "stagnation" formally defined and detected during the evolutionary loop to trigger Behavioural Diversion? Over a typical 500-evaluation run, what proportion of candidate generations is allocated to composition, diversion, and local exploitation, respectively?
  4. The experimental protocol fixes the population size to 10 and the evaluation budget to 500 fitness evaluations. Does the behavioural navigation mechanism scale effectively to larger populations or extended budgets (e.g., 2,000+ evaluations) without suffering from behavioural crowding or drift into unfeasible heuristic spaces?
  5. In the dimension ablation (Table 4 and Table 5), the 3D variant without Adaptivity (-AD) outperforms the full 4D model on large MKP instances (n=500n=500: 103.118 vs. 103.082). Does this indicate that certain behavioural dimensions become redundant or even introduce conflicting guidance when scaling to larger problem instances?
  6. The lack of grounded practical meaning behind the behavioural patterns represents a critical methodological weakness because it severs the causal link between the proposed coordinate space and actual algorithmic performance. Unlike established Quality Diversity paradigms where behavioural descriptors reflect tangible execution dynamics (such as trajectory entropy or operator acceptance rates), the framework relies on subjective, zero-shot LLM ratings inferred statically from raw source code. This creates a severe construct-validity gap: an algorithm labelled with high Uncertainty or Adaptivity may behave completely deterministically or rigidly under actual runtime evaluation. Without demonstrating that these abstract 4D coordinates correlate with physical problem-solving behaviours, it remains ambiguous whether diversifying across this artificial coordinate system truly overcomes algorithmic plateaus or simply mimics the stochastic perturbation effect of randomized prompt jittering.
1 comment
AuthorsOpenReview ↗

Response: Behavioral Grounding, Navigation, and Ablation

We thank the reviewer for the careful reading and detailed feedback. The questions about behavioral grounding, navigation, and dimensionality helped us strengthen the empirical validation and clarify the scope of our claims.

1. Target direction and offspring adherence

Appendix B.1, Table 7 records inputs and re-extracted offspring coordinates in a supplementary CVRP-ACO run, with Diversion threshold zero to increase observations.

Among successfully executed offspring with evaluable coordinates, 36/51 Diversion offspring (70.6%) satisfy (b_child - b_ref) · (b_target - b_ref) > 0, where the dot denotes the vector inner product.

This measures a positive displacement component along the target direction, supporting a directional tendency. Composition achieves 70.5% (103/146) adherence to parental coordinate intervals expanded by 0.10 per side. Local Exploitation achieves 74.3% (55/74) local preservation with maximum absolute coordinate change at most 0.20. These are explicit post-hoc adherence criteria.

2. Extraction consistency and execution-based grounding

Appendix B.1, Table 5 evaluates 160 algorithms (40 per task), each independently extracted five times under the same prompt and model configuration. ICC(2,1) values are 0.900, 0.765, 0.910, and 0.891 for SS, IU, UN, and AD, supporting repeatability.

Table 6 adds execution measurements for 500 CVRP-ACO and 495 TSP-GLS algorithms. SS measures historical solution novelty; IU measures contextual information dependence under controlled interventions; UN measures seed-dependent output variation under fixed inputs. CVRP-ACO AD measures system response to altered path-cost feedback; TSP-GLS AD is an explicitly identified instance-structure response proxy.

At minimum coordinate separation 0.4, UN concordance is 73.9% and 74.8% on the two tasks, TSP-GLS IU is 71.9%, and CVRP-ACO AD is 73.7%. At separation 0.5, SS reaches 70.2% and 75.5%. Runtime ties remain in the denominator. These results connect the coordinates to observable execution characteristics.

3. Diversion triggering and operator allocation

The implementation updates a counter after Composition and Local Exploitation. After Local Exploitation, Diversion is invoked if the counter is at least 5, and the counter is subsequently reset. There are two regular counter updates per outer cycle.

With population size 10 and exploitation rate 0.5, each outer cycle generates 10 Composition and 5 Local Exploitation offspring; each Diversion invocation adds 4 offspring. Overall proportions depend on cycle and Diversion counts.

4. Extended budgets and behavioral crowding

Time constraints prevented experiments with 2,000 or more evaluations. The implementation nevertheless explicitly manages behavioral redundancy. The behavior library stores at most 200 strategy units. When capacity is exceeded, units are processed in fitness order and retained only if their cosine distance from every retained unit is at least 0.05. If more than 200 remain, units in denser regions are removed using mean distance to five nearest neighbors, with fitness resolving density ties.

This limits library growth while preserving behavioral coverage. Section 3.3 also describes Diversion toward sparse frontiers near high-quality behaviors, using a feasible reference implementation. These mechanisms provide a basis for managing longer searches; performance with larger populations and budgets remains to be evaluated.

5. Removing AD on large MKP instances

Section 4.3 and Appendix B.2 already discuss the trade-off between behavioral specificity and generalization. The full representation has the highest mean at small and medium MKP sizes, while removing AD yields 103.118 versus 103.082 at n=500. Reduced dimensionality may allow greater flexibility under scale changes, but excessive reduction loses useful guidance: Table 4's 2D variant obtains 102.417 at n=500. Table 8 reports all dimensional subsets, supporting a balance between behavioral detail and flexibility across settings.

6. Behavioral information versus random perturbation

Section 4.3, Table 4 adds BeNav-Random, retaining strategy extraction and semantic descriptions while replacing coordinates with uniformly sampled vectors in the four-dimensional unit cube, fixed per distinct implementation within a run.

Three-run MKP-ACO objectives (mean +/- standard deviation) are:

Methodn=100 (higher is better)n=300n=500
BeNav23.229 +/- 0.0262.067 +/- 0.20103.082 +/- 0.48
BeNav-Random23.107 +/- 0.1159.282 +/- 3.6595.922 +/- 9.71

Random replacement lowers mean performance at all sizes and substantially increases variability. Together with Tables 5-7, this supports the usefulness of code-derived coordinates.

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

This paper proposes BeNav, a behavior-space navigation framework for automatic heuristic design. BeNav constructs an interpretable four-dimensional behavioral representation and integrates three mechanisms—Behavioral Composition, Behavioral Diversion, and Local Exploitation—to guide LLMs in exploring different regions of algorithmic behavior. The paper presents a clear motivation, a reasonably well-structured framework, and competitive experimental results across multiple optimization tasks.

However, some limitations remain. In particular, the paper lacks direct validation of the reliability of LLM-generated behavioral coordinates and the effectiveness of behavioral navigation. Furthermore, the ablation studies do not sufficiently demonstrate the contribution of the behavioral coordinates.

Critical strengths
  1. This paper proposes a behavior-space navigation approach for automatic heuristic design. Unlike existing methods that primarily rely on historical experience, code structure, or population information to guide algorithm generation, BeNav transforms algorithmic behavior from a descriptive feature into an explicit guidance mechanism, enabling the search process to determine exploration directions in the behavioral space.

  2. The paper constructs a four-dimensional behavioral space consisting of search style, information usage, uncertainty, and adaptivity, and introduces three complementary search mechanisms: Behavioral Composition, Behavioral Diversion, and Local Exploitation. These mechanisms focus on behavioral combination, exploration of new behavioral regions, and local refinement, respectively. Together, they provide a reasonable balance between exploration and exploitation, resulting in a coherent and well-structured framework.

  3. The experimental evaluation is relatively comprehensive. The authors evaluate BeNav on four heuristic design tasks covering different optimization problems and algorithmic frameworks. Comparisons with five representative existing methods, together with component ablation studies and additional analyses, provide preliminary evidence supporting the effectiveness of the proposed framework.

Critical weaknesses
  1. Insufficient validation of the reliability of behavioral representations. The four-dimensional behavioral coordinates are directly assigned by an LLM based on its analysis of algorithm source code. However, the paper does not sufficiently validate the reliability and consistency of these scores or their correspondence with the actual search behavior of the algorithms.

  2. Lack of direct validation of behavioral navigation effectiveness. Although BeNav constructs target behavioral positions, the paper does not quantitatively analyze whether newly generated algorithms actually move toward these targets. Without an analysis of the alignment between the target behavioral coordinates and the behavioral coordinates of generated algorithms, it remains unclear to what extent the proposed mechanisms achieve explicit behavioral navigation.

  3. Lack of computational cost analysis. Although all methods are evaluated under the same algorithm evaluation budget, the paper does not sufficiently account for the additional LLM calls and inference overhead introduced by BeNav.

  4. Insufficient ablation studies to isolate the independent contribution of behavioral coordinates. The paper evaluates the contributions of different components by reducing the number of behavioral dimensions and removing individual navigation modules. However, the current ablation studies do not adequately distinguish whether the performance improvements arise from the effective navigation information provided by the four-dimensional behavioral coordinates, the LLM's inherent ability to understand and generate algorithmic code, or the additional prompts and search strategies.

Questions for authors
  1. The paper uses an LLM to analyze algorithm source code and generate four-dimensional behavioral coordinates. Have the authors evaluated the consistency of the scores obtained from repeated behavioral extraction of the same algorithm? Furthermore, do these code-based behavioral coordinates accurately reflect the actual search behavior of the algorithms during execution?

  2. In Behavioral Diversion, the authors construct target behavioral coordinates and use them to guide the LLM in generating new algorithms. Could the authors quantitatively measure the differences between the behavioral coordinates of the generated algorithms and the intended target coordinates to verify whether the search actually moves in the expected direction?

  3. BeNav requires additional LLM calls for behavioral extraction and navigation guidance during algorithm generation. Could the authors provide statistics on the number of LLM calls, token consumption, or total running time for each method to further evaluate the computational efficiency of BeNav?

  4. The existing ablation studies do not sufficiently isolate the independent contribution of behavioral coordinates. Could the authors conduct additional experiments by removing, randomly replacing, or shuffling the behavioral coordinates to better distinguish whether the performance improvements arise from the four-dimensional behavioral navigation mechanism or from the LLM's inherent generation capabilities and additional prompting information?

  5. The four-dimensional behavioral coordinates are primarily obtained through LLM-based analysis of algorithm source code. Could the authors further validate whether these coordinates accurately capture behavioral characteristics such as search style, uncertainty, and adaptivity by analyzing the actual search trajectories of the algorithms on optimization problem instances?

2 comments
AuthorsOpenReview ↗

Response: Behavioral Validation and Coordinate Ablation

We appreciate the reviewer's careful reading and constructive feedback, as well as the recognition of BeNav's motivation and structure. The questions about representation reliability, navigation effectiveness, and coordinate contributions helped us identify where direct evidence was needed. We have added repeated extraction, execution-based measurements, a navigation case study, and a random-coordinate ablation to address these concerns.

1. Consistency of repeated behavioral extraction

Appendix B.1, Table 5 reports 160 distinct algorithms (40 per task), each independently processed five times using the same prompt and Qwen3-Coder-30B at temperature 1.0. Across 800 valid extractions, the single-measure absolute-agreement ICC(2,1) values are 0.900 for SS, 0.765 for IU, 0.910 for UN, and 0.891 for AD, supporting repeatability under the evaluated configuration.

2. Direct evidence of behavioral navigation

Appendix B.1, Table 7 reports a supplementary CVRP-ACO run recording operator inputs and re-extracted offspring coordinates. Its Diversion trigger threshold is set to zero to increase observations specifically for this diagnostic run.

For successfully executed offspring with evaluable coordinates, explicit post-hoc criteria yield:

OperatorCriterionAdherent / validRate
CompositionParental coordinate intervals expanded by 0.10 per side103/14670.5%
Local ExploitationMaximum absolute coordinate change at most 0.2055/7474.3%
DiversionPositive displacement along the target direction36/5170.6%

The Diversion criterion is (b_child - b_ref) · (b_target - b_ref) > 0, where the dot denotes the vector inner product.

It measures a positive displacement component along the specified direction. The case study supports a directional tendency and provides direct evidence of behavioral combination, local preservation, and directional guidance in generated offspring.

3. Computational overhead

We agree that equal evaluation budgets do not imply equal total inference cost. BeNav introduces overhead through behavioral extraction and behavioral context in generation and reflection prompts. Each extraction request jointly produces the semantic description and all four coordinates; coordinate operations, retrieval, and pruning are local computations. We have not completed a systematic inference-cost comparison covering all baselines. Our performance claims therefore concern solution quality under a common algorithm-evaluation budget, rather than superiority under equal token or wall-clock budgets.

4. Isolating numerical-coordinate contributions

Section 4.3 and Table 4 add BeNav-Random. It retains LLM-based strategy extraction and semantic descriptions but replaces extracted coordinates used for storage and downstream navigation with vectors sampled uniformly from the four-dimensional unit cube. Each distinct implementation receives a fixed random coordinate within a run, using a dedicated random-number generator.

Across three MKP-ACO runs, objectives are reported as mean +/- standard deviation:

Methodn=100 (higher is better)n=300n=500
BeNav23.229 +/- 0.0262.067 +/- 0.20103.082 +/- 0.48
BeNav-Random23.107 +/- 0.1159.282 +/- 3.6595.922 +/- 9.71

Random replacement lowers mean performance at all three scales and substantially increases run-to-run variability, particularly on larger instances. Retaining extraction and semantic descriptions makes this a more direct control for the contribution of code-derived numerical coordinates within the evaluated framework.

5. Correspondence with actual execution behavior

Appendix B.1, Table 6 adds execution measurements for 500 CVRP-ACO algorithms and 495 TSP-GLS algorithms. SS measures explored-solution novelty relative to search history; IU measures decision dependence on background instance information under controlled interventions; UN measures seed-dependent output variation under fixed inputs. CVRP-ACO AD measures system response to altered path-cost feedback; TSP-GLS AD is explicitly identified as an instance-structure response proxy.

We compare coordinate ordering with execution-measurement ordering for algorithm pairs separated by at least a threshold. At separation 0.4, UN concordance is 73.9% on CVRP-ACO and 74.8% on TSP-GLS; TSP-GLS IU reaches 71.9%, and CVRP-ACO AD reaches 73.7%. SS concordance increases with separation, reaching 70.2% and 75.5%, respectively, at 0.5. Runtime ties remain in the denominator. These measurements connect the representation to observable execution characteristics.

Thank you for the additional experiments. The repeated extraction and random-coordinate ablation alleviate some of my concerns, but the core concerns remain insufficiently addressed: some coordinates show weak correspondence with actual execution behavior; the navigation experiment lacks matched controls, and the current adherence criteria do not adequately demonstrate effective behavioral composition or movement closer to the target. Whether the offspring’s actual execution behavior changes as intended has also not been directly verified. Therefore, I find the evidence for the core navigation mechanism insufficient and maintain my original score.

Comments on the paper

1 comment

Dear reviewers,

We have now received the author responses to your raised questions and concerns. Please go through them. Also, have they answered your questions appropriately? Please also look other reviewers' strengths / concerns. Have they changed your perception of the paper? In all cases, please communicate with the authors. It is important that we let them know what relevant issues still exist or have been answered. Please remember we have limited time to get any further responses from them.

Regards,

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.