acceptodds
Under review as a conference paper at ICLR 2027

HypaEvo: Capability-Aligned LLM Scoring for Metaheuristic Hyperparameter Control

Abstract

Dynamic hyperparameter control is a long-standing challenge for population-based metaheuristics. Learning-based controllers achieve strong performance but require expensive training and generalize poorly. Recent zero-shot approaches use Large Language Models (LLMs) to bypass training, but existing architectures route the final control decision through open-ended numerical generation (a regime in which LLMs are comparatively less reliable) while leaving their stronger capabilities in ordinal assessment underexploited. We propose **HypaEvo**, a *capability-aligned dual-LLM scoring architecture* that confines the LLM to bounded ordinal judgment and delegates numerical actuation to deterministic code. Two specialized agents independently rate the urgency of exploration and exploitation on a discrete ordinal scale (0–10); a deterministic coordinator translates the score differential into proportional, range-normalized parameter adjustments via a fixed direction map, enforcing directional priors and parameter bounds by construction. This separates qualitative reasoning from numerical arithmetic, removing open-ended numerical generation from the actuation step. Across four combinatorial optimization problems (TSP, CVRP, FSSP, and UAV trajectory optimization) and three metaheuristic families (GA, PSO, and ACO) with Qwen3-30B as the backbone, our method consistently outperforms representative state-of-the-art learning-based, neural-evolution, and LLM-driven controllers.

Primary area:
optimization
Keywords:
LLMs, Optimization, Metaheuristic algorithm, Automatic Algorithm Design

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

This paper studies online hyperparameter control for population-based metaheuristics using frozen LLMs. It proposes HypaEvo, which separates search-state assessment from numerical actuation. Experiments cover TSP, CVRP, FSSP, and UAV trajectory optimization with GA, ACO, and PSO backbones. Comparisons with AutoEP and other baselines, component ablations, scorer-replacement experiments, and cost analyses support the practical effectiveness of the proposed architecture.

Critical strengths
  1. Separating LLM assessment from numerical actuation is an attractive design choice. Asking an LLM to assess search conditions while deterministic code handles arithmetic, parameter bounds, and update constraints reduces the burden of generating absolute parameter values. This provides a practical and interpretable interface for LLM-based optimization control without controller-specific training.
  2. The experimental design of the paper is relatively rigorous. HypaEvo improves solution quality over the re-run AutoEP baseline on the reported TSP comparisons and extends to CVRP, FSSP, and UAV optimization with different metaheuristic families. In addition, the paper provides sufficient ablation experiments and sensitivity analysis.
Critical weaknesses
  1. A little technical concern. HypaEvo prespecifies "in which direction each hyperparameter should be adjusted to promote exploration or exploitation". However, reliance on fixed directional priors may be brittle when hyperparameter effects are strongly state-dependent or interaction-dominated; the method does not adapt these priors online.
  2. The claim that only ordinal assessment is required is not fully consistent with the actual control rule. Section 4.3 states that the scores need only be ordinally comparable under a shared rubric. However, Section 4.4 uses their difference, z = (u_E - u_X) / 10, to determine both update direction and magnitude, and their sum, a = min(1, (u_E + u_X) / tau_ref), to determine the urgency gate. A shared ordering can determine the sign of the difference, but cannot determine these magnitudes: they also depend on numerical spacing and consistency between the two assessors' scales. An order-preserving nonlinear recoding can preserve score rankings while changing both the update size and the gate. The authors should describe the scores more precisely as bounded numerical control signals and test sensitivity to monotone recoding and cross-role calibration differences.
  3. The contribution of the scoring interface is not yet fully separated from the contribution of action constraints. The same-coordinator experiment establishes the usefulness of LLM-derived scores relative to the tested alternative scorers, while the scoring/direct-generation comparisons evaluate different actuation interfaces. Neither fully isolates whether scoring itself is more reliable than numerical generation once both receive the same local action restrictions, directional priors, step-size bounds, gating, and damping. The LLM-actuator ablation also changes how these structural guarantees are enforced. A stronger control would let an LLM generate a bounded signed update that is then passed through the same deterministic constraints. Without this comparison, some of the gains attributed to capability-aligned scoring may come from restricting and regularizing the action space.
  4. Some empirical conclusions would benefit from more precise uncertainty and budget comparisons. The paper reports repeated runs and variability in several analyses, and Appendix D.1 clarifies the matched search budgets and seed sets. However, the principal TSP table and component-ablation tables mainly report means, making small differences difficult to assess. Some backbone-study quality differences are also small relative to the reported standard deviations. Paired uncertainty estimates or significance tests would strengthen these claims. In addition, reduced runtime relative to AutoEP does not by itself establish an advantage over cheaper non-LLM controllers under equal wall-clock budgets; quality-versus-time comparisons would better characterize when LLM control is worthwhile.
Questions for authors
  1. What numerical comparability is actually required between the Explorer and Exploiter scores? Can the authors test order-preserving nonlinear recodings and role-specific score compression or offsets, and compare the current coordinator with a sign-only or rank-based controller? These tests would clarify the gap between the ordinal interpretation and magnitude-dependent actuation.
  2. Can the authors compare score-based control with direct generation of a bounded signed update under the same direction map, step-size bounds, gate, damping, feature access, and decision schedule? This would help distinguish the effect of the scoring interface from that of action constraints.
  3. How sensitive is HypaEvo to incorrect directional priors or state-dependent reversals in parameter effects? Are there examples where the shared scalar update prevents useful independent changes to different hyperparameters?
  4. Can the authors provide paired uncertainty estimates for the main quality comparisons and evaluate solution quality under equal wall-clock budgets against AutoEP and inexpensive non-LLM controllers?
Reviewer_EdNnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free LLM architecture (called HypaEvo) for dynamic hyperparameter control in population-based metaheuristics. Two specialized assessors score exploration and exploitation urgency from structured search-state features. A deterministic coordinator converts these scores into bounded parameter updates using directional priors, urgency gating, and late-stage damping. Experiments cover four problem classes and three metaheuristic families, with comparisons against existing controllers and extensive component analyses.

Critical strengths

The architecture is a simple but practical improvement over existing methods. Separating assessment from numerical actuation provides explicit control over parameter bounds and update magnitudes. The intermediate urgency scores also make the control process easier to inspect.

Thorough controlled experiments and cross-task validation. The evaluation spans multiple problem classes and metaheuristic families. Matched controller comparisons, scorer replacements under a fixed coordinator, and component ablations provide substantial empirical support for the effectiveness of the combined architecture.

Critical weaknesses
  1. The methodological advance over existing approaches appears relatively limited.

AutoEP already uses online search information and LLM reasoning to adjust metaheuristic parameters without additional controller training. HypaEvo mainly changes how this control is implemented: it separates exploration and exploitation assessment, asks the LLMs to output urgency scores, and uses a deterministic coordinator to update the parameters. These changes provide a useful and practical improvement, and the experiments show their effectiveness. However, the methodological contribution mainly lies in redesigning the control interface within an existing approach. A clearer account of what broader insight this design provides about LLM-based control would strengthen the contribution.

  1. The numerical meaning of the urgency scores needs clearer explanation.

The paper describes the scores as ordinal urgency levels, but the coordinator uses their numerical differences and sums to determine updates. The difference between exploration and exploitation scores controls the direction and size of the adjustment, while their sum controls the urgency gate. The update rule therefore gives the scores a numerical meaning beyond their ordering. This mapping also depends on the two assessors using comparable scales. If the exploration assessor tends to assign higher scores than the exploitation assessor, their difference could favor exploration even when the underlying needs are balanced. The remaining concern is how sensitive that balance is to the way the scores are assigned.

  1. The experiments do not fully separate the capability-alignment explanation from the effects of the coordinator’s update rules.

The paper argues that LLMs are better suited to judging urgency than directly choosing numerical parameter values. However, the comparison changes both the LLM’s prediction task and the parameter update process. HypaEvo applies small steps, scaled by each parameter’s range, along predefined directions. It also uses urgency gating and reduces the step size late in the search. The direct-output variants generate absolute parameter values. Checking these values and keeping them within the allowed ranges does not impose the same restrictions on the size or direction of each update. The performance gain could therefore partly come from more controlled parameter updates. This comparison does not fully establish whether asking the LLMs for urgency scores works better because it matches their capabilities, or because the coordinator restricts how their outputs affect the search.

  1. The proposed method is compared systematically to a set of algorithms of a similar flavor. In order to be appealing to the wider audience that is present at this conference, it is necessary to compare also (or instead of) to other types of algorithms for the approximation a TSP solution.
Questions for authors
  1. What generalizable insight about LLM-based control does HypaEvo establish beyond the specific architectural improvement over existing controllers such as AutoEP? Which findings most directly support that insight?

  2. Why are the numerical difference and sum of the two ordinal urgency scores appropriate signals for parameter updates? How does the method account for possible differences in how the exploration and exploitation assessors use the scoring scale?

Reviewer_7fLuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HypaEvo, a method that uses two LLM assessors to score exploration and exploitation urgency, then maps those scores to bounded hyperparameter updates with a deterministic coordinator. Experiments cover four problem classes and GA, PSO, and ACO backbones.

Critical strengths
  1. Separating assessment from numerical updates gives the method explicit bounds and direction constraints. It is a useful design choice for LLM-based control.

  2. The same-coordinator scorer replacements and agent/interface ablations go beyond a simple comparison with AutoEP. Results across several backbones suggest the idea is useful.

Critical weaknesses
  1. The ordinal-scoring argument is incomplete. The coordinator uses score differences and sums to set update size, so order alone is not enough. Also, score-based control uses bounded local steps while direct generation predicts absolute values; the ablation does not separate scoring from this action-space restriction.

  2. Practical gains need stronger evidence. Most result tables lack paired uncertainty estimates, and the small UAV improvement is difficult to judge. LLM overhead is substantial relative to the base solvers, but there is no equal-wall-clock comparison.

Questions for authors
  1. Does scoring still help against a Direct-Delta baseline with the same direction map, step limits, and damping?

  2. How sensitive is performance to order-preserving score remappings and repeated assessments of the same state?

  3. Can you report paired confidence intervals, equal-time solver comparisons, and the default decision interval and exact inference settings?

Comments on the paper

1 comment
Area_Chair_ygsCOpenReview ↗

On author responses

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.