acceptodds
Under review as a conference paper at ICLR 2027

Representation, Not Search Strategy: Typed DSL Operators Transfer to Unseen Routing Variants Where Free-Form Code Does Not

Abstract

Large language models can evolve local-search operators for vehicle routing, and the published systems that do so — EoH, FunSearch, ReEvo, HiFo-Prompt — all evolve them as free-form source code. We ask whether that choice of representa- tion is what decides whether an evolved operator works on a routing variant the designer never had in mind. We study operator-set evolution: the search starts from an empty operator set, which is pure construction, and each round a newly proposed operator is tried in every position of the current set — replacing one of the k incumbents, or appended to make k + 1 — with the incumbent set replaced only on strict improvement. Because there is no hand-written floor, a failed round cannot be blamed on the floor having already absorbed the available gains. On three unseen variants, a typed DSL front-end with a semantics-preserving com- piler reaches 20.6% mean improvement over its empty-set start, the best of four LLM baselines that evolve free-form code reaches 19.2%, and a free-form arm under the identical 100,000-token budget reaches 4.8%. The gap is not a tuning artefact: it comes with a 11.8×higher cost per proposal for free-form code, a 3/5 static-rejection rate, and candidates that compile, run, and accept zero improving moves. All trajectories, operator sets and the harness are released.

Primary area:
optimization
Keywords:
combinatorial optimization, automatic heuristic design, large language models, local search, domain-specific languages

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_7H4WOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the LLM-based evolution of local-search operator sets for vehicle routing. The main idea is to replace free-form source code generation with a typed DSL and compiler. The search starts from an empty operator set. At each round, a newly proposed operator is tested by replacing each incumbent operator or by being appended to the current set, and an update is accepted only when the objective improves. The authors evaluate EoLS against a raw free-code arm and strategy layers adapted from EoH, FunSearch, ReEvo, and HiFo-Prompt under a common solver host and token budget. The paper calls three routing variants “unseen” because their constraint combinations are not covered by the existing hand-written operator library. The main claim is that the typed DSL transfers better to these variants and that representation is more important than the choice of search strategy. The paper also analyzes token consumption, static rejection, and operators that compile and run but fail to produce improving moves.

Critical strengths
  1. I find the general empirical question useful. It is reasonable to ask whether the representation used for LLM-generated operators affects search efficiency and reliability, and the proposed typed DSL is a promising implementation of this idea.

  2. I appreciate the attempt to go beyond final objective values and inspect failure modes. Measurements such as token cost per proposal, static rejection, and the number of accepted improving moves could be useful for understanding why an LLM-based search succeeds or fails. The empty-set setup is also an interesting attempt to make the contribution of evolved operators easier to inspect.

Critical weaknesses
  1. The central claim is substantially stronger than the novelty and evidence support. The general idea of changing the search representation is not new in LLM-based algorithm discovery. Prior work has already searched over operator design thoughts and combinations [1], semantic concepts [2], knowledge as the primary search object [3], multiple representation spaces [4], and joint code/thought representations [5]. E2OC is particularly close because it already performs progressive replacement inside a multi-operator combination. The new contribution therefore appears much narrower, mainly a routing-specific typed DSL and compiler. More importantly, the empirical comparison used to support the title-level claim is seriously unbalanced. EoLS is evaluated with multiple independent trajectories, while EoH, FunSearch, ReEvo, and HiFo-Prompt are each reported from only a single trajectory. This is especially problematic because EoLS itself shows large stochastic variation, ranging from 10.6% to 25.5% on sd_q_tw_o, while the single EoH run achieves 19.2%, well within that range. Comparing a multi-run mean for EoLS against single-run outcomes for stochastic LLM baselines is not a fair basis for concluding that representation dominates search strategy. In addition, EoLS receives many more proposals under the same token budget and benefits from compiler/backend support that removes many low-level validity failures. These results demonstrate useful engineering advantages, but they do not cleanly establish that the DSL itself produces better operators or that representation is the dominant factor. I am also not convinced by the transfer framing. EoLS appears to search directly on the support units of each “unseen” target variant. In that case, “unseen” seems to mean unsupported by the handwritten library, not unseen to the search process. This is different from the usual source-task to held-out-target notion of transfer. Finally, the empty-set argument is overstated. Once an incumbent set S_k contains evolved operators, a new operator can still have zero marginal improvement because the existing set has already captured the same moves. Replacement reduces this issue, but it does not remove it. The protocol measures contribution conditional on the current set, not the standalone usefulness of an operator.

  2. The manuscript is incomplete, and the experimental reporting is not reliable enough in its current form. The referenced appendix is missing even though the paper points to it for prompts, model details, and additional results. Section 5.1 says that three unseen variants are evaluated, but Table 1 reports only two, while Section 6 later states that the main table covers three. The abstract and conclusion report 4.8% for the free-form arm, whereas Table 1 gives a mean of 0.1%, and the listed EoLS replicate count is also inconsistent with the values shown. There is also a clear mismatch between the framework presented in Figure 1 and what is actually specified and evaluated. Figure 1 includes a population/reproduction process that is not sufficiently described in the methodology, and it presents an implementation-optimization channel with a “Correct and Faster” gate as an integral component, although Section 4.2 states that this channel is not exercised in any reported experiment. The supposedly held-out “selection units” are not reported, and the terminology itself is confusing because they are never used for selection. The “ablation” in Section 5.6 does not ablate any component of EoLS. The paper also contains several clarity and proofreading problems, including undefined terminology and a duplicated paragraph. The authors disclose AI assistance for drafting and editing, which is not an issue by itself, but it does not remove the need to carefully verify and polish the final manuscript. Given these critical issues, the overall manuscript quality is substantially below the standard I would expect for a conference submission.

References

[1] Qiu et al., Evolving Interdependent Operators with Large Language Models for Multi-Objective Combinatorial Optimization, 2026.

[2] Leleu et al., Contrastive Concept-Tree Search for LLM-Assisted Algorithm Discovery, 2026.

[3] Kiet et al., Back to the Beginning of Heuristic Design: Bridging Code and Knowledge with LLMs, 2026.

[4] Guo et al., CoEvo: Continual Evolution of Symbolic Solutions Using Large Language Models, AAAI 2026.

[5] Hu and Zhang, Partition to Evolve: Niching-enhanced Evolution with LLMs for Automated Algorithm Discovery, NeurIPS 2025.

Questions for authors
  1. What exactly is meant by “transfer” and “unseen variant” in this paper? My reading is that EoLS searches directly on the support units of every target variant. If so, what artifact learned on one task is transferred to another task?

  2. What is the complete search procedure used by the EoLS arm? In particular, how is o* generated, how is the operator population in Figure 1 maintained, how are parents selected, what reproduction or mutation operation is used, and what search history is included in the LLM prompt? These should be explained in the methodology and appendix of the manuscript; however, they are currently missing from it.

  3. How is the proposed operator-set evolution different from E2OC’s operator-rotation mechanism? Relatedly, what prior work supports the statement that the “usual way” is to keep a fixed hand-written operator floor and only append one new operator at each round?

  4. What exactly is generated by the LLM in the typed-DSL arm, and what is supplied by the compiler or backend? Could you give some examples of the evolved high-performing operators? Please also clarify how the comparison isolates representation when the DSL receives built-in support for feasibility and low-level execution details. Can the authors provide a matched-proposal-count comparison or statistics such as proposals and tokens per accepted improvement?

  5. Please provide the missing pd_q_tw_o results and the final results on the selection units. Please also reconcile the 4.8% free-form result in the abstract and conclusion with the 0.1% mean in Table 1, and explain the n=3+4n=3+4 entry when five sd_q_o values are listed.

  6. What is the covered variant used in Section 5.6? Please report the per-method and per-replicate results, including the EoLS result. Can the authors also provide actual component ablations for the DSL constraints, compiler support, empty-set initialization, and replacement mechanism?

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

This paper studies whether the representation in which local-search operators are evolved affects their performance on unseen routing variants. It proposes operator-set evolution using a typed DSL with a semantics-preserving compiler. The search starts from an empty operator set; each round, a newly proposed operator is tried in every existing position or appended, and the incumbent set is replaced only on strict improvement. Under a shared token budget, the paper compares this approach with free-form code generation and four LLM-based search strategies. It reports advantages for the DSL approach and attributes them to lower proposal costs, fewer static rejections, and fewer semantically inert operators.

Critical strengths
  1. The paper studies representation as an explicit variable in LLM-driven operator discovery. This is a relevant question for automatic heuristic design on unseen routing variants.
  2. Operator-set evolution combines an empty start, replacement or addition, and strict improvement. This helps assess evolved operators without a fixed hand-written floor.
  3. The paper analyzes proposal costs, static rejection, and semantically inert operators. These diagnostics help explain where free-form evolution loses.
Critical weaknesses
  1. The DSL also provides domain-specific primitives, compiler support, and backend-owned feasibility checks, confounding the claimed representation effect. Stronger evidence is needed that implementation requirements are a major bottleneck for current LLMs.
  2. The abstract reports 4.8% improvement for free-form code, while Table 1 reports 0.1%. The listed DSL replicates also average to 19.4%, rather than 23.8%, weakening confidence in the results.
  3. The presentation in the abstract is confusing, and “DSL” is not defined at first occurrence. The title formatting also differs from the ICLR template.
Questions for authors
  1. Clarity of presentation. The sentence “We ask whether that choice of representation is what decides whether an evolved operator works on a routing variant the designer never had in mind” is confusing. In addition, “DSL” is not defined at its first occurrence in the abstract.
  2. Title formatting. The title uses mixed-case lettering rather than the small caps required by the ICLR 2027 template.
  3. Evidence for the implementation bottleneck. I am not convinced that generating “window lists, frame machinery and feasibility checks,” or satisfying the operator interface, is a major bottleneck for current LLMs. The paper needs stronger evidence that these implementation requirements explain where free-form evolution loses.
  4. Attribution of the representation effect. The typed DSL includes route primitives, compile-time obligations, and backend-owned feasibility checks. The comparison therefore needs to disentangle the contributions of these components to support the claim that “representation, not search strategy, is the dominant factor.”
  5. Internal inconsistencies in the reported results. The abstract reports a 4.8% mean improvement for the free-form arm, whereas Table 1 and Section 5.5 report 0.1%. Furthermore, the per-replicate improvements listed for EoLS on sd_q_tw_o—25.5%, 22.1%, and 10.6%—average to 19.4%, rather than the 23.8% reported in Table 1. These discrepancies need to be reconciled.
Reviewer_DiMCOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether typed DSL representations improve LLM-generated local-search operators for routing constraint combinations not covered by a pre-existing operator library. It proposes operator-set evolution from an empty set, with each proposal evaluated through replacement at every existing position and through appending. The incumbent is updated only under strict improvement. A compiler and backend handle interface obligations, legal composition, and feasibility. The method is compared with direct Rust generation and strategy layers from EoH, FunSearch, ReEvo, and HiFo-Prompt.

Critical strengths

1.Relevant research question. Adaptation to new constraint combinations is an important challenge for automated heuristic design, beyond improving performance on familiar benchmarks.

2.Clear operator-set search formulation. The empty start reduces ambiguity about the contribution of generated operators, while replacement allows weak operators to be removed rather than permanently accumulated.

3.Useful mechanism-level analysis. Proposal token cost, static rejection, and accepted-move counts provide more informative evidence than final objective values alone.

Critical weaknesses
  1. Missing 3rd Variant: The abstract, introduction, and Sec. 5.1 claim evaluation on three unseen variants (sd_q_tw_o, sd_q_o, and pd_q_tw_o), yet Table 1 only reports two variants (pd_q_tw_o is completely absent).

  2. Missing Baselines & n=1 Runs: All four baselines (EoH, ReEvo, FunSearch, HiFo-Prompt) are missing (–) on sd_q_o, and are run with only n=1 trajectory on sd_q_tw_o. Given the massive replicate variance of EoLS (10.6% to 25.5%), drawing strong conclusions against baselines with n=1 (where EoH achieves 19.2%, beating one EoLS replicate) is statistically unsupported.

  3. The supplementary materials cannot be found. The word "appendix" is mentioned in many places, but it cannot be located.

Questions for authors

See Critical Weaknesses

Reviewer_jJW1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies whether, when an LLM evolves local-search operators for vehicle routing, the representation in which operators are expressed determines whether they transfer to routing variants the designer did not anticipate. The authors propose an operator-set evolution formulation: starting from an empty operator set, each round tries a newly proposed operator in every position of the current set (either replacing an incumbent or being appended), and the incumbent set is replaced only on strict improvement; operators are written in a small, self-designed typed DSL. On several unseen variants, EoLS achieves a higher mean improvement over the empty-set start than EoH and the free-form code.

Critical strengths

None

Critical weaknesses
  1. The typed DSL compresses the search space into a closed set of predefined primitives, which may run counter to the trend toward general, free-form LLM generation, and the paper does not discuss the resulting expressiveness ceiling. The typed DSL restricts neighborhoods to combinations of a small number of hand-written primitives such as split, join, rev, and orient; introducing a new problem class or constraint requires designing new primitives and compile-time obligations in advance, and this manual engineering cost grows with the problem domain. Free-form code, by contrast, is not bounded by a predefined closure and requires no bespoke language, making it more likely to generalize as model capability improves. The paper neither argues that this primitive set scales to new constraints nor reports the manual-design cost of adding a new problem class; consequently, the claim that representation (rather than the strategy layer) is the dominant factor is confined to the hand-built primitive set used here, and its external validity is questionable.
  2. The experimental validation is insufficient: the sample size and key figures do not support the strong claims. Each of the four strategy baselines (EoH/ReEvo/FunSearch/HiFo) runs only a single trajectory (n=1), with no replicates and no significance test; EoH's 19.2% is already close to EoLS's 22.2%, and this gap falls inside EoLS's own replicate range (10.6%–25.5%). The abstract and conclusion report 4.8% improvement for the free-form code under the identical budget, whereas Table 1 and §5.5 report 0.1%, and the manuscript never explains this discrepancy in measurement. The paper claims three unseen variants, yet the entire pd_q_tw_o column is missing from the table. The coverage ablation in §5.6 reports only a 0.27%–0.78% range with no per-arm quantitative values. Moreover, only deepseek-flash is used, so the experiments do not examine whether free-form code still fails under a stronger model.
Questions for authors

The reported experimental results are inconsistent and difficult to reconcile. what do the 4.8% (abstract/conclusion) and 0.1% (Table 1) figures for the free-form code correspond to, respectively, and why is there no result for pd_q_tw_o? Given n=1 baselines and the absence of a significance test, how do the authors rule out that the gap between EoH (19.2%) and EoLS (22.2%) is merely sampling variance, while still supporting the title-level claim that representation, rather than the strategy layer, is the dominant factor?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.