{"slug": "from-hand-crafted-to-llm-based-variation-operators-in-metaheuristics", "title": "From Hand-Crafted to LLM-Based Variation Operators in Metaheuristics", "summary": "Researchers at Apeiron Intelligence, the Artificial Intelligence Research Institute (IIIA-CSIC) and Universitat Politècnica de Catalunya (UPC) published a tutorial introducing an operator-level framework that classifies large language models used as variation operators in metaheuristics by two descriptors: the type of prompt-conditioning information (Numeric, Symbolic, Linguistic) and artifact persistence (Transient, Amortized, Transfer). The tutorial presents a worked build template, a method survey, an evidence table and a cost-aware decision guide, arguing that conditioning choice shifts cost across validation, evaluation and inference. The authors state that most real methods are hybrids, with only the load-bearing channel naming the method.", "body_md": "From Hand-Crafted to LLM-Based Variation Operators\nin Metaheuristics\n\nWhen a language model writes the next candidate, variation stops being a fixed kernel and\nbecomes a model call conditioned on information you choose. This tutorial gives\nyou the framework to classify that choice, build it, and decide whether it earns its cost.\n\n1Apeiron Intelligence, Barcelona ·\n2Artificial Intelligence Research Institute (IIIA-CSIC), Bellaterra ·\n3Universitat Politècnica de Catalunya (UPC), Barcelona\n\nLarge language models (LLMs) are increasingly being employed as variation operators in\nmetaheuristics, generating or modifying candidate solutions, heuristics, or programs inside\niterative search loops. This shift reframes variation as a model call conditioned on\ndifferent types of information. We introduce an operator-level framework\nwith two descriptors: (1) the type of prompt-conditioning information at variation time\n(Numeric, Symbolic, Linguistic), and (2) artifact\npersistence, identifying what survives the model call (Transient,\nAmortized, Transfer). The tutorial shows how to classify, build, and\nselect these operators through a worked build template, a method survey, an evidence table,\nand a cost-aware decision guide.\n\nlarge language modelsmetaheuristicsvariation operatorsevolutionary computationcombinatorial optimizationprompt conditioningartifact persistence\n\nThe question this tutorial answers\n\nOne operator, three prompts. Which information should condition it?\n\nConsider a practitioner building an LLM-assisted heuristic for the Travelling Salesman Problem.\nThe loop is the same in all three cases — ask the model for the next construction heuristic.\nWhat changes is what goes into the prompt. That single choice shifts cost across validation,\nevaluation and inference, and it is the axis the literature has not been organized around.\n\nWalkthrough\n\nThe whole tutorial, read aloud\n\nA two-voice walkthrough of the paper, in English, twenty-three minutes long. It follows the\nsame route as the text: the two descriptors, the placement rule and its audit, the build\ntemplate with its validator and bounded repair, the two traced iterations, the map and its\nempty cells, the evidence table, and the decision guide.\n\n01 — The organizing lens\n\nConditioning channels\n\nA method is classified by the dominant channel in its variation prompt: the one\nthat most drives the proposal. Most real methods are hybrids, so the full set of channels is\nreported and only the load-bearing one names the method.\n\nNumeric\n\n```\nh1: 427\nh2: 411\nh3: 419\n```\n\nscore-only trajectory over heuristics\n\nThe encoded search information classical metaheuristics already consume: solution encodings,\ninstance features, scalar scores, ranked parents, score trajectories. Evaluative content —\nhow good is this? — easiest to check, least portable.\n\nSymbolic\n\n``` python\ndef step(T, U):\n    c = nearest(T[-1], U)\n    return T + [c]\n```\n\nexecutable construction rule\n\nMachine-interpretable artifacts with formal structure — code, syntax trees, formal rules,\nstructured graphs — whose syntax or behaviour a parser, validator, compiler or executor can\ncheck. Denotational content — what does this compute or do? — verifiable by\nexecution.\n\nLinguistic\n\n```\n\"NN assigns tours to clusters;\n try a savings merge;\n then apply 2-opt repair\"\n```\n\nstrategy-level change to the heuristic\n\nOperative natural language: critiques, reflections, diagnoses, design principles, strategy\nnotes that steer later proposals. Propositional content — why does this work, and\nwhat should change? — travels furthest, least mechanically checkable.\n\nThe three channels sit side by side; none subsumes the others.Symbolic and Linguistic both add more\nexplicit structure than Numeric, but they are incomparable with each other: code can\nbe more structure-preserving than prose, while prose can express abstractions absent from the\ncode. Fixed task headers and static problem statements do not by themselves make a\nmethod Linguistic.\n\nArtifact persistence\n\nThe second descriptor asks a different question: what survives the model call?\n\ni\n\nTransient\n\nThe LLM sits inside the loop. Every step pays a model call, and nothing outlives the run.\n\nii\n\nAmortized\n\nThe model is called offline to emit an artifact — a heuristic, a program — that then runs at near-zero runtime cost.\n\niii\n\nTransfer\n\nSource-domain evidence is distilled into an inspectable artifact or principle, re-bound in a new domain without rediscovery. Still an emerging frontier — few methods reach this far.\n\n02 — The map\n\nWhere the field actually sits\n\nRepresentative methods placed on the two descriptors. Read each cell as a coordinate pair: what\nconditions the call, and what survives it. The empty cells are gaps in the current\nlandscape, not impossibility claims.\n\nNumeric\n\nSymbolic\n\nLinguistic\n\nTransfer\n\nopen\n\nopen\n\nLAPTHiFo-Prompt\n\nAmortized\n\nopen\n\nFunSearchLLaMEA\n\nEoHReEvo\n\nTransient\n\nLMXLLMOAEvoLLMOPROEvoPrompt\n\nMOLLEOLLEMAGI-LLM\n\nSGEMAEF\n\n† prompt-optimization boundary cases whose placement depends on prompt design.\nMethods cluster on and below the diagonal because higher persistence tends to require more\ndomain-independent conditioning — a structural tendency the current body of methods reflects,\nnot a law.\n\n03 — Inside the tutorial\n\nFour things you can take to your own loop\n\n01\n\nA worked build template\n\nThe full prompt → sample → parse-and-validate → bounded-repair → select loop, written out as\nan algorithm with an explicit repair cap, and mirrored by runnable code.\n\n02\n\nA method survey on one axis\n\nRepresentative and extended coverage of the literature classified by dominant conditioning\nchannel and persistence — 30 methods placed against their primary sources.\n\n03\n\nAn evidence table that refuses to be a leaderboard\n\nEach method's own headline result, tagged with whether it is peer-reviewed or preprint, and\nwhether the comparison is matched, source-local, or qualitative.\n\n04\n\nA cost-aware decision guide\n\nA staged flowchart whose first branch is the null choice: add an LLM operator only when it\nadds a capability a classical operator does not. The test is a capability gap, not model\navailability.\n\n04 — Evidence summary\n\nWhat each method actually reports\n\nReported results are source-specific. Compare across rows only where the evidence column says\nmatched — same evaluator, same harness. Everything else is a\nuseful signal, not settled comparative evidence.\n\nMethod\n\nConditioning\n\nEmitted\n\nCall locus\n\nEvidence\n\nReported result\n\nPR = peer-reviewed · pre = preprint · own = source-specific benchmark ·\nmatched = same evaluator · qual = qualitative only · † boundary case.\nBPP = bin packing problem · BBOB = the black-box optimization benchmark suite ·\np.p. = percentage points.\n\n05 — Appendix results\n\nPlacing every method: the drop-channel audit\n\nA classification is only as good as its rule. The audit applies one operation to every method:\nremove or neutralize a channel while keeping a valid scaffold, and see whether the operator still\nhas anything to work with. This is what resolves the awkward cases — methods that\nemit code yet are steered by natural language.\n\nMethod\n\nChannels present\n\nDominance\n\nDrop-channel reasoning\n\nNL = natural language · AHD = automatic heuristic design, the offline design of a\nheuristic by search.\n\nValidation of the placement rule\n\nTwo checks accompany the companion code — one for reproducibility, one for whether a channel is\nactually load-bearing. The second returned a negative result, and it is reported as such.\n\n3independent LLM codersOne instance each from Claude Fable 5, GPT-5.5 and DeepSeek-V4-Pro, queried separately and shown only the placement rule and the channel descriptions.\n\n9/12methods coded unanimouslyThe three divergences fall where the channel description is least determinate: OPRO and LMX, whose bare “solution” descriptions under-determine the representation, and ReEvo, where reflection and code-substrate readings compete.\n\n0.73Gwet's AC1 agreementChance-corrected agreement across the three coders — the measure that discounts the agreement you would get by luck alone. The labelling is largely reproducible; its contested cases are exactly the annotated boundary ones.\n\nLoad-bearing test — a negative result, reported\n\nA drop-channel ablation compared an EoH-style operator on online bin packing with and without\nits natural-language idea, holding parent code and scores fixed. On this easy task\nboth conditions reached the optimum (mean gap 0%; 6 of 6 candidates\noptimal). The ablation did not demonstrate performance necessity for the linguistic channel\non that task.\n\nEoH is still reported as Linguistic: the idea is the source-described\ndesign-intent signal that structurally distinguishes it from code-only heuristic design.\nPlacement is reproducible; how much a present channel contributes is task-dependent.\n\nExtended coverage map\n\nA broader set of methods placed on the lens, beyond the representative ones, to show the field's\nbreadth. Each entry is read against its primary source. It shows the shape of the field; the\nsystematic census lives elsewhere. Entries marked adjacent fall outside the lens: they translate a specification\nrather than vary a candidate, and are listed only as foils.\n\nMethod\n\nDominant conditioning\n\nPersistence\n\nConditions on → emits\n\n06 — Companion code\n\nRuns offline. No API key, no network.\n\nA deterministic mock model returns completions from a fixed pool, so the whole\nbuild → sample → validate → repair → select loop executes end to end and reproduces exactly.\nAn optional adapter is included for real model runs. Python 3.9+, standard library only.\n\nsearch.pythe reference build loop, with bounded repair\n\nllm.pypluggable interface: offline mock + optional real backend", "url": "https://wpnews.pro/news/from-hand-crafted-to-llm-based-variation-operators-in-metaheuristics", "canonical_source": "https://camilochs.github.io/semantic-turn-metaheuristics/", "published_at": "2026-09-13 12:54:24+00:00", "updated_at": "2026-09-13 13:10:25.248624+00:00", "lang": "en", "topics": ["large-language-models", "machine-learning", "ai-research", "ai-tools"], "entities": ["Apeiron Intelligence", "Artificial Intelligence Research Institute (IIIA-CSIC)", "Universitat Politècnica de Catalunya (UPC)", "Travelling Salesman Problem"], "alternates": {"html": "https://wpnews.pro/news/from-hand-crafted-to-llm-based-variation-operators-in-metaheuristics", "markdown": "https://wpnews.pro/news/from-hand-crafted-to-llm-based-variation-operators-in-metaheuristics.md", "text": "https://wpnews.pro/news/from-hand-crafted-to-llm-based-variation-operators-in-metaheuristics.txt", "jsonld": "https://wpnews.pro/news/from-hand-crafted-to-llm-based-variation-operators-in-metaheuristics.jsonld"}}