{"paper":{"title":"Large Language Models as Optimization Controllers: Adaptive Continuation for SIMP Topology Optimization","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"A large language model can act as a real-time controller for SIMP topology optimization by choosing parameters from current state observations instead of a fixed schedule.","cross_cats":["cs.AI"],"primary_cat":"cs.CE","authors_text":"Jun Wang, Shaoliang Yang, Yunsheng Wang","submitted_at":"2026-03-26T07:14:31Z","abstract_excerpt":"We present a framework in which a large language model (LLM) acts as an online adaptive controller for SIMP topology optimization, replacing conventional fixed-schedule continuation with real-time, state-conditioned parameter decisions. At every $k$-th iteration, the LLM receives a structured observation$-$current compliance, grayness index, stagnation counter, checkerboard measure, volume fraction, and budget consumption$-$and outputs numerical values for the penalization exponent $p$, projection sharpness $\\beta$, filter radius $r_{\\min}$, and move limit $\\delta$ via a Direct Numeric Control"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The LLM agent achieves the lowest final compliance on every benchmark: -5.7% to -18.1% relative to the fixed baseline, with all solutions fully binary.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The structured observation vector (compliance, grayness index, stagnation counter, checkerboard measure, volume fraction, budget consumption) plus the LLM's learned mapping is sufficient to produce superior real-time parameter decisions that the ablation study attributes to the agent's intervention rather than schedule geometry alone.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"An LLM acting as real-time controller for SIMP topology optimization parameters outperforms fixed schedules and heuristics, delivering 5.7-18.1% lower compliance on 2D and 3D benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A large language model can act as a real-time controller for SIMP topology optimization by choosing parameters from current state observations instead of a fixed schedule.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d35b57c77c1cb25f281b85757b603b44ee94eb8ba097809c2b9bca86b75f6dde"},"source":{"id":"2603.25099","kind":"arxiv","version":2},"verdict":{"id":"8dbb6518-5ba2-4f4e-baaa-1777e7cb0a43","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:29:57.153375Z","strongest_claim":"The LLM agent achieves the lowest final compliance on every benchmark: -5.7% to -18.1% relative to the fixed baseline, with all solutions fully binary.","one_line_summary":"An LLM acting as real-time controller for SIMP topology optimization parameters outperforms fixed schedules and heuristics, delivering 5.7-18.1% lower compliance on 2D and 3D benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The structured observation vector (compliance, grayness index, stagnation counter, checkerboard measure, volume fraction, budget consumption) plus the LLM's learned mapping is sufficient to produce superior real-time parameter decisions that the ablation study attributes to the agent's intervention rather than schedule geometry alone.","pith_extraction_headline":"A large language model can act as a real-time controller for SIMP topology optimization by choosing parameters from current state observations instead of a fixed schedule."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2603.25099/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":74,"sample":[{"doi":"10.1038/nature23911","year":2017,"title":"Lazarov, and Ole Sigmund","work_id":"47521b4d-9ae3-493d-8d06-4a9b8d83ee3a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1016/j.compstruc.2020.106283","year":2020,"title":"Abueidda, Seid Koric, and Nahed A","work_id":"39d67515-30e2-477f-ac64-6275d9802c71","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1613/jair.1.13922","year":2022,"title":"Automated dynamic algorithm configuration.Journal of Artificial Intelligence Research, 75:1633–1699, 2022","work_id":"02a02dd1-c9a5-487c-af38-c1515ec673f6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1007/s00158-010-0594-7","year":2011,"title":"Lazarov, and Ole Sig- mund","work_id":"e4cc86b8-99bf-40a7-bf14-ed288e84c566","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"The Claude model card and evaluations","work_id":"536847b6-3412-470f-80e3-0fde56208b74","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":74,"snapshot_sha256":"c0c90095409891924a71c937e5b42297b83a4448adfc112ae6965581def9a322","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c88e6c9e71682d3ce47884d86323adaac2d823a971252b6f8bfc59ab5e8c2aa3"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}