EvoCoT: Co-evolving generation and verification for LLM-based optimization autoformulation

Formulating optimization models from natural-language specifications is a major bottleneck in engineering decision-making, requiring both domain expertise and strict modeling discipline. Existing large language model (LLM) approaches often produce formulations that are executable yet semantically incorrect, or structurally invalid under solver constraints. Evolutionary Chain-of-Thought Optimization (EvoCoT) is introduced as a co-evolutionary autoformulation framework that treats model synthesis as guided discrete search over reasoning–code pairs rather than one-shot generation. EvoCoT couples three mechanisms in a single feedback loop: reasoning-first evolutionary search over a population of decoded programs, memory-guided exemplar retrieval, and a co-evolving verifier that provides structural-faithfulness feedback beyond static solver checks. Experiments on eight benchmark suites show consistent improvements over strong zero-shot baselines and competitive performance against agentic and fine-tuned systems, with the largest gains on difficult industrial-scale tasks. Attribution analyses under matched search budgets show that the population-level evolutionary search at the core of the framework accounts for these accuracy gains. The verifier and memory complement this search with reliability and triage rather than additional accuracy: a graded, execution-grounded trust signal for ranking surviving formulations, and an auditable, leakage-controlled library of validated formulations.

Authors

Institutions

Publication Details

Journal
Swarm and Evolutionary Computation
Published
2026-11-17
DOI
https://doi.org/10.1016/j.swevo.2026.102613
Primary Topic
Evolutionary Algorithms and Applications
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

EvoCoT: Co-evolving generation and verification for LLM-based optimization autoformulation

Xi Vincent Wang, Ruixiang Zheng, Fubo Qi, Tianyu Wang et al.
Swarm and Evolutionary Computation
Evolutionary Algorithms and Applications
article

EvoCoT: Co-evolving generation and verification for LLM-based optimization autoformulation

Xi Vincent Wang, Ruixiang Zheng, Fubo Qi, Tianyu Wang, Lihui Wang, Mian Li, Xu Ge
article en

Abstract

Formulating optimization models from natural-language specifications is a major bottleneck in engineering decision-making, requiring both domain expertise and strict modeling discipline. Existing large language model (LLM) approaches often produce formulations that are executable yet semantically incorrect, or structurally invalid under solver constraints. Evolutionary Chain-of-Thought Optimization (EvoCoT) is introduced as a co-evolutionary autoformulation framework that treats model synthesis as guided discrete search over reasoning–code pairs rather than one-shot generation. EvoCoT couples three mechanisms in a single feedback loop: reasoning-first evolutionary search over a population of decoded programs, memory-guided exemplar retrieval, and a co-evolving verifier that provides structural-faithfulness feedback beyond static solver checks. Experiments on eight benchmark suites show consistent improvements over strong zero-shot baselines and competitive performance against agentic and fine-tuned systems, with the largest gains on difficult industrial-scale tasks. Attribution analyses under matched search budgets show that the population-level evolutionary search at the core of the framework accounts for these accuracy gains. The verifier and memory complement this search with reliability and triage rather than additional accuracy: a graded, execution-grounded trust signal for ranking surviving formulations, and an auditable, leakage-controlled library of validated formulations.

Swarm and Evolutionary ComputationVol. 111
Shanghai Jiao Tong University (CN), KTH Royal Institute of Technology (SE)
National Natural Science Foundation of China
Evolutionary Algorithms and Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon