Engineering design often requires redesigning an existing solution when material costs or system requirements change. Such redesigns typically rely on simulators that are too complex or computationally intensive for traditional numerical solvers, which lack domain knowledge and struggle to guarantee global solution quality. Large language models (LLMs) can interpret semantic and engineering information, guiding the search; recent work shows that embedding this knowledge into optimization improves performance.\ \ We propose and evaluate an integrated LLM‑optimization framework for redesign tasks. Three ways of leveraging LLMs are compared:\
- warm start – using LLM‑generated initial solutions as the search seed;\
- variable fixing – letting the LLM fix a subset of variables to shrink the search space;\
- ReAct – having the LLM provide dynamic suggestions and corrections during the search.\ Four black‑box solvers are tested as baselines: Bayesian optimization, genetic algorithms, mesh adaptive direct search (MADS), and random search. The case study minimizes the cost of an electric‑vehicle battery pack across four redesign scenarios.\ \ The integrated approach matches the feasible‑design discovery of the solver baselines while cutting simulator evaluations by roughly 80% and tripling the fraction of runs that reach near‑optimal solutions. The optimal LLM‑solver pairing and the way the prior solution is exploited depend on the structure of the requirement change; the benefit of larger LLM models grows as the LLM exerts more control over the search.\ \ These results demonstrate clear potential for LLMs in redesign tasks, either to guide or replace traditional numerical solvers.\ \ Review