The authors compare three inference settings for translating natural-language optimization problems into executable solver programs: direct generation, OptiCycle, which organizes code generation, execution, feedback interpretation, and revision in a persistent Python environment, and OptiState, which adds an explicit, revisable formulation state.
Execution-guided revision improves LLM optimization modeling without extra training
Two lightweight inference harnesses run and repair solver code, outperforming adapted agentic workflows and post-trained models across nine benchmarks and 11 Qwen variants.
Chinese Tech
Yunxiang Luo · Xin Xu · Tianhao Chen · Zhixi Xiong · Peiqi Jiang · Fan Zhang · +4 more
The Hong Kong University of Science and Technology · Huawei
Research Digest··2 min read
The authors present OptiCycle and OptiState, two inference harnesses that let a frozen language model iteratively execute and revise solver code.
§