Dynamic Adaptation of the LLM Context for Generating Routines with Coupled Semantics

To address the limitations of static binding in large language models when generating code with runtime-coupled semantics, researchers proposed a dynamic context-adaptive method. This method uses verification agents to extract structured diagnostic information from execution traces, providing gradient-based guidance for the generation agents. It also employs knowledge graphs based on problem descriptions to provide semantic constraints, while using simulated annealing mechanisms to avoid greedy collapse. Experimental results show that this method outperforms baseline models such as zero-sample, Reflexion, and OpenEvve on seven eight-item test problems, achieving the best performance in 300 to 600 evaluations; it also achieved the highest score in 1,000 evaluations of the main motivation problem (cross-coupling optimization). Ablation experiments further confirm that structured execution feedback is the main driver of performance improvement.