When one function's correctness depends on another's runtime behavior, structured execution-trace diagnostics beat Reflexion and OpenEvolve on 7 of 8 problems
The paper names static binding as the failure where meaning is defined by joint execution behavior across components and cannot be resolved from textual descriptions, and shows it appears in packing, routing and symbolic search, not just explicitly coupled optimizers. Dynamic context adaptation runs a validation agent that extracts structured diagnostics from execution traces as gradient-like guidance to a generation agent proposing multiple candidates per iteration, with a knowledge graph from the problem description supplying semantic constraints and simulated annealing preventing greedy collapse. It beat zero-shot, Reflexion and OpenEvolve on seven of eight problems at both 300 and 600 evaluations (p < 0.01), and the ablation confirms structured execution feedback, not the search wrapper, is the driver.
↳ Follow the thread