Article written to `papers/procedural-graphs-self-evolving-execution-structures.html`. 2,228 words covering the problem of procedural incoherence in LLM agents, the graph representation, generative guidance mechanics, the four-step self-evolution loop with rejection memory, results across 7 benchmarks and 4 LLMs (21 of 24 first-place finishes), EnterpriseArena survival analysis, scratch versus expert-prior construction, efficiency ablations, and the comparison with seven baseline methods. All numbers match the paper.

Read the paper on arXiv