The authors train two small code models (a 100M-parameter Rainbow-Pony-100M and a fine-tuned Qwen2.5-Coder-0.5B) on Flutter/Dart code in both direct whole-file generation and iterative diff/search-replace regimes, evaluating on roughly 1,790 held-out tasks each. Direct generation wins broadly on compilation pass rate, bits-per-byte, character similarity, and LLM-judge scores, but diff-based generation is competitive and sometimes better for short, spatially localized edits like refactors and error fixes — a limitation the authors term ‘task locality.’
