A final-image score cannot identify the failing module
The DARS preprint, submitted August 20, models instruction-based image editing as a VLM planner followed by a diffusion renderer. A poor final image does not reveal whether training should focus on the edit plan or rendering, nor which decision inside a free-form plan caused the problem.
DARS compares reward variation across multiple plans and multiple renders to route optimization softly between the two modules. The planner emits four structured fields, allowing prefix-gated rewards and token-level advantage reweighting to localize outcome feedback. The authors report that DARS outperformed a Joint RL baseline across five benchmarks with the same backbone, data, reward model and rollout budget, with its largest gains on reasoning-intensive edits.
Boundary before product claims
This is a preprint, and the abstract does not provide absolute scores or real-user preference results. Improvement on selected benchmarks does not establish performance on commercial editing, identity preservation, text accuracy or safety. The notable contribution is the training architecture that separates planning failure from execution failure rather than treating one final score as sufficient supervision.