VETR (Verified Execution Trajectories for Refinement) is a framework that enables visual agents to recursively self‑improve over executable world representations. Existing symbolic graphics reasoning agents rely only on final renders for supervision, which limits further improvement. In each task VETR keeps aligned program‑render pairs together with bounded runtime observations, and introduces an execution refiner. The refiner compiles observed writes and state changes into constrained alternative programs, then validates their end‑state and preservation conditions against the captured runtime state. A standard visual verifier finally judges whether the alternatives achieve the goal. Repair construction does not invoke additional language‑model calls, although it may require extra executions. Evaluations on BlenderGym (280 tasks) and BlenderBench (27 released tasks) using GPT‑5.6‑luna and Qwen3.8‑27B show that VETR consistently outperforms prior refinement methods such as BlenderAlchemy and VIGA. Controlled ablations under fixed generator‑interaction budgets demonstrate the separate contributions of execution refinement and role‑asymmetric evidence organization.
Review