Visual agents improve symbolic graphics reasoning using verified execution traces
VETR: Execution-Grounded Self-Improving Agent for Symbolic Graphics Reasoning
Artificial Intelligence
Summary
It is hard for computer programs to understand and fix drawings made from symbols because they only see the final picture, not how it was made. The authors created a new method called VETR that helps these programs learn by looking closely at each step they take to change the drawing. By checking what worked and what didn’t during their attempts, the method helps the program improve without extra guesswork. When tested on many drawing tasks using advanced language models, VETR did better than older methods that tried to fix drawings.
What this means in practice
- •For graphics software developers: Improve automated correction tools for procedural graphic design by using execution-feedback to refine symbolic scene representations.
- •For ai system integrators: Build more reliable visual reasoning agents that self-improve by verifying edits through runtime execution in interactive settings.
- •For technical content creators: Enable tools that generate more accurate symbolic graphics by leveraging execution-grounded refinements without additional language model calls.$Commercial implications: Makes it possible to offer AI-assisted graphic creation tools with fewer errors, enhancing creative workflows for designers and artists.
Authors
Yingbo Ma, Kuangshi Ai, Gordon Dai, Chenglong Yu, Chaoli Wang, Hua Shen
Abstract
Visual agents for symbolic graphics reasoning depend on executable world representations, yet continued improvement remains bottlenecked by ambiguous supervision from final renders. Recent agentic systems refine scenes through repeated interaction, but they lack mechanisms to recursively extract actionable evidence from execution traces, overlooking which proposed symbolic edits survived, disappeared, or interfered during program execution. To address this, we propose VETR (Verified Execution Trajectories for Refinement), a framework that enables agentic recursive self-improvement over executable world representations for symbolic graphics reasoning. Within each task, VETR retains aligned program/render records with bounded runtime observations and introduces an execution refiner that compiles observed writes and state changes into constrained alternative programs. The refiner validates their endpoints and preservation conditions against captured runtime state; the ordinary visual verifier then assesses goal relevance. Repair construction makes no additional language-model calls, but can require additional executions. Evaluations on BlenderGym (280 tasks) and BlenderBench (27 released tasks) with GPT-5.6-luna and Qwen3.8-27B show that VETR consistently outperforms existing refinement strategies, including BlenderAlchemy and VIGA. Controlled ablations evaluate the contributions of execution refinement and role-asymmetric evidence organization under fixed generator-interaction caps.