Publications
カンファレンス (国際) Is Feedback All You Need? Benchmarking Random Search vs. Iterative Refinement for Parametric Icon Design
Naoki Kimura
The 31st International Conference on Intelligent User Interfaces (ACM IUI 2026)
2026.7.16
Vision–language models (VLMs) can score images against natural-language intents. This capability makes “self-refining” design agents possible: a candidate is generated, feedback is produced, and the candidate is revised. For low-dimensional parametric graphics (e.g., facial icons), it is unclear whether this expensive loop beats a simple generate-and-filter workflow. We test VLM-guided parametric icon design under a strict budget of N = 5 distinct candidates per run. Two strategies are compared: Random Top-k (sample N candidates and select the best) and a rule-based Self-refine loop (revise one candidate using VLM feedback). Using GPT-4o across four emotion tasks, Random Top-k matches or outperforms Self-refine (Joy +0.4, Sadness +4.4; ties on Anger/Neutral). We observe threshold-like score jumps rather than smooth gradients, which makes local refinement unreliable. These results suggest that for constrained parametric design, a lightweight generate-and-filter interface can be more robust and efficient than conversational iterative refinement.
Paper :
Is Feedback All You Need? Benchmarking Random Search vs. Iterative Refinement for Parametric Icon Design
（外部サイト）