TL;DR of the paper, generated automatically.

Framework that tests visual assistants against simulated users, so replies affect what happens next.

• Benchmark: IVG-Bench, 315 scenarios: cooking, household tasks, mobile interfaces.
• Results: best model, Gemini 4 Argon, completes 53.4% of tasks within the time limit.
• Training: proof-of-concept GRPO run improved Qwen3.5-4B's cooking results.