Release · Sep 30, 2026
Einsia AI's Navers Lab Releases PPTBench, a Visual Coding Benchmark for Coding Agents
PPTBench, a new visual coding benchmark released by Einsia AI's Navers Lab, asks coding agents to recreate flowcharts and architecture diagrams from real papers as editable PPTX files using native PowerPoint objects. It spans 500 tasks across 13 fields. The evaluation pairs an Agentic Judge scoring semantics, readability, and visual similarity with an Artifact Gate, covering 10 models, 36 configurations, and 54,000 reviews. GPT-6 Astra High led with 77.34 points, ahead of Kimi K3 High at 67.80; GPT-5.6 Sol Max and Qwen 3.8 Max XHigh scored 49.28 and 47.69. Flow-semantics errors were the main source of lost points.
Original sources (Chinese)
GPT-6 Astra也会「看错」流程图? Einsia AI发布PPTBench，视觉编程还有多远？