ByteDance's WorkflowGym Exposes GUI Agents: Top Models Pass Only 30%
机器之心 · wechat · 2026-07-23
ByteDance's Seed evaluation team has released WorkflowGym, a new benchmark that tests GUI agents on professional software like FreeCAD and Blender, revealing the true boundaries of current model capabilities.
Evaluation Design & Results:
- The benchmark covers 56 professional applications and 338 real-world, long-horizon workflows, averaging over 100 operational steps.
- In these rigorous tasks, top-tier models like Gemini 3.1 Pro and Kimi K2.6 achieved a single-pass success rate of just over 30%. Lighter models almost completely failed on tasks requiring more than 60 steps.
Core Model Flaws:
The paper identifies four major dead spots in current agent frameworks: loss of long-context consistency, cascading errors without self-correction, goal drift, and a lack of understanding of professional UI paradigms. This suggests that scaling parameters alone won't fix these architectural bottlenecks, making human-AI collaboration the most practical product approach for now.
More from Models
- Gemma 4 26B-A4B runs offline on a 16 GB Mac: 33/40 on TypeScript tasks at 31 tok/s — Mission-Sir-1614 · 2026-10-26
- Claude safety classifier blocks SS7 attack analysis, dev pivots to abliterated GLM — l4rz · 2026-10-26
- Sakana AI to pretrain first domestic model Sakana Sawara for defense and cybersecurity by 2027 — tkasasagi · 2026-10-26
- 31M-param English TTS model phonon-31m-en trends on Hugging Face — gradium · 2026-10-26
- NVIDIA tutorial: post-training Nemotron for Europe with Persona datasets and LoRA — NVIDIA Developer · 2026-10-26
- Opus 5 hailed as the 'most emo, greatest artist' model by AI community users — repligate · 2026-10-26