Workflow Gym introduces a realistic, long‑horizon benchmark covering 56 professional applications and 338 real‑world workflows, revealing that top GUI agents like Gemini 3.1 Pro achieve only about 30% single‑run success and exposing key failure modes such as consistency breaks and lack of domain knowledge.
The article analyzes why GUI agents trained on both desktop (mouse‑click) and mobile (touch) interactions suffer from behavior collapse and catastrophic forgetting, introduces the UI‑MOPD framework that assigns platform‑specific teachers through on‑policy distillation, and shows an 8B model surpassing a 235B baseline on OSWorld and MobileWorld benchmarks while preserving general GUI understanding.
The SE‑GA framework introduces hierarchical memory and a two‑stage self‑evolution training pipeline that enable GUI agents to retain critical context, learn from past successes, and achieve state‑of‑the‑art performance on long‑horizon benchmarks.
The paper introduces LaSM (Layer‑wise Scaling Mechanism), a training‑free weight‑scaling patch that restores GUI agents’ focus by selectively amplifying attention and MLP weights in critical middle layers, effectively defending against pop‑up‑style environment injection attacks without retraining or altering model architecture.
The author explains why he transitioned from search advertising algorithms to foundation model research, outlines the four typical activities of base‑model teams, and shares detailed technical insights, experimental practices, and scaling strategies that led the EvoCUA GUI Agent to achieve open‑source SOTA on OSWorld.
The article compiles insights from the AIR 2025 conference and related talks, covering the evolution of agents from reinforcement‑learning to LLM‑driven systems, novel agent architectures like AIDE, GUI agents, natural‑language reinforcement learning, and scaling advances in large language models such as Qwen, while highlighting key algorithms, benchmarks, and open research questions.
