RESEARCH · RESEARCH · #921
GUI-SD-v2 extends on-policy self-distillation to multi-turn GUI agents
arXiv paper (2609.27307v1) introduces GUI-SD-v2, a two-stage training framework that extends on-policy self-distillation (OPSD) from GUI grounding to multi-turn GUI interaction by jointly optimizing rollouts with and without privileged guidance and selectively distilling step-specific reasoning and memory guidance. The authors report that GUI-SD-v2 improves Pass@1 and Pass@3 over OPSD baselines and evaluated state-of-the-art methods on the AndroidWorld and MobileWorld benchmarks; code and training data are slated for public release.
KEY POINTS
- arXiv paper (2609.27307v1) introduces GUI-SD-v2, a two-stage training framework that extends on-policy self-distillation (OPSD) from GUI grounding to multi-turn GUI interaction by jointly optimizing rollouts with and without privileged guidance and selectively distilling step-specific reasoning and memory guidance.
- The authors report that GUI-SD-v2 improves Pass@1 and Pass@3 over OPSD baselines and evaluated state-of-the-art methods on the AndroidWorld and MobileWorld benchmarks; code and training data are slated for public release.
- Improving OPSD's privilege-following and selective distillation for multi-turn interactions can yield more reliable action choices and long-horizon memory in GUI agents, advancing practical interactive agent performance.
WHY IT MATTERS
Improving OPSD's privilege-following and selective distillation for multi-turn interactions can yield more reliable action choices and long-horizon memory in GUI agents, advancing practical interactive agent performance.