NTU's ExpActivator boosts GUI agent step success 28% by activating only applicable experience
NanyangTechnologicalUniversity · hf · 2026-10-13
Researchers from NTU & Alibaba show retrieved history for personal GUI agents mainly helps the first step, offering little benefit over the remaining 90% of steps—relevance at task level doesn't imply applicability at decision time. Their training-free ExpActivator matches each screen to historical states in the frozen backbone's latent space and activates recurring intents only when context supports it: +28% within-trajectory step success across four backbones, 1/5 the history tokens, and 2.3x MCC of the best proactive baseline.
More from coding & agent
- Open-source Claude people-search skill delivers verified leads at $0.029 each — ramagetime · 2026-10-13
- micro1 launches virtual Eval Hackathon on Oct 30 with $10,000 prize pool — Exp_Mark · 2026-10-13
- Dev builds Bops: business agents with their own computer, phone number and email — TheMoonMidas · 2026-10-13
- ARCS benchmark: ambiguous text-to-SQL stumps models, best at just 51% accuracy — megagonlabs · 2026-10-13
- Alibaba's TerminalHorizon synthesizes long-horizon agent data, lifting Qwen3.5-27B across 12 benchmarks — alibabagroup · 2026-10-13
- Sentry CEO: AI codes fine, but Bevy's 3D gaps like bone mesh systems hurt — zeeg · 2026-10-13