KnowAct-GUIClaw Hits 64.1% SR on MobileWorld, Beats GPT-5.5

A new agent-framework paper from the Lychee Team at Harbin Institute of Technology, Shenzhen puts a concrete number on how much the memory-plus-skill-library story is actually worth on mobile GUI tasks. Running on the open-source Kimi-K2.6 model, their KnowAct-GUIClaw system reportedly hits 64.1% on the MobileWorld GUI-only benchmark, ahead of Seed-2.0-Pro (63.2%), GPT-5.5 (62.4%), Claude-Opus-4.7 (56.4%) and a raw Kimi-K2.6 baseline (55.6%). The authors report it beats every closed-source agentical model in their comparison table.

The mechanism is a four-stage loop the authors call Know-Route-Act-Reflect. A host agent decomposes a long-horizon request, a lightweight GUI subagent called GUIClaw does the actual tapping and swiping over a hybrid action space of raw GUI primitives, distilled skills, validated Android deeplinks and intents, and intervention calls back to the user; a reflection stage after each run turns useful traces into an experience-memory entry or a state-validated skill. Two persistent stores, a memory-and-history store and a skills-and-shortcuts store, feed the next task. Extending OpenClaw is the framing: the paper argues OpenClaw-style local assistants lack both cross-platform GUI support and any built-in self-evolution mechanism.

The transfer result is the one worth watching. Distilling memory and skills from Kimi-K2.6 trajectories and dropping them into a Qwen3.5-35B-A3B executor lifts that model's score from 24.8% to 41.0% on the same benchmark, a 16.2-point gain, with no weight update. Cross-platform runs report 76.2% on 63 HarmonyOS tasks and 70.0% on 30 Windows desktop tasks spanning browser, file, office, terminal, and system-control workflows.

The honest caveat is that these are the authors' own runs, not third-party replications, and the strongest numbers use a specific configuration where Kimi-K2.6 serves as both the host and the GUI executor. What the paper does not give you is the compute cost of building the transferable skill library in the first place, whether persistent user-profile memory holds up in privacy-sensitive deployments, or how the state-validated skills degrade when target apps push UI updates. What it does establish, at least in the Kimi-to-Qwen direction, is that reusable skills and distilled experience genuinely travel across model families, which is the part that shifts where the next round of GUI-agent work will concentrate.