Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
The system beat GPT-5.5 and Seed-2.0-Pro on a mobile-and-desktop benchmark by pairing memory with a skill library instead of a bigger model.
Why now: Covered in the July 16 briefing on GUI agent research.
A system called KnowAct-GUIClaw scored 64.1% on the MobileWorld benchmark, beating closed-source Seed-2.0-Pro and GPT-5.5, per the paper posted at arXiv 2607.12625. It was tested across Android, iOS, HarmonyOS, and Windows.
The number that matters isn't the leaderboard spot. It's the 8.5% score bump the paper reports when learned skills transfer to a different base model than the one that trained them. That's the detail with stakes for anyone building GUI agents instead of API-based ones. It suggests the skill library survives a model swap even when the model underneath doesn't.
Most GUI agent work treats skills as tied to the model that learned them. Swap the model and you start over. This paper pairs an experience-attributable memory system with a pluggable, self-evolving skill library. The library is built to work as a separate asset from the model running it.
That distinction matters for lock-in. Picking a base model has felt like a long-term bet, one where you inherit whatever skills the agent built along the way. Portable skills would break that link. You could swap to a better model without throwing out what the agent already learned to do.
I'd want to see this replicate outside the paper's own benchmark before trusting it. MobileWorld is one test suite. An 8.5% transfer gain there isn't proof skills survive a jump from a Claude-based agent to a Gemini-based one in production. If cross-model transfer numbers start showing up as a standard metric in other GUI agent papers, that's the sign this is real.
Each link below shares sources, entities, or timing with this story.
Claude Code supports Android / Shared entity: GPT / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Code supports Android); both cover GPT; overlapping topics (chang, different, model).
Linked by a graph relationship (Claude Code supports Android); both cover GPT; overlapping topics (different, model).
Linked by a graph relationship (Claude Code supports Android); both cover GPT; overlapping topics (beating, model).
Linked by a graph relationship (Claude Code supports Android); both cover GPT; overlapping topics (base, model).
Linked by a graph relationship (Claude Code supports Android); both cover GPT; overlapping topics (different, model).
Linked by a graph relationship (Claude Code supports Android); both cover GPT; overlapping topics (different, model).
Claude Code supports Android / Shared entity: GPT / Same source domain / Earlier coverage
Linked by a graph relationship (Claude Code supports Android); both cover GPT; reported by the same outlet (arxiv.org).
Claude Code supports Android / Shared entity: GPT / Earlier coverage / Tension
Linked by a graph relationship (Claude Code supports Android); both cover GPT; earlier GPT coverage from 2026-07-10.