Relevance Does Not Imply Applicability: Experience Activation for Personal GUI Agents
Read the paperThis item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.
ExpActivator gates GUI history by the current screen and recurring intents by time and scenario evidence.
What it does
Without retraining the GUI backbone, ExpActivator matches a current screen to historical latent states and supplies the corresponding action as a reference. It activates a recurring intent only when current time and scenario support it, otherwise abstaining.
Why it matters
The study finds relevant history helps the opening decision but offers little sustained benefit over the remaining 90% of steps. Across four backbones, it reports 28% average within-trajectory step improvement, about one-fifth the history tokens and approximately 2.3 times the strongest proactive baseline’s MCC.
How it applies
Compare an activation gate with the existing history strategy on held-out routines. Measure step success, inappropriate proactive actions, history tokens and actual end-to-end latency/cost separately. The token reduction does not itself establish a particular bill saving.
The short version
Should this past action apply to the screen in front of the agent now?
My take
A useful trial should count both successful steps and wrongly activated routines. Token efficiency is helpful only when it preserves the evidence needed for the next decision.
How it connects
- Selective retrieval and abstention
- Temporal evidence for proactive actions
- Token footprint versus billed cost
Bottom line
Evaluate activation and abstention on held-out routines before changing the history pipeline.
Takeaways
- Task-level relevance is different from decision-time applicability.
- Screen matching and recurring-intent activation solve different decisions.
- Abstention is part of the proactive decision policy.
- The reported history footprint is about one-fifth in the tested comparison.
- API price and wall-clock savings require their own target measurements.
Historical Brendon Score: 8.1/10
Recorded for a previous shelf selection. Editorial corrections do not constitute a new assessment.
- Quality: 7.5/10 — base
- Authority: 7.0/10 — +0.20
- Freshness: 4.8/10 — +0.00
- Engagement: 7.7/10 — +0.27
- Relevance: 8.0/10 — +0.00
- Corroboration: 1.9/10 — +0.14
- Sum: 8.11
- Total (rounded): 8.1/10
Why this is here
Independently surfaced by 1 community: HuggingFace Papers.
First seen: .