- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper observes that mobile GUI agents waste most of their latency on redundant per-step chain-of-thought tokens, and proposes to replace that repeated reasoning with "muscle memory": reusable reduction recipes, learned offline from past successful executions, that crop the screen to the few relevant GUI elements and are invoked at runtime by a lightweight dual-gated (OOD-distance + learned scorer) trigger. MEMOIR thereby compresses both the VLM's input and its thinking output, cutting per-step latency substantially on AndroidControl while slightly improving step accuracy.
Critical strengths
Accurate problem framing: identifying output-side thinking tokens as the neglected real bottleneck. Prior pruning- and shortcut-based methods compress only the input (visual tokens, multi-step actions). By measuring the prefill/decode throughput gap and breaking down output tokens, the authors show that thinking tokens make up over 84% of the output and that existing methods barely reduce output tokens — explaining why earlier speedups fall short. The contribution is a shift in perspective: from "letting the VLM see less" to "letting the VLM think less."
Critical weaknesses
-
"Without sacrificing accuracy" is thinly supported. High-level success moves only 10.8% → 11.3% over 1,543 episodes with no significance test, and the aggressive calibration setting collapses accuracy to 63.5%, so accuracy preservation looks conditionally tuned rather than guaranteed. The claim that cropping reduces "distraction" is also unverified: no controls (cropping with thinking preserved, random crops of equal size, shuffled candidate order) rule out alternative explanations.
-
Only per-step speedup is demonstrated, not faster task completion. Because evaluation is offline, the agent never executes actions: inputs are recorded screenshots and episode length is fixed to the human trajectory, so retrying and error correction cannot occur. Latency evidence is therefore per-step only, and the sole end-to-end-sounding claim ("cuts end-to-end latency by 3.25×") is drawn directly from that per-step column.
-
The memory library is built from successful trajectories and validated with ground truth, so in a static dataset the state always lies in-distribution; the reported hit rate is an artifact of the protocol, and the policy-induced distribution shift that threatens this method in real execution is never tested.
Questions for authors
How would the memory hit rate change under interactive execution?
The memory library is built from past successful trajectories, and the appendix states that ground-truth information is used during the offline filtering stage. In a static dataset the current state always lies within the seen distribution, so the reported 92.4% hit rate can be regarded as an artifact of the protocol. Under real execution, where the agent's own actions alter the interface and produce states absent from human trajectories, by how much would the b_full fallback frequency rise, and how far would the speedup degrade?