- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies device-cloud GUI grounding, where an on-device model offloads uncertain predictions to a cloud model. The proposed method, SCOPE, combines two scores. The scale-profile score measures the variation of top-1/top-5 log-probability gaps across coordinate digits, gated by the mean decoding entropy. The view-consistency score measures the displacement between the original prediction and a re-grounded prediction on a local crop. The combined score is calibrated with split conformal prediction into a click radius measured in target-element radii, and inputs whose radius exceeds one are offloaded. Across three benchmarks and three edge models, the method achieves a mean AUCG of 0.785, versus 0.708 for the strongest baseline, and offloads 58.4% of inputs at α = 0.1.
Critical strengths
-
The scale-profile score reuses decoding logits without extra inference and achieves a mean AUCG of 0.720, outperforming all baselines. The full method adds 0.20× extra inference cost. Experiments cover 9 configurations and 100 random splits.
-
Calibration normalized by element size, and useful failure analysis. Expressing the conformal radius in units of the target-element radius is meaningful. Table 12 shows that wrong predictions confirmed by re-grounding account for 37.9% of all errors but 81.0% of the errors retained on-device.
-
The paper is well written, and easy to follow
Critical weaknesses
-
PSS already uses coordinate-digit logits for confidence estimation. AutoFocus (arXiv:2605.02630) uses coordinate-token perplexity together with crop-based re-grounding. PACE (arXiv:2609.24277) explicitly weights per-digit entropy by place value. Neither AutoFocus nor PACE is cited or compared. Moreover, the steepness term is a standard deviation across digit positions and is therefore invariant to permutations of those positions. It cannot distinguish uncertainty in higher-order digits from uncertainty in lower-order digits, which is the paper's own motivation. Comparisons with simple alternatives, such as coordinate-only entropy or place-weighted entropy, are missing.
-
The guarantee does not match click correctness, which tilts the SafeGround comparison in SCOPE's favor. Eq. (3) bounds , where is half the box diagonal. Clicks outside the target box but within of it are therefore not counted as violations, even though correctness requires clicking inside the box. The guarantee is also marginal over all inputs and does not bound the error rate among retained predictions. SafeGround, by contrast, controls the false discovery rate (FDR) among accepted predictions, which is a stricter guarantee. Applying the proposed rule to SafeGround's scores reduces its offloading rate from 94.2% to 74.3% (Table 3). Meanwhile, SafeGround's native rule retains predictions with higher accuracy (0.911 vs. 0.891 for SCOPE). Much of the reported 58% versus 94% gap therefore stems from the different guarantees rather than from better uncertainty estimation.
-
The deployment results contradict the device-cloud motivation. Table 4 shows that cloud-only inference is more accurate than SCOPE (0.782 vs. 0.773), uses less computation (27.9 vs. 46.5 edge + 18.3 cloud TFLOPs), and has lower latency (3.99 vs. 5.98 s). The scale-profile-only fallback suggested in §4.4 does not fix this: at its 74.3% offloading rate, we estimate its latency at roughly 2.17 + 0.743 × 3.99 ≈ 5.1 s, which is still slower than cloud-only. On ScreenSpot-Pro, 97.3–100% of inputs are offloaded (Table 7). The chosen cloud model activates only about 3B parameters per token, and no monetary cost comparison is provided.
Questions for authors
-
What proportion of incorrect predictions satisfies and therefore does not violate the constraint in Eq. (3)? What is this proportion among the retained predictions?
-
Can the authors compare with SafeGround under the same type of guarantee, either by applying FDR-controlled calibration to their scores or by comparing retained accuracy at matched offloading rates?
-
Can the authors compare against PACE, coordinate-token-only entropy, and a higher-order-digit-only score? How does the scale-profile score behave on samples whose higher-order digits are uncertain but whose lower-order digits are confident?
-
Why was Qwen3.5-35B-A3B chosen as the cloud model? Would the method still reduce cost with a substantially stronger and more expensive cloud model, when monetary cost is measured?