acceptodds
Under review as a conference paper at ICLR 2027

TopoRerank: Bounded Structural Reranking for Fixed-Candidate GUI Action Selection

Abstract

Graphical user interface (GUI) agents may choose the wrong control even when the target is present, as semantically similar candidates receive close scores. End-to-end accuracy conflates candidate coverage, ranking, target assignment, and final selection. To isolate this decision, we propose TopoRerank under a protocol that fixes the candidate set, Stage–1 scores, target matcher, split, and evaluator. The method adds a bounded, instruction-independent residual from candidate-local accessibility-tree features, changing only close competitors. On package-disjoint AndroidControl, it improves a text-retrieval baseline by 3.62 percentage points across three seeds, with a positive package-equal interval; the gain falls to 0.53 percentage points under a stronger learned text-only ranker. An area-free nearest-center control is positive under the primary retrieval baseline but inconclusive under lexical scoring; app-disjoint GUIrilla results are positive after target-domain refitting. Overall, local hierarchy offers a conditional correction to frozen rankings, not a replacement for semantic ranking or candidate recall.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI grounding, GUI agents, GUI action selection, accessibility trees, structural reranking, learning to rank, structured interface understanding, cross-platform replication, fixed-candidate protocol

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vpvwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether local accessibility-tree features improve GUI action selection when candidates and upstream scores are fixed. TopoRerank adds a bounded, instruction-independent residual using a 209-parameter MLP. On AndroidControl, it improves BM25 by 3.62 percentage points and a learned text-only ranker by 0.53 points. The paper also reports area-free matcher controls, structural ablations, an external UGround reference, and positive GUIrilla results after target-domain refitting.

Critical strengths
  1. The fixed-candidate protocol isolates final selection from candidate coverage and upstream ranking. Three-seed results, package-level bootstrap analyses, matcher controls, and explicit reporting of null results provide credible support for the narrowly framed empirical claim.

  2. The paper includes meaningful alternative explanations and baselines. The area-free BM25 control remains positive, GUIrilla provides a positive refitted result with target node identifiers, and Table A22 compares against bounded linear listwise and pairwise residuals. The authors appropriately qualify transfer and end-to-end applicability.

Critical weaknesses
  1. The methodological advance is modest. The paper already discusses structure-aware GUI models and candidate-ranking methods such as RelationAgent, Mind2Web, and RankGround. Its main addition is a bounded MLP correction over local tree and geometry features. The controlled analysis is careful, but the new method and resulting insights offer a limited advance over these established ideas.

  2. The benefit becomes less convincing with the stronger upstream ranker: the gain is 0.533 percentage points, with a package-equal confidence interval crossing zero. Table A21 also gives no clear advantage over other structural corrections. Although the head outperforms linear controls under lexical scoring, its usefulness beyond that simpler setting remains insufficiently established.

Questions for authors
  1. What is the package-equal gain and 95% confidence interval on top of the learned text-only ranker? A paired comparison with the controls in Table A21 would help interpret the small differences.

  2. Table A22 shows an advantage over linear corrections under lexical scoring, but this advantage is less clear with the learned upstream. What do the authors see as the main benefit of the proposed head in that setting?

3 comments
AuthorsOpenReview ↗

Response to the reviewer: learned upstream and correction controls

Thank you for the careful assessment of the method and alternative explanations. We added a stronger learned upstream and paired comparisons to clarify the contribution.

Requested learned-upstream interval. Replaying the three saved hashed-token checkpoints on GPU reproduces the reported per-seed baseline/corrected hit counts and hit-change counts. The package-equal change is +0.71 pp [95% CI +0.20, +1.25]. The CPU package-equal point estimate agrees within 0.01 pp; row-level runtime differences are described in Appendix H.4. These checkpoint-replay intervals resample 78 packages conditional on fitted checkpoints (revised Table 1).

The submission stated that this interval crosses zero; we cannot verify that statement from the available archived aggregate results. The revised interval is the new checkpoint-replay estimate, not a reconstruction of the original row-level predictions.

Paired comparison on the same learned upstream. We also replayed the saved pointwise-deep control from the submitted Table A21 (revised Table A22) against TopoRerank on the same test rows and frozen hashed-token scores. The CPU checkpoint replay gives MLP minus pointwise-deep −0.05 pp package-equal [95% CI −0.63, +0.50], with a −0.07 pp row-weighted difference over three seeds. The interval includes zero. The pointwise checkpoints reproduce the reported final hit counts; some baseline hit and ranking-change counts differ in the current runtime (revised Appendix H.4).

Benefit above a stronger learned scorer. A MindAct-initialized DeBERTa text scorer fine-tuned on AndroidControl reaches 48.29% top-1; adding the structural correction reaches 50.49%, or +2.20 pp row-weighted and +1.31 pp package-equal [95% CI +0.69, +2.03], across three seeds with fixed candidates and evaluation (revised Table 1). This is a distinct, stronger text-only operating point, not a replacement for the hashed-token setting.

On the hashed-token upstream, the correction has a positive package-equal replay estimate, while the paired comparison does not establish an advantage over the pointwise-deep head. The DeBERTa experiment tests a separate, stronger text-only operating point.

What the correction adds. TopoRerank adds candidate-local structural information with a 209-parameter, instruction-independent head, without retraining the frozen scorer or changing its candidates. Because each correction has magnitude at most B, an upstream score gap greater than 2B cannot be reversed. This explicit bound makes the intervention inspectable: close decisions can change while sufficiently separated pairs retain their order. The same guarantee applies to the bounded pointwise control; the replay does not establish an accuracy advantage for either head. The matched operating-point and matcher controls test when bounded structural correction translates into accuracy gains.

On the hashed-token upstream, the topology-shuffle control gives a +0.03 pp row-weighted gain, compared with +0.53 pp for the unshuffled head. This contrast is consistent with sensitivity to the tested structural features at that operating point.

Relation to the closest work. RelationAgent models spatial, functional, and hierarchical relations in a grounding–relation–navigation pipeline; RankGround ranks screenshot crops for downstream grounding; Mind2Web studies candidate-based web action selection. TopoRerank studies an already-proposed accessibility-tree candidate set: upstream scores and target matching are held fixed while an instruction-independent bounded residual changes the final ordering. This design isolates the contribution of candidate-local structure to selection.

Simple heads and the requested comparisons. At BM25 with the released matcher, the MLP minus a train-fitted linear prior over child count, depth, and area is −0.07 pp package-equal [95% CI −1.19, +0.86]; MLP minus a ten-feature linear head is +0.20 pp [−0.38, +0.74]. These are BM25 comparisons; both intervals include zero. Together with the DeBERTa result, they demonstrate the value of evaluating the structural intervention across operating points and using paired controls to assess head design.

Correction to the tie analysis. We also corrected Appendix F: the learned tie-only gain is approximately +10.04 pp on the 1,195 BM25-tied rows, versus +5.523 pp for smallest-area ordering on those same rows; +2.739 pp used all 4,381 rows as denominator.

The checkpoint interval, paired head comparison, and stronger text-scorer experiment address your questions about the learned upstream, alternative corrections, and the setting in which the head adds value.

Thank you for the additional experiments. The checkpoint-replay interval for the learned text-only upstream and the new DeBERTa result (+1.31 pp package-equal, [+0.69, +2.03]) show that the correction can still help above a stronger text ranker, and I appreciate the honest reporting. However, the paired comparisons also show no clear advantage of the proposed head over the pointwise-deep control (−0.05 pp, interval includes zero) or over simple linear heads at BM25. This supports my original concern: the main contribution is a bounded structural correction whose benefit can largely be obtained with simpler heads, which is a limited advance over existing work. I will keep my score.

AuthorsOpenReview ↗

Clarifying the contribution of bounded structural correction

Thank you for considering the additional experiments and recognizing the stronger text-upstream evidence. Your distinction between the structural correction and the particular head architecture helps clarify our contribution. The paired comparisons do not show a clear package-level advantage of the MLP over the pointwise-deep control or the tested linear heads.

We view the contribution as a bounded, instruction-independent structural intervention on frozen candidate scores, together with controlled evidence of when that intervention changes selection. The fitted correction can be attached without retraining the upstream scorer or changing its candidates. Its bound makes the intervention inspectable: an upstream score gap greater than twice the bound cannot be reversed, while close candidates can be reordered using local structure. The simpler bounded heads share this guarantee. At the tested BM25 operating point, the three-feature linear prior recovers most of the observed package-equal gain.

The DeBERTa experiment provides a stronger operating point, with a +1.31 pp package-equal gain [95% CI +0.69, +2.03] over its frozen text scorer. Together with the matched prior and matcher controls, it identifies a setting where structural correction is useful; the choice of head parameterization remains an empirical question. We would welcome your assessment of whether this more precise positioning addresses your concern about the scope of the contribution.

Reviewer_wJZVOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TopoRerank adds a bounded, instruction-independent MLP residual from local hierarchy and geometry features to frozen GUI candidate scores. Holding candidates, matching, and evaluation fixed, it improves AndroidControl accuracy by 3.62 pp over BM25 and 0.53 pp over a learned text ranker. Matcher sensitivity, structural controls, and GUIrilla refitting experiments examine the correction’s dependence on upstream quality and evaluation conditions.

Presentation: 3 — Clear and easy to follow. The formulation and experimental scope are clear. The BM25 tie-analysis inconsistency below requires correction.

Informativeness: 2 — Learned something interesting or useful. The study provides observations about structural priors under weak upstream ranking. Their implications for stronger GUI grounding systems remain limited.

Soundness: 2 — Minor or fixable concerns. The controlled comparisons support the narrow BM25 result. Concerns remain about interpreting structural gains, evidence under stronger upstreams, and inconsistent denominators in the tie analysis.

Rating: 1 — Clear rejection. The submission does not establish sufficient methodological novelty or a substantial empirical contribution relevant to stronger GUI grounding systems. The conditional BM25 improvement is credible, but diminishing gains under the learned upstream and incomplete comparisons with simpler structural controls leave the overall contribution insufficient for acceptance.

Critical strengths
  1. The paper provides controlled evaluation. Fixing candidates, scores, matching, and evaluation isolates ranking improvements. Reporting row-weighted and package-equal estimates exposes application heterogeneity.

  2. The results support robustness evidence. BM25 achieves a positive package-equal gain of +1.263 pp, with 95% CI [+0.260, +2.325]. Positive nearest-center and GUIrilla refitting results extend the evidence beyond one matcher and accessibility schema.

Critical weaknesses

1. The experiments do not establish sufficient relevance to contemporary GUI grounding. Most gains occur over lexical or BM25 ranking. Under the learned text ranker, the gain falls to +0.53 pp and its package-equal interval crosses zero. Existing work such as GUI-Actor evaluates candidate verification within VLM-based grounding, while TopoRerank lacks a comparable strong-upstream experiment. Consequently, the paper does not demonstrate a convincing incremental benefit at an upstream quality level relevant to current systems.

2. The methodological contribution is limited, and the empirical evidence does not sufficiently compensate for it. The method combines handcrafted descriptors, a small bounded MLP, and a listwise loss. Bounded linear fusion already achieves +4.246 pp versus +5.691 pp, without uncertainty reported for their difference. Size and hierarchy features also align with the smallest-containing-box matcher. The nearest-center control lacks comparisons against simpler learned priors, leaving the proposed head’s added value insufficiently established.

Questions for authors

Additional validation: Can the correction improve a stronger pretrained semantic or VLM-based ranker under the same fixed-candidate protocol? Paired package-level gains would directly address its relevance to current grounding systems.

Additional validation: How does the MLP compare with bounded linear fusion and a train-fitted leaf/depth/size prior under both BM25 matcher conditions? Please provide intervals for their paired differences.

Clarification/correction: Appendix F compares +2.739 pp with +5.523 pp on BM25 ties, but Table A17’s +2.739 pp uses all rows. Dividing by the tie fraction, 1195/4381, gives approximately +10.04 pp, matching Figure 3B. Please reconcile the denominators and the claimed ordering reversal.

2 comments
AuthorsOpenReview ↗

Response to the reviewer: stronger scorer and matched head controls

Thank you for finding the protocol clear and for recognizing that the matched comparisons support the BM25 result. We added a stronger pretrained semantic upstream and paired simple-head controls, and corrected the tie comparison to address the concerns behind your overall recommendation.

Stronger upstream. A MindAct-initialized DeBERTa-v3-base scorer, fine-tuned on AndroidControl, reaches 48.29% Stage-1 top-1. With the bounded structural residual it reaches 50.49%, a paired improvement of +2.20 pp row-weighted and +1.31 pp package-equal [95% CI +0.69, +2.03] over the same 4,381 test rows, with the same candidates, matcher, split, and evaluator (revised Table 1). The interval is a package bootstrap conditional on the three fitted checkpoints. This establishes a benefit for this pretrained text scorer. DeBERTa is text-only; our visual-upstream evaluations have not established a benefit from the correction under the tested protocols. The 209-parameter correction uses B=0.10, as in the main bounded head, and operates on frozen candidate scores without retraining the scorer. Separately, app-disjoint GUIrilla on macOS also improves after target-domain refitting.

Matched simple heads and matcher sensitivity. At BM25 with the released matcher, the train-fitted linear prior using child count, depth, and area gains +1.72 pp package-equal [95% CI +0.70, +2.92]. The MLP minus this prior is −0.07 pp [−1.19, +0.86]; MLP minus a ten-feature linear head is +0.20 pp [−0.38, +0.74]. Both intervals include zero: simple learned structural corrections are competitive across packages. The three-feature prior combines size and hierarchy, distinct from a pure position-only baseline.

For the lexical comparison you highlighted, the bounded listwise linear head gains +4.246 pp and the tree-position MLP +5.691 pp, a row-weighted difference of +1.445 pp (revised Appendix Table A23). A replay of the saved checkpoints with current code matches both heads' per-seed aggregate accuracies and the MLP's flip counts. Its paired package-equal MLP-minus-linear difference is −0.26 pp [95% CI −1.80, +1.28], resampling 78 packages conditional on the checkpoints. This is a current-source sensitivity analysis, not an exact reconstruction of historical row predictions; the feature-source version differs from the historical MLP audit (revised Appendix H.5).

Under nearest-center matching with area removed, the BM25 area-free head gains +1.84 pp package-equal [95% CI +0.89, +3.04]. This positive result under an alternative matcher and without area reduces the concern that the gain depends only on smallest-containing-box alignment. For a matched simple-head comparison under this matcher, both heads are retrained using the same six tree-only features: the linear head gains +0.89 pp package-equal [−0.12, +1.99] and the MLP +0.67 pp [−0.11, +1.57]. MLP minus linear is +0.63 pp row-weighted [+0.19, +1.07] and −0.22 pp package-equal [−0.93, +0.60]. This six-feature comparison is a different configuration from the preceding area-free head (revised Appendix Table A26; released-matcher controls are in A24–A25).

Across the two matcher settings, the paired package-level comparisons do not establish an MLP advantage over the tested linear heads. The stronger-upstream and matcher controls instead show when a bounded structural correction can improve selection.

What is new? TopoRerank brings bounded, instruction-independent structural correction to fixed GUI candidates and measures its effect across matched rankers and target matchers. The 2B score-gap property specifies exactly which pairwise orders the correction can change. The stronger DeBERTa result and the paired simple-head controls turn this into a testable finding: a lightweight structural intervention improves selection on a strong text-based operating point, while head complexity and target-matching conditions affect the observed gains.

Tie correction. Your reconciliation is correct: +2.739 pp uses all 4,381 rows. On the 1,195 tied rows, the learned tie-only correction is approximately +10.04 pp, compared with +5.523 pp for smallest-area ordering. Appendix F now uses the same denominator for both.

We would appreciate your reassessment in light of these matched results.

AuthorsOpenReview ↗

Follow-up: an exploratory GUI-Actor diagnostic

Visual-upstream evidence. To address the concern about stronger visual grounding, we evaluated TopoRerank with a released GUI-Actor checkpoint. Its screenshot-and-instruction attention scores were mapped to the same accessibility-tree candidates before and after correction. On 442 GUIrilla source-train examples from 106 applications, the bounded correction improved fixed-candidate top-1 accuracy from 61.76% to 64.25%. The app-equal gain was +2.36 pp, with a post-hoc app-bootstrap 95% interval of [+0.84, +4.17], conditional on three fitted heads and one frozen visual checkpoint.

Protocol. The candidate set and visual scores are held fixed, so the comparison measures the effect of adding structural correction at this operating point. The cohort comprises source-train examples with recoverable target geometry, excluding applications in the supplied prior manifests and five applications inspected visually before scoring. We scaled the visual density score by 0.25 before applying the bounded residual; the scale was selected using earlier development analysis, and both scale and head were frozen before this evaluation. This fixed-candidate diagnostic is distinct from GUI-Actor's end-to-end evaluation and GUIrilla's official test split. Some changed cases have possible instruction–target mismatches or ambiguities that require further verification.

Implication. The GUIrilla result extends the evidence beyond text-only upstreams: local structure can improve selection above a visual scorer without changing its candidate set or retraining it. In a separate AndroidControl setting using an Android-trained visual checkpoint and a different density protocol, the raw structural heads decreased accuracy by 2.62 pp on average. Score scale changes the range of gaps a bounded residual can reverse; the two diagnostics differ in several respects, so they do not isolate scale as the cause of their different outcomes.

Reviewer_2QXAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoRerank changes the scores of existing GUI candidates. It adds a bounded residual based on candidate-local accessibility-tree features to fixed Stage–1 scores. The candidate set and target matcher stay the same. The paper tests several Stage–1 rankers and also tests GUIrilla after training the reranker on that platform. The average gain is 3.62 percentage points with BM25 and 0.53 points with a learned text-only ranker.

Critical strengths
  1. The paper states what its results do and do not show. It separates finding a target from ranking an existing target. It also makes clear that the GUIrilla result requires training on that platform. The reported uncertain and negative results help readers understand where the method works.
Critical weaknesses
  1. The method's added value is limited under the stronger ranker. Its larger gains occur with lexical and BM25 scores. With the learned text-only ranker, the gain is 0.533 percentage points, and the package-equal interval includes zero (Section 7; Table A20). Here, package-equal means that each application package has the same weight in the average. The small MLP, listwise objective, and bounded residual are also standard parts. A simple learned position-prior baseline would help show whether the method learns more than a preference for certain candidate locations.
  2. The exact-tie comparison appears to mix two denominators. Page 20 compares a learned gain of 2.739 points with 5.523 points for smallest-area ordering. Table A17 reports 2.739 over all 4,381 examples. Table A11 reports 5.523 only over the 1,195 examples with exact ties. On that same subset, the learned gain is about 10.04 points, as Figure 3 shows. The claim that the learned method loses on exact ties therefore needs correction. This error does not cancel the overall BM25 gain.
Questions for authors
  1. Clarification: please compare the page-20 exact-tie results using the same denominator for both methods.
  2. Additional validation: how does TopoRerank compare with a simple position prior learned from the training data? Please also report the numerical package-equal confidence interval for the learned text-only Stage–1 ranker.
2 comments
AuthorsOpenReview ↗

Response to the reviewer: stronger upstream, prior, and tie correction

Thank you for identifying the denominator issue and asking for a learned prior and package-level analysis. We corrected the comparison and added matched controls and a stronger text scorer.

Exact ties: corrected comparison. On the same 1,195 BM25-tied rows, the learned tie-only residual improves top-1 by approximately +10.04 pp, versus +5.523 pp for smallest-area ordering. We withdraw the submission's claim that smallest-area ordering outperforms the learned residual on exact ties: its comparison used +2.739 pp normalized over all 4,381 rows. On the tied rows, the learned gain is about 1.8 times the heuristic's point estimate. Revised Appendix F uses the common denominator; the +3.62 pp all-row gain of the always-on correction is unchanged.

Learned location prior. We trained a bounded linear prior on two candidate-center coordinates, using the same BM25 scores, candidates, released matcher, split, and three seeds. It gains +1.45 pp row-weighted and +1.30 pp package-equal [95% CI +0.10, +2.58]. The ten-feature MLP minus this prior is +2.36 pp row-weighted [row-bootstrap CI +1.64, +3.07], and +0.34 pp package-equal [package-bootstrap CI −1.07, +1.69] (revised Appendix Tables A24–A25). Location alone provides a useful signal, while the MLP has a larger row-weighted gain in this comparison; the package-level difference remains uncertain. Both the features and head form differ, so this contrast does not isolate the contribution of additional features. Coordinates are normalized by the maximum node-center coordinate per hierarchy, not by measured screen dimensions. Intervals resample rows or packages, respectively, conditional on the fitted checkpoints.

A train-fitted structural prior. On the same BM25 rows and seeds, a linear head using child count, depth, and area gains +1.72 pp package-equal [95% CI +0.70, +2.92]. The MLP minus this head is −0.07 pp [−1.19, +0.86]. The interval includes zero, showing that a compact learned structural prior is a competitive correction across packages. The prior combines size and hierarchy features; it is not a pure screen-position-only baseline (revised Appendix Tables A24–A25).

Learned upstream and the requested interval. For the hashed-token scorer, a GPU replay of the three saved checkpoints reproduces the reported per-seed baseline/corrected hit and hit-change counts. The replay gives +0.71 pp package-equal [95% CI +0.20, +1.25]; the CPU package-equal point estimate agrees within 0.01 pp. Both intervals resample 78 packages conditional on the fitted checkpoints; row-level runtime differences are described in revised Appendix H.4 (see also revised Table 1).

The submission stated that this interval crosses zero; we cannot verify that statement from the available archived aggregate results. The revised interval is the new checkpoint-replay estimate, not a reconstruction of the original row-level predictions.

As a separate stronger text-only operating point, a MindAct-initialized DeBERTa ranker fine-tuned on AndroidControl improves from 48.29% to 50.49% with the correction: +2.20 pp row-weighted; +1.31 pp package-equal [95% CI +0.69, +2.03]. Candidates, matcher, split, and evaluator remain fixed (revised Table 1). This result strengthens the evidence above a substantially stronger learned scorer; it is separate from the hashed-token replay.

Contribution. TopoRerank combines a learned local-structure correction with an explicit bound on its effect: if the original score gap exceeds 2B, two candidates cannot reverse order. It applies directly to fixed candidate scores and makes its decision changes inspectable. The matched controls then examine when the correction helps under different upstreams and target matchers.

Ethics flag. Regarding the ethics review flag, we would be grateful for clarification if there is a specific concern to address. Our experiments evaluate recorded data offline and do not execute actions.

The corrected tie comparison, learned priors, and checkpoint-replay interval address your questions about the denominator, location signal, and learned upstream, respectively.

Area_Chair_n9ZEOpenReview ↗

Reviewer 2QXA] Please specify what is your ethics concern specifically.

Dear Reviewer 2QXA,

Reviewer 2QXA marked “Flag for Ethics Review: Yes,” but the AC cannot find an ethics concern for the review.

Please specify what is your ethics concern specifically if you have.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revisions and new matched controls

Thank you for the careful reviews. We address the shared questions with a stronger protocol-matched upstream, matched simple-head controls, and a correction to the exact-tie comparison. All gains below are percentage-point (pp) changes in top-1 accuracy; package intervals resample the 78 test packages, conditional on the fitted checkpoints.

1. The correction also helps a stronger learned upstream. We fine-tuned a MindAct-initialized DeBERTa-v3-base text ranker on AndroidControl and kept candidates, target matcher, split, and evaluator fixed. Its Stage-1 top-1 is 48.29%, versus 39.44% for the hashed-token ranker. The structural correction raises DeBERTa to 50.49%: +2.20 pp row-weighted and +1.31 pp package-equal, 95% CI [+0.69, +2.03], averaged across three seeds. The previously reported hashed-token gain is +0.53 pp. Thus a smaller correction on one learned upstream does not imply that it vanishes for every stronger upstream. Separately, the app-disjoint GUIrilla evaluation on macOS is positive after target-domain refitting.

2. Matched learned-head controls. At the BM25 operating point with the released matcher, using the same rows and seeds:

For context, the untrained smallest-area heuristic gains +1.51 pp row-weighted on the same BM25 setting; its package interval is not included in the table below.

HeadRow ΔPackage Δ [95% CI]
Linear (node-normalized x/y centers)+1.45+1.30 [+0.10, +2.58]
Linear (child count, depth, area)+2.79+1.72 [+0.70, +2.92]
Linear (10 features)+3.06+1.45 [+0.43, +2.50]
MLP (10 features)+3.81+1.65 [+0.67, +2.67]
Paired: MLP − 3-feature linear+1.02−0.07 [−1.19, +0.86]
Paired: MLP − 10-feature linear+0.75+0.20 [−0.38, +0.74]
Paired: MLP − x/y linear+2.36+0.34 [−1.07, +1.69]

The x/y prior uses candidate-center coordinates normalized by the maximum node-center coordinates within each hierarchy, not measured screen dimensions. Its package interval is positive; the paired MLP-versus-x/y package interval includes zero, while the row-weighted contrast is positive. The three-feature prior uses child count, depth, and area. The practical contribution is a bounded, instruction-independent intervention that can be attached to a frozen scorer, together with controlled evidence on when structural information changes selection. The DeBERTa comparison is in revised Table 1; the head controls, including the new x/y prior, are in revised Appendix Tables A24–A26.

3. Exact-tie denominators. Reviewers 1 and 2 correctly identified a mismatch. The tie-only residual's +2.739 pp is normalized over all 4,381 rows; on the same 1,195 tied rows it is approximately +10.04 pp, versus +5.523 pp for smallest-area ordering. The corrected Appendix F compares like with like. The always-on BM25 result (+3.62 pp over all rows) is unchanged.

We also reorganized the revised manuscript to present the bounded correction and matched operating points more directly, with detailed controls and replay diagnostics in the appendix.

Area_Chair_n9ZEOpenReview ↗

Urgent Reminder: Discussion Period Ends November 18

Dear Reviewers,

The discussion period will end on November 18, 2026. Please carefully review the authors' responses and actively engage in discussions.

If you have not yet initiated a discussion, please do so immediately. Reviewers are expected to address the authors' rebuttals, provide constructive follow-up comments, and update their evaluations when appropriate.

Please also ensure that your reviews reflect independent human judgment and that any AI assistance is properly documented in accordance with the review guidelines.

Active participation is an essential reviewer responsibility and is critical to ensuring a rigorous and fair review process.

Thank you for your cooperation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.