acceptodds
Under review as a conference paper at ICLR 2027

VecMALT: Fixed-Capacity Global–Local Representations for Geospatial Vector Geometry

Abstract

Encoding geospatial vector geometry for reuse requires a fixed-size code to retain both overall form and spatially localized detail. We introduce VecMALT, which stores a global embedding and a fixed number of local chart records, each pairing a feature with its spatial frame. Coverage-constrained measure allocation bounds token participation, jointly assigns evidence to charts, and supervises their support along sampled boundaries. Across five code widths, local capacity improves complex-polygon reconstruction, while six interventions reveal allocation dependent trade-offs. We also freeze the codes and train independent readers. Accurate reconstruction does not ensure easy attribute access: for VecMALT, chart-record reading reduces window-occupancy MAE by 31.8% relative to a flat-code MLP, and feature–frame reassignment raises error by 16.3% after reader retraining. On three public polygon cohorts, with geometries available during representation learning and task labels held out, record-aware readers reduce boundary-count MAE in all nine local-method/dataset comparisons. These results distinguish geometric fidelity, spatial support, and independent access as complementary tests of reusable geometry codes.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
geometric representation learning, geospatial vector geometry, global-local representations, implicit neural representations, optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_s2XFOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces VecMALT, a fixed-width global-plus-local representation for geospatial vector geometry, with local records pairing features and spatial frames. It combines constrained allocation, transport-based chart construction, refinement, and sampled boundary supervision. Experiments cover reconstruction across code widths, frozen-code attribute readers, feature–frame reassignment, public downstream tasks, and storage/runtime. The results support the broad value of local capacity for difficult polygon reconstruction and suggest that record-aware readers can improve access to frozen codes, but the specific benefits of VecMALT’s allocation and support mechanism are not isolated.

Critical strengths
  • The global-plus-local representation and explicit feature–frame records address a meaningful problem: retaining both global geometry and localized spatial information under a bounded latent width.
  • The evaluation distinguishes reconstruction, spatial support, and independent readout, and reports useful negative results and protocol qualifications rather than presenting a uniformly favorable picture.
Critical weaknesses

Critical Weaknesses:

  • Major Issue 1 — The central method comparison is not causally isolated. In Section 4 and Appendix A.7, Uniform-local, Demand-adaptive, and VecMALT differ in allocation, transport cost, refinement, regularization, schedules, and boundary supervision. The experiments therefore establish at most that particular complete systems behave differently; they do not establish that coverage-constrained allocation causes the gains. The broad local-versus-global conclusion is more credible than VecMALT’s specific superiority.
  • Major Issue 2 — The independent-reader comparison is substantially confounded. In Section 3.4 and Appendix H.2, Flat MLP, code-chunk attention, and chart-record attention differ in grouping, parameter sharing, normalization, architecture, and parameter count. Thus, the 31.8% window-MAE reduction and the reassignment result support the usefulness of one record-aware interface, but do not cleanly establish that learned spatial frames or VecMALT codes themselves provide the improvement. Area-priority checkpoint selection (Appendix A.4) is an additional confound because area is a headline readout target.

Minor Issues:

  • Most experiments use a single training seed, while the reported bootstrap intervals condition on fitted models and do not capture seed variability. This makes small differences and intervention rankings difficult to interpret.
  • The paper should distinguish fixed latent width from total serialized capacity, since pose and header information add substantial storage (Equation (1), Tables A1, and Appendix G.3).
  • The external topology evidence is limited (Appendix B.3: 17 polygons with holes and 3 multipart polygons), and the public-task protocol is transductive because test geometries participate in representation learning. Claims about general topology preservation and unseen-geometry reuse should therefore be narrowed.
  • Comparisons with Geo2Vec combine different acquisition, fitting, compute, and extraction protocols; the scope of this comparison should be made explicit.
Questions for authors
  • Clarification that could change the assessment: Can the central B=512 comparisons be repeated across multiple representation and reader-training seeds, with uncertainty that includes seed variability?
  • Can the allocation ablations hold localization, refinement, regularization, and boundary supervision fixed while changing only the allocation component?
  • Can the reader comparison match normalization, parameter count, architecture, and training budget, including controls with randomized spatial frames or feature–frame pairings?
  • Are pose and structural-header fields available to downstream readers, and do the results hold when total serialized storage—not only latent width—is matched?
  • Can external results be reported separately by source, scale, and topology, and can the public results be clearly labeled as transductive?
5 comments

Response to s2XF — Part 1 of 3: Allocation and representation-training variation

The new experiments address the two attribution questions separately: a one-factor marginal control tests allocation, and a same-architecture intervention tests explicit frame inputs. Additional representation seeds evaluate training variation.

Major Issue 1: A matched allocation-only test

The new control changes only λmax⁡\lambda_{\max} from 0.25 to zero, making the target token marginal uniform. Architecture, localization, transport cost and solver, refinement, regularization, boundary supervision, the remaining schedule settings, the 100,000-step budget, and checkpoint selection retain their Full settings. This differs from historical Uniform-local, which changes other settings, and from no learned importance, which retains the geometric prior in the marginal. The control protocol is given in revised Appendix M.4.

Full and control each use representation seed 42. Their codes are frozen before training readers with paired seeds 42–46. Evaluation uses the same 20,000 external polygons, with 64 windows per object. Values are mean ± sample SD across five reader runs.

Table S1. Window MAE in the allocation-only comparison.

Reader conditionVecMALT FullUniform-marginal control
Intact chart-record attention0.032829 ± 0.0016480.034014 ± 0.001932
Flat-code MLP0.044994 ± 0.0019420.045715 ± 0.001991
Code-chunk attention0.034911 ± 0.0012750.033490 ± 0.001463
Frame-masked chart-record attention0.032953 ± 0.0014520.034026 ± 0.001322
Reassigned chart-record attention0.033386 ± 0.0013660.034297 ± 0.001295

With intact records, Full lowers mean MAE by 3.48%; the paired Full-minus-control difference is −0.001185 ± 0.000781. RMSE also decreases, from 0.060352 to 0.059593. Full has lower mean MAE in four conditions; code-chunk attention favors the control.

The control's intact MAE is higher than historical Uniform-local's (0.034014 versus 0.032360); it retains Full's other settings to isolate marginal adaptation rather than reproducing that configuration.

Additional representation seeds

We completed seeds 43–46 at B=512B=512 for all five methods: 20 additional representation runs, with reconstruction and area/window evaluation. Table S2 reports polygon metrics across those four runs. Canonical distances are scaled by 10310^3; the last column is in meters.

Table S2. Polygon reconstruction, mean ± sample SD.

ConfigurationIoU (%) ↑Canonical CD ×10³ ↓Canonical HD95 ×10³ ↓HD95 (m) ↓
Adapted Geo2Vec98.4188 ± 0.15892.0217 ± 0.31035.4945 ± 0.96340.4030 ± 0.0321
Global-only98.7880 ± 0.01171.4768 ± 0.01994.2447 ± 0.06120.7969 ± 0.0311
Uniform-local99.2287 ± 0.26201.0866 ± 0.31072.6771 ± 0.77600.5477 ± 0.1160
Demand-adaptive99.3362 ± 0.03270.9497 ± 0.03632.5187 ± 0.15710.3982 ± 0.0661
VecMALT99.3733 ± 0.00870.9180 ± 0.00692.2196 ± 0.00470.4737 ± 0.1033

All runs are included. VecMALT exceeds Uniform-local's polygon IoU in all four paired runs, has the lowest mean canonical polygon distances, and averages 0.5853 percentage points above Global-only in IoU. Uniform-local's seed-44 IoU is 98.835692%, versus approximately 99.35–99.36% in its other runs, accounting for its larger spread. Demand-adaptive retains the lowest mean meter-valued polygon HD95.

The corresponding reconstruction results are in revised Table 4A and Appendix M.2. Part 2 separates reader-seed variation on fixed codes from variation across new representation/reader pairs.

Response to s2XF — Part 2 of 3: Reader controls and checkpoint selection

Major Issue 2: Same architecture, normalization, and training for the frame intervention

The new frame-masked control uses the original 34,689-parameter chart-record reader. For 512=128+24(2+3+11)512=128+24(2+3+11), we reuse the intact condition's task-training normalization, then zero the two anchor and three metric channels of each standardized record. The eleven feature channels, global embedding, query input, input width, and architecture remain unchanged.

Masking applies to training, validation, and test inputs before reader retraining. Intact and masked runs use paired reader seeds, the same maximum of 50 epochs, and validation-MAE checkpoint selection. Reassignment preserves both feature and frame sets while changing their correspondence. The encoder and reconstruction decoder remain frozen; each condition starts from the same cached codes and applies its specified input transformation.

Two experimental groups are reported separately. Fixed-code runs keep representation seed 42 and vary reader seeds 42–46. New representation/reader pairs use matching seeds 43–46 at both stages. Evaluation comprises 20,000 external polygons and 64 windows per object. Tables S3–S4 focus on VecMALT and Uniform-local; ± denotes sample SD across the runs in each row.

Table S3. Window MAE under matched frame interventions.

SettingConfigurationIntactFrame-maskedReassigned
Fixed codes; five readersUniform-local0.032360 ± 0.0011390.032908 ± 0.0018680.032693 ± 0.001574
Fixed codes; five readersVecMALT0.032829 ± 0.0016480.032953 ± 0.0014520.033386 ± 0.001366
Four new trained pairsUniform-local0.031691 ± 0.0019350.032398 ± 0.0017550.033175 ± 0.002021
Four new trained pairsVecMALT0.031191 ± 0.0007910.032497 ± 0.0021210.033916 ± 0.000915

For VecMALT, masking raises MAE in 4/5 fixed-code runs but gives a small mean increase (+0.38%). In the new pairs, the paired masking increase is +0.001306 ± 0.002663 (3/4 positive); reassignment gives +0.002725 ± 0.000677 (4/4 positive). Original correspondence has repeated support, while the explicit-coordinate effect varies across runs. Revised Table 4B and Tables M8–M9 report the intervention and interface comparisons.

The interface benefit is distinct from the frame-only effect

Flat-code MLP, code-chunk attention, and chart-record attention have 74,497, 28,353, and 34,689 registered parameters. Their comparisons test complete interfaces; only the intact/masked/reassigned controls hold the chart-record architecture fixed.

Table S4. VecMALT paired interface comparisons; negative differences favor intact reading.

SettingIntact − Flat MAELower inIntact − Chunk MAELower in
Fixed codes; five reader seeds−0.012164 ± 0.0023235/5−0.002082 ± 0.0026554/5
Four new representation/reader pairs−0.013395 ± 0.0006304/4−0.003524 ± 0.0018124/4

Intact reading improves over flat in 9/9 comparisons and chunk in 8/9, across these two designs. Reductions computed from mean MAEs are 27.0% and 30.0% versus flat. These comparisons measure record organization together with reader architecture, sharing, and normalization. The matched frame controls above examine the explicit channels; retained features may still encode spatial information.

Area-priority does not select on the downstream area regressor

Area-priority selects representation checkpoints using internal polygon-family, complexity-tier-weighted field errors and polygon field/geometry terms (Appendix A.4). The independent area MLP, window-reader errors, and downstream test labels are not selection inputs. Reader fitting follows checkpoint selection, and all conditions reading the same code share that checkpoint.

Response to s2XF — Part 3 of 3: Storage, external strata, and evaluation scope

Total storage and reader inputs

At B=512B=512 in FP32, a local code stores 512 bytes of global features, 480 bytes of frames, and 1,056 bytes of local features: 2,048 bytes in total. All frame coordinates count toward the code budget. Independent readers receive the code and, for windows, the query rectangle; pose and structural headers are not reader inputs.

Table S5. Measured storage, including metadata.

ConfigurationLearned code (bytes/object)Mean payload (bytes/object)With amortized decoder (KiB/object)
Global-only2,0482,896.33.122
Uniform-local2,0482,896.33.067
Demand-adaptive2,0482,896.33.067
VecMALT2,0482,896.33.067

Payload includes 16 identifier/index bytes and mean pose/header storage of 832.3 bytes. Decoder amortization uses 54,379 objects, as in submitted Table G7. Thus the measured mean object payload is matched across the four encoder configurations; shared networks and compute are separate resources.

Source-specific external results

The following use the original seed-42 reconstructions, not the new-seed averages.

Table S6. External reconstruction by source.

ConfigurationDelhi buildings: IoU (%)São Paulo land use: IoU (%)Cairo curves: CD (m)
Adapted Geo2Vec98.893896.86270.4081
Global-only99.213796.85320.3606
Uniform-local99.625998.24940.4071
Demand-adaptive99.566798.20070.4980
VecMALT99.626698.24000.4021

The columns contain 32,373 polygons, 7,100 polygons, and 14,903 curves. Three additional São Paulo curves remain in the all-object evaluation. Local configurations improve polygon IoU in both polygon sources; Global-only leads Cairo curve CD. Geography and feature type are coupled, so these are source-specific effects.

Scale and topology

Table S7. Polygon HD95 in meters, from Table J2. The common-scale subset uses reference pose scale ≤362.0883 m (the population's 99th percentile). P95 is computed across finite per-object HD95 values; the final column gives the subset's finite-distance count.

ConfigurationFull meanFull P95Common-scale meanSubset finite n
Adapted Geo2Vec0.40141.14010.206139,078
Global-only0.77141.58390.228039,078
Uniform-local0.65170.65560.106339,077
Demand-adaptive0.36820.67690.118739,077
VecMALT0.59800.64210.104439,078

VecMALT has the lowest full-population P95 and common-scale mean. One retained large polygon contributes 53.73% of its summed meter-valued HD95, explaining the sensitivity of the full mean to scale. The diagnostic subset was already reported in Table J2.

Table S8. Polygon IoU (%) on the available topology subsets.

ConfigurationWith holes (n=17)Multipart (n=3)
Adapted Geo2Vec81.948992.2330
Global-only66.671474.5068
Uniform-local81.709493.6014
Demand-adaptive77.541293.4860
VecMALT82.669893.8915

These estimates describe 17 hole-containing polygons and three multipart polygons; the sample counts limit conclusions about topology preservation.

Transductive public tasks and Geo2Vec's protocol

External reconstruction and area/window evaluation use held-out geometries. Public Shape/Edge learns representations from all cohort geometries without task labels, then holds out labels for reader evaluation. This transductive protocol tests label readout on known geometries (Section 4); Edge counts stored segments, including collinear vertices. The exploratory Shape-head revision is separately identified in Appendix K.4.

On external objects, adapted Geo2Vec fits new codes for 1,000 steps with its decoder frozen, whereas encoder methods use forward passes over prepared evidence. On public cohorts it exports jointly learned codes. These are comparisons under the declared adaptation and extraction pipelines, not equal acquisition compute. Acquisition and cached-code reading are reported separately.

Reviewer_s6gvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VecMALT, a fixed-capacity global–local representation framework for geospatial vector geometry. It combines global embeddings with spatially anchored local features and introduces coverage-constrained evidence allocation through bounded token participation, optimal transport, and boundary supervision. Experiments evaluate geometric reconstruction, spatial coverage, and independent downstream readout.

Critical strengths

S1: The proposed global–local representation explicitly associates local features with spatial anchors and geometric frames.

S2: The independent-reader and feature–frame reassignment experiments provide valuable insights into representation usability beyond reconstruction accuracy.

Critical weaknesses

W1: Table 1 shows that VecMALT does not achieve the best performance on any metric, with substantial gaps on some metrics.

W2: The boundary-coverage loss is applied only to cached reference-boundary samples. It provides a soft constraint rather than a guarantee of continuous boundary coverage. This limits the strength of the claimed coverage guarantees.

W3: Eq.6 uses the same support gate for both global detail and local residual branches. Queries outside all chart supports receive only coarse predictions, making global detail reconstruction vulnerable to local allocation errors.

W4: The local encoder receives both the object-level summary and learnable chart identities, but the training objective does not explicitly enforce complementarity.

W5: No code is released for reproductivity.

Questions for authors

NA

1 comment

Response to s6gv: Reconstruction, coverage, and specialization

W1 — Table 1 and the distribution of boundary error

Table 1's aggregate rankings are complemented by the boundary-error distribution in Table J2: VecMALT has the lowest full-set P95 of per-object HD95 and the lowest mean on the common-scale subset. The full-set mean and these distributional measures are shown together below.

Table R1. Original seed-42 polygon HD95 in meters (Table J2). P95 is across finite per-object HD95 values. The common-scale subset uses reference pose scale ≤362.0883 m, the population's 99th percentile; its finite-distance counts are shown.

ConfigurationFull meanFull P95Common-scale meanSubset finite n
Adapted Geo2Vec0.40141.14010.206139,078
Global-only0.77141.58390.228039,078
Uniform-local0.65170.65560.106339,077
Demand-adaptive0.36820.67690.118739,077
VecMALT0.59800.64210.104439,078

One large polygon contributes 53.73% of VecMALT's summed meter-valued HD95. It remains in the full-set statistics. The common-scale subset is defined by reference geometry, not prediction error, and was already reported in the submission.

The four additional representation seeds 43–46 at B=512B=512 give polygon IoU of 99.3733 ± 0.0087% for VecMALT, 99.2287 ± 0.2620% for Uniform-local, and 98.7880 ± 0.0117% for Global-only (mean ± sample SD). VecMALT exceeds Uniform-local in each of these four paired IoUs. These new runs and the original scale analysis answer different aspects of reconstruction quality.

W2 — Target-mass bounds versus sampled support

The bound μin≥(1−λmax⁡)uin\mu_{in}\geq(1-\lambda_{\max})u_{in} applies to the target token marginal (Eq. 2; Appendix A.6), with λmax⁡=0.25\lambda_{\max}=0.25. Spatial support is separately trained through soft losses at cached boundary samples (Eq. 5; Appendix A.5). The former bounds evidence participation; the latter supervises sampled spatial reach. Neither guarantees continuous boundary coverage, and zero uncovered samples is an empirical diagnostic.

W3 — What the shared gate controls

Eq. 6 retains a coarse global field everywhere and gates both global-detail and local-residual corrections. Outside all chart supports, only the coarse field contributes. Boundary supervision targets this support dependency. An ungated global-detail alternative has not been evaluated. Independent readers do not apply the decoder gate, so their results test access through a different reading pathway.

W4 — Context and specialization

Chart-relative coordinates and chart identities distinguish local encoding contexts, while the object summary supplies shared context. Appendix A.4 specifies supervision of the coarse field and detail residual. These provide spatial conditioning and output-level targets for specialization; they do not enforce statistical independence or strict global–local disentanglement. Feature–frame controls test whether the learned organization is useful, rather than whether all stored information is nonredundant.

W5 — Reproducibility

We plan to release the code and trained models officially at a later date. Appendices A and H document the model and reader implementation, and Appendix M now records the added experimental groups and controls.

Eqs. (2)–(4) and Appendix A specify the code layout, encoder/decoder parameterization, allocation and transport, Sinkhorn iterations, temperature, bounds, optimizer, schedule, and checkpoint selection. Appendix A.5 gives the sampled boundary-coverage objective; Table A2 identifies the method configurations. Appendices B and H–K describe data splits and sources, metrics, the original reader architectures and training settings, and the Geo2Vec adaptation.

Tables C2–C3 report the original 20-model capacity sweep and six interventions; H2 lists all 36 original external area/window configurations. Table 3 and K5–K6 report public-task results and the Shape-reader variants. Appendix M adds the representation-training repeats, uniform-marginal and frame-masking protocols, per-run IoUs, and paired reader statistics; Tables M10–M11 report the allocation control.

Comments on the paper

3 comments

Response to NecL — Part 1 of 2: Capacity, fitting variation, and allocation

We address Q1 through training repeats, the original component interventions, and a matched marginal control. All ± values denote sample SD across the runs identified below.

Budget dependence and additional representation seeds

The original sweep gives the following VecMALT-minus-Uniform-local polygon-IoU differences:

Code width BCharts KIoU difference (pp)
1286+0.0644
25612+0.0214
51224−0.0011
76836+0.0285
102448−0.0285

The small-budget gains are consistent with allocation having more leverage when records are scarce. Because the pattern is non-monotonic and global width grows with chart count, this is a budget-dependent observation. At 512 dimensions, Uniform-local and VecMALT have zero sampled gaps and essentially tied IoU.

We completed four additional representation seeds, 43–46, at B=512 for all five methods, including reconstruction and area/window evaluation. For the configurations central to Q1:

Table N1. Polygon IoU across new representation seeds.

ConfigurationMean ± sample SD (%)
Global-only98.7880 ± 0.0117
Uniform-local99.2287 ± 0.2620
VecMALT99.3733 ± 0.0087

VecMALT exceeds Uniform-local's IoU in all four pairs; every run is included (revised Table 4A; Table M4). This measures representation-training variation beyond the original conditional bootstrap.

Boundary supervision and the floor

Removing boundary supervision changes polygon IoU from 99.3772% to 99.3695% (−0.0077 pp; Table 2), a small single-run reconstruction change. The loss directly supervises sampled chart support; its separate support effect was not measured in the disabled-loss run (Table D1). Removing the explicit floor improves IoU by 0.0109 pp and lowers meter-valued HD95.

Paired window uncertainty and the competitive uniform configuration

For the original seed-42 models/readers, VecMALT-minus-Uniform-local MAE is −0.000469, with a 95% whole-object bootstrap interval of [−0.000548, −0.000387]. The 2,000 resamples retain all 64 windows per selected object among 20,000 objects. This interval conditions on the fitted models, not training-seed or spatial-block variation; the original object win rate is 51.06%.

Table N2. Intact-reader window MAE in the two repeat designs.

SettingUniform-localVecMALT
Original representations; reader seeds 42–460.032360 ± 0.0011390.032829 ± 0.001648
New representation/reader pairs 43–460.031691 ± 0.0019350.031191 ± 0.000791

In the new pairs, the paired difference is −0.000500 ± 0.001155, with each method lower in two runs. Uniform-local remains competitive for window prediction.

Completed allocation-only test

We change only λmax⁡\lambda_{\max} from 0.25 to zero, retaining Full's localization, transport, refinement, regularization, boundary supervision, remaining schedules, 100,000-step budget, and selection rule. This uniform-marginal control is distinct from historical Uniform-local and no learned importance (revised Appendix M.4). Full and control each use representation seed 42, with five paired reader seeds 42–46 on their frozen codes.

Table N3. Window MAE for the matched allocation comparison.

Reader conditionVecMALT FullUniform-marginal control
Intact chart-record attention0.032829 ± 0.0016480.034014 ± 0.001932
Flat-code MLP0.044994 ± 0.0019420.045715 ± 0.001991
Code-chunk attention0.034911 ± 0.0012750.033490 ± 0.001463
Frame-masked chart-record attention0.032953 ± 0.0014520.034026 ± 0.001322
Reassigned chart-record attention0.033386 ± 0.0013660.034297 ± 0.001295

Full lowers intact mean MAE by 3.48%, with paired difference −0.001185 ± 0.000781. Its lower mean persists under masking and reassignment; chunk attention favors the control. The control is also worse than historical Uniform-local (0.034014 versus 0.032360): it preserves Full's other settings to isolate marginal adaptation, not to reproduce Uniform-local.

Response to NecL — Part 2 of 2: Same-reader frame controls and reuse

Q2/W2: The requested control on existing frozen codes

Frame masking retains the chart-record reader's 34,689 parameters, 512=128+24(2+3+11)512=128+24(2+3+11) input layout, global pathway, query input, training budget of at most 50 epochs, and checkpoint-selection rule. We reuse the intact condition's task-training normalization, then zero the five frame channels in every standardized record, retaining the eleven feature channels. Masking applies to every split, and the reader is retrained with paired seeds. Reassignment retains both feature and frame sets while changing their correspondence.

The primary experiment keeps the original representation seed 42 fixed and trains reader seeds 42–46. A separate experiment uses four new representation/reader pairs at seeds 43–46. The following tables focus on VecMALT and Uniform-local. Window MAE is evaluated on 20,000 external polygons, each with 64 windows; values are mean ± sample SD.

Table N4. Same-architecture reader conditions.

SettingConfigurationIntactFrame-maskedReassigned
Original codes; five reader seedsUniform-local0.032360 ± 0.0011390.032908 ± 0.0018680.032693 ± 0.001574
Original codes; five reader seedsVecMALT0.032829 ± 0.0016480.032953 ± 0.0014520.033386 ± 0.001366
Four new trained pairsUniform-local0.031691 ± 0.0019350.032398 ± 0.0017550.033175 ± 0.002021
Four new trained pairsVecMALT0.031191 ± 0.0007910.032497 ± 0.0021210.033916 ± 0.000915

Table N5. Paired increases over intact reading; counts show runs with positive differences.

SettingConfigurationMasked − IntactPositiveReassigned − IntactPositive
Original codesUniform-local+0.000548 ± 0.0023893/5+0.000333 ± 0.0020004/5
Original codesVecMALT+0.000124 ± 0.0024784/5+0.000557 ± 0.0027094/5
New pairsUniform-local+0.000707 ± 0.0034842/4+0.001484 ± 0.0018114/4
New pairsVecMALT+0.001306 ± 0.0026633/4+0.002725 ± 0.0006774/4

For the original VecMALT codes, masking worsens four of five runs, but one larger reverse effect leaves a small mean penalty of 0.38%. In the new pairs, masking raises mean MAE by 4.19%, while reassignment raises it by 8.74% and is worse in all four runs. The explicit-coordinate effect is variable; the original feature–frame correspondence has repeated support in the new representations. This correspondence benefit is also observed for Uniform-local.

The complete interface is the more consistent result

Table N6. VecMALT paired interface comparisons; negative differences favor intact reading.

SettingIntact − Flat MAELower inIntact − Chunk MAELower in
Original representation; five reader seeds−0.012164 ± 0.0023235/5−0.002082 ± 0.0026554/5
Four new representation/reader pairs−0.013395 ± 0.0006304/4−0.003524 ± 0.0018124/4

Intact reading beats flat in 9/9 comparisons and chunk in 8/9 across the two settings, with reductions of 27.0% and 30.0% from the mean flat-reader MAEs. These comparisons evaluate record organization together with reader architecture, sharing, and normalization; the features retained under masking may still encode spatial information.

Reuse depends on the query and code view

Table G3 provides a direct comparison on 256 common external retrieval queries, without a trained adapter:

ConfigurationFull-code widthFull-code standardized cosine: nDCG@10Global widthGlobal raw cosine: nDCG@10
Global-only5120.50275120.5297
VecMALT5120.34451280.5274

VecMALT's 128-dimensional gg has a raw-cosine nDCG@10 close to Global-only's 512-dimensional gg. In this evaluation, global vectors serve object-level similarity, while the complete-record reader serves window-occupancy prediction. The full-code and global protocols also differ in standardization, so their difference is not a component-only ablation. Table G1 reports weaker complete-code linear probes on several targets. Public Edge tests stored segment counts, including collinear vertices, under its transductive protocol.

Overall response

We thank all reviewers for their constructive feedback. VecMALT addresses two connected questions: how to distribute a fixed geometry-code budget, and how a model other than the original decoder can access the resulting representation. The discussion now includes training repeats, a matched marginal intervention, and same-reader frame controls.

1. Repeated reconstruction and the existing scale analysis

Four additional representation seeds (43–46) were completed for all five B=512B=512 methods, including reconstruction and area/window evaluation. Polygon IoU is:

ConfigurationMean ± sample SD (%)
Global-only98.7880 ± 0.0117
Uniform-local99.2287 ± 0.2620
VecMALT99.3733 ± 0.0087

VecMALT exceeds Uniform-local's IoU in all four pairs. In the original Table J2, VecMALT also has the lowest P95 of finite per-object HD95 values (0.6421 m) and the lowest common-scale mean (0.1044 m). A large-object tail raises its full-set mean; the replies reproduce all five methods and retain that comparison.

2. Adaptive versus uniform participation within Full

The new control sets only λmax⁡:0.25→0\lambda_{\max}:0.25\rightarrow0, retaining the remaining architecture, objectives, training settings, and selection rule. Full and control use representation seed 42 and five paired reader seeds:

Intact chart-record readingWindow MAE, mean ± sample SD
VecMALT Full0.032829 ± 0.001648
Uniform-marginal control0.034014 ± 0.001932
Paired Full − control−0.001185 ± 0.000781

Full lowers mean MAE by 3.48%. Historical Uniform-local has lower MAE than this control (0.032360); the control isolates marginal adaptation within Full rather than reproducing that separate configuration. The original no-boundary-loss ablation changes IoU by −0.0077 pp in one run. It measures reconstruction sensitivity; its separate support diagnostic was not recorded.

3. Structured reading, explicit frames, and correspondence

The frame intervention retains the same 34,689-parameter reader, normalization, input width, queries, and training protocol. The five standardized frame channels are zeroed in every split, and readers are retrained on the training split. Reassignment instead changes feature–frame correspondence.

VecMALT settingMasking: change in mean MAEReassignment: change in mean MAEIntact beats flat / chunk
Original representation; five reader seeds+0.38%; worse in 4/5+1.70%; worse in 4/55/5 / 4/5
Four new representation/reader pairs+4.19%; worse in 3/4+8.74%; worse in 4/44/4 / 4/4

The complete-interface improvement is consistent across these comparisons; the explicit-coordinate effect is more variable. In Table G3, VecMALT's 128-D global component reaches raw-cosine nDCG@10 of 0.5274, close to Global-only's 512-D score of 0.5297. Global vectors support that object-similarity evaluation, while complete-record readers support window queries. Full-code standardized-cosine retrieval and several linear probes remain weaker.

Changes in the revised manuscript: Section 5.1 and Table 4A add representation-training repeats; Section 5.2 and Table 4B integrate repeated readers and frame controls; Section 5.3 and Table 4C report the uniform-marginal comparison. Appendix M separates experiment groups and paired statistics, A.4 clarifies checkpoint selection, and H.2 specifies frame masking. Original seed-42 results remain identified as such. Tables G3 and J2 supply the retrieval and scale analyses discussed above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.