acceptodds
Under review as a conference paper at ICLR 2027

EQUAL INFORMATION IS NOT ENOUGH: ENCODER RANKINGS DEPEND ON INTERFACE REALIZATION

Abstract

Encoder comparisons are often interpreted architecturally once the information supplied to each model has been matched. We show that information matching is necessary but not sufficient: the same information can be realized through heterogeneous model interfaces in ways that induce different learning problems and different encoder rankings. We separate declared information φ(X) from its architecture-specific realization ρA, writing performance as Y (A, ρA(φ(X)); T ) under a fixed learning protocol T . In camera–LiDAR correspondence, all controlled inputs encode exactly the same occupied-cell set; an audit over 13,935 validation frames finds zero mismatched cells. Holding architecture, parameter count, optimization, comparator, gallery and evaluation fixed, changing only the realization reverses SparseConv3D’s ranking against a raster encoder on RADIATE (−0.0729 → +0.1206) and Boreas (−0.0115 → +0.0618); all four comparisons are resolved under the preregistered criterion. The effect extends to PointPillars on both datasets: on RADIATE an unresolved +0.0073 becomes a resolved +0.0569, and on Boreas an already-resolved +0.0032 becomes +0.0451, with direct realization effects of +0.0496 and +0.0419. The SparseConv3D–PointPillars ordering reverses on both datasets as well. We therefore distinguish information equivalence from interface realization and introduce a realization-aware Declare–Realize–Audit protocol. Architectural attribution requires controlling not only what information is available, but how that information is exposed to the learner.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation Learning, input representation, Point-cloud Encoders, Preregistration, Architecture Comparison, Camera-LiDAR correspondence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HKpaOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I think it has to do with encoders in 3d representation, but I could not tell, the paper is very hard to parse.

Critical strengths

It has 9 pages.

Critical weaknesses

This paper is too hard to parse for me to make a review. The problem is never introduced clearly, the proposed solution is never proposed clearly. The paper is poorly organized.

Questions for authors

Why did you submit a paper this hard to read? Why submit a paper in this state, this is irrespecutful of the reviewers' time.

Here are my notes while "reading":

Reading the abstract I have not idea what the problem is. 4 sentences in we talk about lidar. This abstract is near impossible to parse.

Introduction: it is not improving, first paragraph and I still dont understand the problem. wow third paragraph is quite something.

5 comments

We are sorry and we take the point that the paper was hard to read, and we have rewritten the abstract and opening of the introduction so that the problem is stated in plain language before any notation appears. In short:

Two encoders are given exactly the same information: the same set of occupied cells on a map, verified cell by cell. The only thing that differs is the format in which that set is handed to each encoder, for example whether a cell’s coordinates are written into the encoder’s input features or left for the encoder to infer from where the cell sits. That formatting choice alone changes which encoder wins, by a wide margin, on two different datasets. Comparisons between encoders usually report what information each model received, not how it was formatted, so the effect is invisible in how results are normally written up. The paper measures this effect, identifies what causes it, and proposes a simple reporting rule that makes the formatting choice explicit.

The revised first page is written at this level, and the formal definitions now start in §3, which is also shorter. We would be grateful if the reviewer could look at the revised first page and tell us whether any specific passage remains unclear.

Maybe it is me, but that first sentence to me is confusing. I do not know much about this problem of comparing encoders. The first sentence feels like 2 ideas being put together (in the introduction). Overall the new version of the paper does not clarify my confusion. Do you have a related work I could look at for me to make sense of the work? So far I am not inclined to change my score since I believe research should be accessible even if you are not an expert.

Thank you for reading the revision and for pointing to exactly where it loses you. You are right: the first sentence tried to say two things at once, that we are comparing two encoders and that we are matching what they are given. Here is the problem with one concrete example.

Suppose we want to compare two neural-network encoders fairly. As an example, take two encoders that read LiDAR data. We give both encoders exactly the same information, namely which cells of a top-down grid contain a LiDAR point. If one encoder does better, it is natural to credit its architecture. We found that “the same information” can still be handed to an encoder in two ways. In one, a cell’s position is only implied by where it sits in the grid. In the other, the cell’s coordinates are also written out as numbers in the encoder’s input. Both ways carry the same information, yet which encoder appears better can depend on which way was used.

That is the problem our paper studies. Many encoder comparisons describe what information each model receives, but not always how that same information is encoded for each model. If that choice changes the result, a reader cannot tell whether the winner has the better architecture or simply received the information in a form that is easier for that architecture to use. Our paper measures this effect on two datasets, gives evidence that much of the effect comes from how easily each encoder can use position from its input, and proposes that this presentation choice be stated and controlled when encoders are compared.

For background, two papers from our Related Work may help. Goyal et al., “Revisiting Point Cloud Shape Classification with a Simple and Effective Baseline” (ICML 2021), show that conclusions about which 3D architecture is better can change when details of the comparison are controlled. Liu et al., “An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution” (NeurIPS 2018), show that a convolutional network can have position implicit in the spatial layout and still find it much easier to use when coordinates are supplied explicitly.

Our paper connects those two ideas: even after the information itself is matched, can the way it is presented to different encoders change the conclusion about the architecture? We have rewritten the opening of §1 in the revised PDF(uploaded now) so that it begins with this example, in short sentences, with each idea stated on its own.

Reviewer_RRqAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper distinguishes the information declared for an encoder comparison from the way that information is realized at each model interface. In camera-LiDAR correspondence, two realizations preserve exactly the same occupied-cell set but expose cell-derived position differently. This changes the relative ranking of SparseConv3D and a raster encoder on RADIATE and Boreas. The authors propose a Declare-Realize-Audit protocol for reporting such comparisons.

Critical strengths
  1. The study isolates a well-defined comparison variable. The two input realizations preserve the same occupied cells, while the architecture, parameter count, optimization, comparator, and evaluation protocol remain fixed. The frame-wise audit supports the claim of information equivalence.
  2. The ranking reversal is substantial and reproduced on two independently trained datasets. Five-seed comparisons and controls for gallery support, parameter budget, and comparator choice make the result credible within the stated camera-LiDAR correspondence setting.
  3. The Declare-Realize-Audit protocol is a useful methodological reminder for heterogeneous encoder comparisons. The paper clearly separates a conditional ranking claim from a representation-independent architectural claim and states the limits of its own experiment.
Critical weaknesses
  1. The empirical scope is narrow relative to the broad methodological message. Both datasets concern automotive camera-LiDAR data, and the scored task is token-level correspondence retrieval. The study does not establish whether similar ranking changes arise in other domains or downstream detection and segmentation.
  2. The result depends on the chosen class of admissible interface realizations. The two tested policies are informative counterexamples, but the protocol does not provide an independent criterion for selecting a preferred policy or characterizing sensitivity across other defensible realizations.
  3. The mechanism behind the family-dependent effect remains unresolved. Explicit positional accessibility is plausible, but the current experiments do not separate it from optimization geometry or other architectural interactions. This limits what can be learned beyond the demonstrated ranking reversal.
  4. The rankings are measured under a fixed compute budget with compact encoder representatives, and the RADIATE score uses a preselected subset of validation frames. Stability under longer optimization, stronger positional supervision, or larger models remains unclear.
Questions for authors
  1. What rule would select an admissible realization policy before observing results, and how should a comparison report sensitivity when several policies are equally defensible?
  2. Do the available training checkpoints or configuration sweeps indicate whether the ranking reversal persists with longer optimization or stronger positional supervision?
  3. Can the authors further isolate positional accessibility from other changes in learning dynamics when cell-derived coordinates are written into native feature fields?
  4. How stable is the RADIATE ranking when more validation frames are scored, and what evidence suggests that the finding transfers beyond correspondence retrieval?
5 comments

We thank the reviewer for the accurate summary and for separating the conditional ranking claim from the architectural one. We answer the four questions below, and the scope point at the end.

Q1 — a rule for selecting a realization before results are seen (W2). Revised §7 now states one. (1) Matched access: enumerate the native semantic fields the compared interfaces expose; any such field that is deterministically recoverable from the declared statistic and natively exposed to one architecture is reconstructed for every architecture with a corresponding native field, and where none exists a minimal declared adapter (here, two input channels) serves as the sensitivity control. Arbitrary derived functions of the statistic are not added. (2) Report both: Δ against the comparator and Δρ within each family, together. (3) Architectural only if invariant: a ranking is called architectural only if it holds, under the resolution criterion, for every admissible realization reported; otherwise it is reported as realization-conditional. Rule 1 was applied before any new result was read: the raster comparator now receives the same coordinates as the sparse families (Table 3, lower block). Under matched access SparseConv3D leads it on Boreas (+0.0204, S 11.3, resolved) and nominally on RADIATE (+0.0161, S 7.6, clustered interval includes zero). Rule 3 applied to our own headline: the SparseConv3D–comparator ranking is not realization-invariant (§3) and is reported as a conditional estimand rather than an architectural one (§8). For sensitivity across defensible realizations, the revision reports the ranking under each realization run, including a graded family of five (§6).

Q3 — isolating positional accessibility (W3). We measured it three ways, and the three agree on both datasets. A linear position probe on the output tokens: SparseConv3D decodes position worst under ρ_str (S 45 RADIATE, S 116 Boreas) and best under ρ_if; the padding families decode it at two-thirds of chance error, so their structural route is partial. The Stage-0 alignment loss: worst for SparseConv3D under ρ_str, indistinguishable from the comparator under ρ_if. A coordinate-quantization sweep on SparseConv3D: P@1 unchanged to 4-cell quantization, resolved drops at 8 and 16 cells (S 2.7, 8.0), Spearman −0.90. Two controls close the alternatives: removing Stage 0 leaves the deficit at 93% of its value (S 20.8), and every configuration change tried raises the arm by +0.0091 to +0.0394, the best recovering about half (50.1%) of the deficit, with the four that widen the route to position giving 2.6–4.3× the gain of the one that does not (Appendices K.1, K.2). We therefore describe positional accessibility as the measured mediator, and say in §8 that we have not decomposed it further into inductive-bias alignment and optimization geometry.

Q2 — longer optimization, stronger positional supervision (W4). Tripling the Stage-1 schedule on the RADIATE SparseConv3D pair leaves the structural deficit negative and resolved at every length (−0.0766, −0.0811, −0.0674 at 6, 12, 18 epochs; S 31, 29, 15); the 6→18 change is not resolved (S 1.8) while all three arms improve resolvedly, so the plateau is in the gap, not in learning (Appendix K.4). The epoch-6 point also reproduces the submitted arms on current code within their seed floors. On supervision: Stage 0 already aligns every tower contrastively to a positional encoding of the cell coordinates, identically for every arm, and the deficit persists after it (−0.0787) and without it (−0.0732, S 20.8). Whether a detection head would remove it is untested here, and §8 says so.

Q4 — more frames; transfer (W1, W4). Every RADIATE arm is now scored on all 15,002 validation frames with the token-pooled estimator used for Boreas; the submission scored the first 300 frames of the unshuffled validation order with a per-frame mean, and the revision says so. The full-population estimator was registered as primary before any number was read; every resolution verdict is unchanged (one sign changes, the PointPillars ρ_str tie, unresolved under both estimators), several RADIATE magnitudes shrink substantially (by a third to a half), and floors tighten 2–3×, so every contrast is better resolved (Appendix F). We also add a second leg to the resolution criterion, a cluster bootstrap over validation sequences (RADIATE) and 64-frame blocks (Boreas), because the seed floor measures training uncertainty only. It demotes two RADIATE contrasts that clear the seed floor, the SparseConv3D–PointPillars gap under ρ_if and SparseConv3D's lead over the coordinate-equipped comparator. Both are small leads whose sequence-clustered interval includes zero, and with only 12 clusters that leg is coarse and conservative by construction. The first was an auxiliary claim of the submission (that SparseConv3D overtakes PointPillars on RADIATE); we withdraw it and report the pair as tied there. The second was never a claim. The central SparseConv3D–raster reversal is resolved on both datasets under both legs. On Boreas, every headline verdict holds within each traversal separately. On transfer: we do not claim it. Restricting the gallery to a 5×5 window, so that global position no longer discriminates among candidates, attenuates the structural deficit by 76% with a resolved local remainder (Appendix K.5), so most of the effect measured here is global positional discrimination; whether it recurs in detection or segmentation is untested (§8).

W1 — scope. We agree, and §8 states it plainly: two automotive camera–LiDAR datasets, one task family, compact encoders. What the paper shows is that the failure of comparison invariance exists, is large, is family-dependent in a measured way, and replicates, in one controlled setting; the protocol and the selection rule are the part we expect to carry over.

We hope these answer the reviewer's questions, and we would be glad to clarify anything further during the discussion.

Dear Reviewer RRqA,

As the discussion period closes tomorrow, we wanted to check whether anything in our response to your four questions remains unclear. The revised paper and supplement include the three-clause realization-selection rule (§7), the 6/12/18-epoch schedule comparison (Appendix K.4), the local-window scoring that attenuates global positional discrimination (Appendix K.5), and the full 15,002-frame RADIATE evaluation with the added cluster-bootstrap leg (§4, Appendix F). A further revision uploaded today only rewrites the opening of §1 in plainer language, in response to another reviewer; no result or claim changed. We would be glad to clarify any point before the discussion closes.

Thank you again for a thorough and constructive review.

Reviewer_XR3eOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether giving two encoder architectures the same input information is enough to compare them fairly. The authors argue that it is not, because the same information can be presented to each model in different ways, and that choice can change which architecture wins. They test this on camera-LiDAR patch retrieval in bird's-eye view (BEV), using RADIATE as the main dataset and Boreas as a replication. Every model receives the same set of occupied grid cells, and the authors verify this across 13,935 validation frames.

The experiment compares two input encodings for the sparse encoders. In the first, each occupied cell carries a constant feature of 1.0, so the model can infer position only from where the cell sits in the sparse grid. In the second, the cell-centre (x, y) coordinates are also written into the input features. Adding the coordinates moves SparseConv3D from below a raster CNN baseline to above it on both RADIATE (-0.0729 to +0.1206 P@1) and Boreas (-0.0115 to +0.0618). It also makes the PointPillars gain over the baseline significant under the paper's test, and it reverses the order of SparseConv3D and PointPillars. All encoders are small (about 0.2M BEV parameters) and are trained with 5 seeds. From these results, the authors propose a protocol for reporting how inputs are presented to each model in encoder comparisons.

Critical strengths

Quality

The experiment is carefully controlled. Model size, training schedule, data order and the baseline are held fixed across the two input encodings, so the only difference between conditions is whether the sparse encoders receive coordinates. The authors verify frame by frame that every model sees the same occupied cells, fix their decision criteria before running the experiments, and repeat the full study on a second dataset recorded with a different vehicle, sensor suite and city.

Raster baseline

The raster CNN is a sound fixed baseline. It is not retrained between conditions, it receives an identical input in both, and it is evaluated on the same retrieval gallery. The paper reports both a dense and a sparsity-invariant masked variant, so the comparison does not rest on a single weak baseline, and a post-hoc check in Table 4 shows that the choice of raster baseline does not change any of the conclusions.

Significance

The reporting recommendation is correct and easy to adopt. A comparison between architectures should state how the shared information is presented to each model, not only which information is shared, and benchmarking papers would benefit from following this practice.

Critical weaknesses

Some key technical limitations and suggested improvements:

  1. The main result is largely explained by a known property of sparse convolution. With constant input features, submanifold SparseConv3D only sees the local pattern of occupied cells. Because the operation is translation-equivariant and, unlike dense convolution, has no zero-padding boundary that could reveal absolute location, the model cannot determine where a cell lies on the map, which is exactly what the 576-cell retrieval task requires. Adding the cell-centre coordinates as features is essentially CoordConv (Liu et al., 2018) applied to sparse convolution, and its benefit on position-dependent tasks is well established. The paper acknowledges in Sec. 6 that the two input encodings give the model different sets of learnable functions, and it cites usable information (Xu et al., 2020). I agree with this reading, but it implies that the ranking change is the expected consequence of removing one model's access to absolute position rather than a new phenomenon. I am also not convinced that the constant-feature encoding is what a careful practitioner would choose for this task. FCGF uses constant features because it needs location-independent local descriptors, which is the opposite of what retrieval against a fixed BEV grid requires. The paper would be more convincing if it found two input encodings that change the ranking without changing what the model is able to represent.

  2. The comparison with coordinates is not balanced, because only the sparse encoders receive them. The raster baseline gets the same input in both conditions and never sees the cell-centre coordinates. Sec. 7 itself states that this kind of imbalance should be justified or removed, yet a raster baseline with coordinate channels is left as future work (Sec. 8). Without that control, the result can equally be read as "giving coordinates to one model family but not the other changes the ranking." The evidence is also narrow: a single retrieval task, encoders with roughly 0.2M parameters, 300 scored RADIATE frames, and no detection or segmentation results. A CoordConv-style raster baseline looks inexpensive to add and would be the most informative missing experiment, and even one downstream task would strengthen the general claim.

Questions for authors
  1. If the raster CNN baseline also receives cell-centre coordinates as extra input channels, does it again outperform SparseConv3D? This is the cleanest way to separate "the input encoding changes the ranking" from "only one model received coordinates," and the answer would affect my score the most.

  2. Is there a pair of input encodings that changes the ranking without changing the set of functions the model can represent, for example by providing the same coordinates through different but equally accessible input fields? Such a case would show an effect beyond the known role of absolute position.

  3. Does the ranking still change on a task that does not depend on absolute position, or with larger encoders?

5 comments

We thank the reviewer for a careful and precise review. We address the two weaknesses and the questions below. The coordinate-equipped comparator the reviewer asked for has been trained and scored on both datasets.

Q1 / W2 — the coordinate-equipped raster comparator. We trained the raster comparator with two additional input channels carrying the cell-centre coordinates, bit-identical to the values the sparse families receive under ρ_if (audited cell by cell on all 15,002 RADIATE validation frames, zero mismatches; +1,152 parameters, still the smallest model in the study), five seeds and both datasets.

Does it outperform SparseConv3D? No, on either dataset. SparseConv3D ρ_if leads the coordinate-equipped comparator by +0.0204 on Boreas (S 11.3, block-clustered CI excludes zero) and by +0.0161 on RADIATE (S 7.6). On RADIATE the clustered leg has only 12 validation sequences to resample, so it is coarse and conservative, and its interval for this lead includes zero; we let the stricter leg govern and do not call the RADIATE lead resolved. What symmetric access removes is the size of SparseConv3D's lead, not the erasure of its structural deficit (−0.0787, S 27.6; −0.0115, S 10.4; both resolved under both legs).

Is the realization effect just "one family received coordinates"? No. The comparator's own gain from the coordinates (its coordinate-access gain, since the arm adds two channels and is not a same-architecture Δρ) is +0.0362 (RADIATE) and +0.0414 (Boreas), indistinguishable from PointPillars' Δρ on both datasets; SparseConv3D's Δρ is +0.1309 and +0.0732. The excess is resolved under both legs on both datasets (C2: +0.0947, S 26.7; +0.0319, S 15.1). The SparseConv3D–PointPillars interaction on Boreas (+0.0313, S 16.1) was pre-registered before that arm existed.

Standard CoordConv. On the sparsity-invariant comparator used throughout, coordinates on every cell and coordinates on occupied cells only are the same arm by construction (the masked convolution never reads empty cells); five seeds trained under both modes are bit-identical. On the dense comparator they differ, and this pair may be the clearest instance of the paper's point. The occupied-only input is the CoordConv input multiplied by the occupancy channel the model already receives; the declared information and the architecture are identical; and the CoordConv arm, the one holding the superset input, is resolvedly worse on both datasets (D-cc − D-if: −0.0320, S 7.6 RADIATE; −0.0094, S 3.9 Boreas) and does not reach the masked R-if arm on either (D-cc − R-if: −0.0380, S 8.8; −0.0091, S 3.8). Strictly, a linear first layer cannot form x·m from m and x, so the two arms are equivalent only up to that gating; within that caveat the gap cannot be information withheld from one model, and the direction has a plain reading: coordinates written on empty cells are frame-independent constants that a dense trunk must learn to ignore, and the penalty is larger on RADIATE, where 86% of cells are empty, than on Boreas, where 56% are (Appendix K).

W1 / Q2 — "largely explained by a known property." The reviewer is right that positional accessibility explains much of the effect in this task. What is new is that this known module-level dependence changes an architecture verdict under exact information equality, and does so by a family-dependent amount; the revision now measures it rather than naming it. (i) A linear probe from each encoder's output tokens to the cell-centre coordinates: under ρ_str SparseConv3D decodes position worst (20.4 cells RMSE vs 17.6 raster, 17.5 PointPillars; S 45 on RADIATE; 21.0 vs 16.4 / 16.4, S 116 on Boreas). Even the padding families decode position at only two-thirds of chance error, so their structural route is partial, which is why the comparator gains from explicit coordinates. Under ρ_if SparseConv3D becomes the best-decoded family on both datasets, matching its retrieval lead (Appendix K.1). (ii) Quantizing the reconstructed coordinates to k cells gives a graded family of admissible realizations: P@1 is unchanged to k = 4 (≈1.6 m) and falls resolvedly at k = 8 and 16 (S 2.7, 8.0), monotone by Spearman −0.90 with a clustered interval that never reaches zero. (iii) Removing Stage 0 entirely leaves the deficit at −0.0732 (S 20.8) against −0.0787 with it, so the protocol's positional pretext is not what handicaps the sparse arm (Appendix K.2).

The reviewer asked whether two realizations can change the outcome without changing the set of functions the model can represent. We cannot guarantee exact equality of the induced hypothesis classes under different realizations; our own formalism (§3, §6) makes that dependence explicit. The closest control we have is the dense D-if/D-cc pair above, which keeps the architecture, the parameter count and the declared information fixed and changes only where the same deterministic coordinate field is written, and still yields a resolved difference on both datasets. The graded ρ_k family then shows a smooth dependence on positional granularity. We therefore do not claim a phenomenon beyond positional accessibility, and the revised §6 says so. The contribution we do claim is the one the reviewer's own observation supports: a property well understood at the module level, and unreported at the comparison level, decides benchmark verdicts, by a family-dependent amount, with a measured mediator and a replication. We agree that a careful practitioner would choose ρ_if for this task. The comparisons we have in mind do not say which realization they used, and the structural default is what a standard sparse-tensor pipeline produces when the choice is not made deliberately, which is why §7 asks that the choice be declared and matched (rule 1); §5 now says this directly.

Q3 — a task where position is not the label; larger encoders. We have not added a detection or segmentation task and do not claim transfer; §8 states this. We did run the nearest thing the harness allows: restricting the gallery to the 5×5 window around the true cell, so that every candidate is a near neighbour and global position no longer discriminates. That attenuates the structural deficit by 76% (seed floor 0.06), and the remainder (−0.0186, S 3.8) is still resolved under both legs (Appendix K.5). Global positional discrimination accounts for most of the effect, with a resolved local residual. Larger encoders are untested and stated as such.

Scope. The 300-frame subset the reviewer noted is gone: every RADIATE arm is scored on all 15,002 validation frames, token-pooled, registered as primary before any full-population number was read; every resolution verdict is unchanged (one sign changes, on a contrast unresolved under both estimators), with several magnitudes down substantially and floors tighter (Appendix F).

We hope the symmetric control and the measured mediator answer the reviewer's questions, and we would be glad to clarify anything further during the discussion.

Dear Reviewer XR3e,

As the discussion period closes tomorrow, we wanted to check whether anything in our response remains unclear. The revised paper and supplement include the coordinate-equipped raster comparator (R-if) you identified as the most important missing control, the dense CoordConv comparison (D-cc versus D-if), the position-decodability and quantization measurements of the mechanism, and the full 15,002-frame RADIATE evaluation (§5–6, Table 3, Appendix K). A further revision uploaded today only rewrites the opening of §1 in plainer language, in response to another reviewer; no result or claim changed. If any point would benefit from clarification, we would be glad to provide it before the discussion closes.

Thank you again for your time and for a review that directly improved the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers (revised PDF is uploaded)

We thank the reviewers. Both technical reviews rate soundness as having only minor or fixable concerns. Reviewer XR3e identifies one control as decisive, the raster comparator having been realized once while the sparse families were realized twice; Reviewer RRqA asks four questions. We have run that control and answered each question with new data, on both datasets, with every criterion registered before the corresponding run was launched (hash chain, Appendix K.3). The revised PDF marks every changed sentence in blue; a numbered change list is in the supplement (Appendix M) and summarized below.

1. Symmetric comparator (XR3e W2/Q1; RRqA). The raster comparator now receives the same cell-centre coordinates as the sparse families (R-if), bit-identical on every occupied cell. It does not overtake SparseConv3D on either dataset: SparseConv3D leads it by +0.0204 on Boreas (S 11.3, resolved) and +0.0161 on RADIATE (S 7.6; the sequence-clustered interval includes zero, so this lead is not called resolved). The structural deficit (−0.0787 RADIATE, −0.0115 Boreas) is erased under ρ_if whether or not the comparator is coordinate-equipped. Standard CoordConv on the dense raster is worse than coordinates on occupied cells only, on both datasets, and does not reach the masked R-if arm (§5, Table 3 lower block, Appendix K).

2. Mechanism (XR3e W1/Q2; RRqA W3/Q3). We agree positional accessibility is the principal mechanism and now measure it three ways, on both datasets: a linear position probe (SparseConv3D decodes position worst under ρ_str, S 45 RADIATE / S 116 Boreas, and best under ρ_if), the Stage-0 alignment loss, and a graded coordinate-quantization sweep (P@1 unchanged to 4-cell quantization, resolved drops at 8 and 16 cells; Spearman −0.90, clustered CI [−1.0, −0.7]). Removing Stage 0 leaves the deficit at 93% of its value; the best configuration change recovers about half (50.1%) of it and remains resolvedly behind (§6; Appendices K.1, K.2).

3. Scoring population and stability (RRqA W4/Q2/Q4; XR3e W2). Every RADIATE arm is now scored on all 15,002 validation frames with the token-pooled estimator used for Boreas. The submission scored the first 300 frames of the unshuffled order with a per-frame mean, and stated that the audit covered a superset of the scored frames; for RADIATE it did not, and the audit now covers exactly the scored frames (Appendix F). Under the registered S > 2 rule every resolution verdict is unchanged (one sign changes, on a contrast unresolved under both estimators); several RADIATE magnitudes shrink substantially (by a third to a half) while the floors tighten 2–3×. The shrinkage has a known source: a per-frame mean weights sparse frames, which score higher, equally with dense ones, and the 300-frame prefix over-represents the earliest validation sequences; the token-pooled estimator removes both (Appendix F). We then add a second, cluster-bootstrap leg to the criterion (sequences on RADIATE, 64-frame blocks on Boreas); it demotes two RADIATE leads whose sequence-clustered interval includes zero. One of them was an auxiliary claim of the submission, that SparseConv3D overtakes PointPillars on RADIATE under ρ_if; we withdraw that claim and now report the pair as tied on RADIATE. The other, SparseConv3D's lead over the coordinate-equipped comparator on RADIATE, was never a claim. The central SparseConv3D–raster reversal is resolved on both datasets under both legs. Tripling the Stage-1 schedule leaves the structural deficit resolved at every length (§8, Appendix K.4).

4. Selection rule (RRqA W2/Q1). §7 now states one: matched access, report Δ and Δρ together, call a ranking architectural only if it holds under every admissible realization reported.

5. Presentation (HKpa; XR3e). The abstract and the opening of §1 are rewritten in plain language before any notation; §3 is shortened by a third.

The revised numbers reproduce from the registered artifacts on demand (every number in both PDFs is generated from one script; the three headline arms and both dense comparators re-trained on current code reproduce within their seed floors, Appendix L).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.