acceptodds
Under review as a conference paper at ICLR 2027

Constraining Shortcut Pathways: A Front-Door-Inspired Interaction Bottleneck for End-to-End Autonomous Driving

Abstract

End-to-end autonomous driving exchanges information among traffic participants, map elements, and ego planning states through high-dimensional latent representations. Although expressive, these representations can also encode contextual cues that are predictive during training but weakly tied to physical interaction. When propagated across entities, such cues can form shortcut pathways that make planning sensitive to distribution-specific context. To constrain this information flow, we introduce the Causal Interaction Bottleneck (CIB), a front-door-inspired module for cross-entity interaction. For each source, CIB predicts type-specific kinematic or geometric hypotheses and associated mode probabilities. Each hypothesis is re-encoded before CIB performs target-conditioned attention, with the mode probabilities serving as priors. Raw source features do not bypass this mediated path. At the network level, this design enforces source-side structural mediation: newly transmitted source information can affect the target only through the mediator. Under deterministic mediation, the mediator cannot increase contextual information after conditioning on the physical anchor and entity type. We further establish stability of the resulting attention rule and derive explicit bounds for the discrete kinematic update. Across sequential, parallel, and iterative architectures, CIB consistently improves planning performance. Matched controls indicate that these gains are not fully explained by generic bottleneck regularization. Representation probes and paired interventions further show that CIB reduces sensitivity to appearance-only changes while preserving responses to behaviorally relevant physical changes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
End-to-End Autonomous Driving, Shortcut Learning, Physics-Guided Learning, Robust Representation Learning, Interaction Bottleneck, Motion Planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_i1omOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CIB, which constrains cross-entity communication through predicted physical states to reduce contextual shortcut reliance in end-to-end driving. Experiments report planning improvements across multiple interaction structures and driving benchmarks, with controlled analyses examining sensitivity to appearance and physical changes.

Critical strengths
  1. The paper introduces an explicit physical-state interface for cross-entity interaction. Controlled comparisons show consistent planning improvements across different interaction structures and evaluation settings.
Critical weaknesses
  1. The evidence that CIB addresses harmful shortcut reliance remains indirect. Appearance decodability and trajectory sensitivity do not establish that contextual cues cause planning errors.
  2. Evidence for CIB’s additional value on stronger recent models remains limited. It is unclear whether the benefits of physical constraints or additional supervision observed on weaker baselines persist in models with stronger prediction and interaction representations. Although modest improvements are reported on SparseDriveV2, repeated runs or confidence intervals are not provided for these gains, making it difficult to distinguish them from training and evaluation variability. Controlled comparisons and shortcut analyses on strong baselines are also limited, leaving uncertainty about whether the observed improvements stem from the proposed communication constraint.
Questions for authors
  1. What evidence connects the baseline’s appearance sensitivity to actual planning errors, and shows that CIB reduces these errors by limiting harmful shortcut reliance?
  2. Could you evaluate CIB across multiple strong recent planners under matched conditions, examining both planning performance and responses to appearance and physical interventions? Please provide repeated-run results or confidence intervals to assess whether the improvements are reliable across these models.
6 comments
AuthorsOpenReview ↗

Post 1 of 5: Q1. Connecting appearance sensitivity to planning errors

Thank you for distinguishing trajectory sensitivity from harmful planning behavior. We added collision-error measurements under matched appearance conditions, with progress reported alongside them. This directly evaluates whether a changed visual condition accompanies additional errors, beyond changing the predicted trajectory.

The table shows reference-to-changed collision rates in percent and the within-model error increment in percentage points. The reported intervals are 95% confidence intervals. Each comparison uses the same host and evaluation conditions.

HostHost collision rateCIB collision rateHost error increment, CICIB error increment, CI
VADv214.8 to 19.113.4 to 16.24.3 [1.0, 8.0]2.8 [0.0, 5.5]
DriveTransformer-L23.7 to 27.421.9 to 24.23.7 [0.2, 7.7]2.3 [-0.6, 5.5]
HiP-AD12.6 to 15.411.8 to 13.62.8 [0.1, 5.4]1.8 [-0.2, 4.2]
AlignDrive11.4 to 13.711.2 to 12.92.3 [-0.2, 4.8]1.7 [-0.4, 3.9]
SparseDriveV210.8 to 13.59.9 to 11.82.7 [0.0, 5.0]1.9 [-0.2, 4.0]

For VADv2, DriveTransformer-L, and HiP-AD, the host's appearance-induced error increment has an interval above zero. This supplies a direct behavioral connection for the tested conditions. Full CIB has smaller point estimates of both changed-condition collision error and the appearance-induced increment on all five hosts.

We checked progress because a planner can reduce collision exposure by making less progress. Under changed appearance, progress increases from 89.7 to 91.9 for VADv2, 85.8 to 88.8 for DriveTransformer-L, 94.0 to 95.2 for HiP-AD, 95.4 to 96.0 for AlignDrive, and 95.4 to 96.3 for SparseDriveV2. These paired patterns make uniformly more conservative behavior an incomplete explanation of the collision-error results.

VeteranAD+ supplies a separate native NAVSIM-v1 diagnostic. RGB corruption reduces the host's PDMS from 90.9 to 88.7, a loss of 2.2 with interval [0.6, 4.2]. CIB changes from 91.1 to 89.5, a loss of 1.6 with interval [0.1, 3.3]. We report this as a native corruption test, separately from the CARLA collision analysis.

The revised empirical claim is specific: the tested appearance changes affect planning errors, and physically mediated models show favorable error and progress point estimates under those changes. To isolate the role of communication content, we additionally compare raw-feature and physical branches directly, as described in our next reply.

Manuscript locations: Table 31 on p. 46, Table 33 on p. 48, Table 35 on p. 50, Table 37 on p. 52, Table 28 on p. 43, Table 39 on p. 53; Section 4.4 on p. 8.

AuthorsOpenReview ↗

Post 2 of 5: Q1. Testing communication content with raw-feature controls

We added a raw branch to the complete CIB model while retaining its physical predictions, supervision, and interaction modules. A matched physical branch adds capacity while keeping source communication in the physical representation. The comparison tests the consequence of restoring raw-feature access, beyond the physical inductive bias and auxiliary loss shared by these variants.

The following point-estimate diagnostics report clean DS/SR on Bench2Drive; training-seed variability is not reported for these ablations. Auxiliary denotes adding physical supervision to the original host. They are distinct from the repeated main comparisons in our third reply.

HostPhysical auxiliaryFull CIBCIB + rawCIB + physical
VADv275.4/50.576.1/51.476.0/51.476.2/51.8
DriveTransformer-L62.3/36.863.3/38.263.4/37.763.5/38.6
HiP-AD86.3/69.186.6/69.586.8/69.586.7/70.0
AlignDrive88.4/71.888.4/72.388.7/72.388.5/72.7
SparseDriveV289.2/70.589.6/71.489.5/71.489.6/71.8

Clean scores alone give a limited separation between the full and raw-branch variants. We therefore use the appearance-induced collision-error increment from the preceding reply and calculate direct differences of increments. A negative value favors the physically mediated comparator.

HostPhysical branch minus raw, 95% CIFull CIB minus raw, 95% CI
VADv2-1.2 [-3.8, 1.4]-1.0 [-3.6, 1.6]
DriveTransformer-L-1.1 [-4.1, 1.8]-0.9 [-3.9, 2.0]
HiP-AD-0.9 [-2.7, 1.2]-0.7 [-2.6, 1.4]
AlignDrive-0.5 [-2.4, 1.6]-0.3 [-2.2, 1.8]
SparseDriveV2-0.8 [-2.8, 1.2]-0.5 [-2.5, 1.5]

VeteranAD+'s corresponding native corruption-loss contrast is -0.6 [-1.6, 0.5] for the physical branch minus raw, and -0.5 [-1.5, 0.7] for Full CIB minus raw. These are PDMS-loss differences, not collision-rate differences.

The point-estimate pattern is informative: adding a physical branch preserves the smaller error increment, while adding a raw branch moves the response toward the original host. Because every direct interval includes zero, this is directional evidence for the communication-content hypothesis; it does not statistically isolate the incremental contribution of the restriction.

We accordingly sharpened the paper's contribution to a physically grounded communication interface with local structural guarantees. The name is now Constrained Interaction Bottleneck. On replaced edges, source-dependent messages use predicted physical states and mode probabilities. The network's retained target, temporal, and other pathways remain part of the planner. The controlled experiments evaluate this architectural hypothesis within their stated scope.

Manuscript locations: Table 27 on p. 42; Table 28 on p. 43; Table 30 on p. 45; Table 31 on p. 46; Table 32 on p. 47; Table 33 on p. 48; Table 34 on p. 49; Table 35 on p. 50; Table 36 on p. 51; Table 37 on p. 52; Table 38 on p. 53; Table 39 on p. 53; Section 3.2 on p. 4, Section 4.3 on p. 7.

AuthorsOpenReview ↗

Post 3 of 5: Q2. Multiple strong planners and uncertainty across training runs

We expanded the study to VADv2, DriveTransformer-Large, HiP-AD, AlignDrive, SparseDriveV2, and VeteranAD+. This includes planner families beyond the SparseDrive hosts used in the original controlled study. HiP-AD is treated as layerwise trajectory refinement; VeteranAD+ is evaluated in its native NAVSIM-v1 setting. We keep these settings explicit rather than treating every refinement mechanism as temporal autoregression.

For each main host-versus-CIB comparison, we report three training seeds, 1, 42, and 123. Within-seed CIB-minus-host differences determine the paired Student-t 95% interval with two degrees of freedom. The full tables include individual runs and mean plus sample standard deviation.

HostMean DS, host to CIBDS difference, 95% CIMean SR, host to CIBSR difference, 95% CI
VADv274.433 to 75.6001.167 [-4.058, 6.391]49.533 to 51.4001.867 [-4.965, 8.698]
DriveTransformer-L61.367 to 63.0671.700 [-4.014, 7.414]35.767 to 38.2002.433 [-7.768, 12.634]
HiP-AD85.667 to 86.1670.500 [-3.498, 4.498]68.033 to 69.4001.367 [-4.349, 7.082]
AlignDrive87.933 to 87.833-0.100 [-4.098, 3.898]70.933 to 71.8330.900 [-2.578, 4.378]
SparseDriveV288.700 to 89.3330.633 [-2.351, 3.618]69.367 to 71.4002.033 [-1.952, 6.018]

VeteranAD+'s mean PDMS is 90.900 for the host and 91.133 with CIB; the paired increment is 0.233 [-0.771, 1.237]. SR differences are percentage points.

The repeated results expose both transfer and variability. Mean SR improves on all five Bench2Drive hosts, while AlignDrive's mean DS is nearly unchanged. Every paired interval crosses zero, so the current three-seed evidence does not resolve a reliable gain for each strong model. The revised claim is that CIB transfers across independent implementations and exhibits host-dependent benefits, with uncertainty stated explicitly.

The original controlled topologies were also repeated. Their reported three-seed means are:

SparseDrive variantL2, host to CIBPDMS, host to CIBDS, host to CIBSR, host to CIB
Sequential0.612 to 0.56381.7 to 85.143.7 to 46.517.4 to 19.4
Parallel0.581 to 0.54782.9 to 86.345.0 to 48.118.7 to 21.2
Iterative0.621 to 0.56583.6 to 87.747.2 to 50.320.4 to 22.0

These favorable aggregate directions complement the strong-host estimates. Together with the direct branch controls, they make clear where the design has been implemented, what was compared, and how much the available measurements establish.

Manuscript locations: Table 30 on p. 45, Table 32 on p. 47, Table 34 on p. 49, Table 36 on p. 51, Table 27 on p. 42, Table 38 on p. 53, Table 1 on p. 7; Section 4.5 on p. 8.

Reviewer_5troOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies architectural shortcut pathways in end-to-end autonomous driving and proposes the Causal Interaction Bottleneck (CIB). On selected cross-entity interaction edges, source latent features are mapped to type-specific kinematic or geometric hypotheses and mode probabilities; these mediators are re-encoded and used for target-conditioned attention, while raw source keys/values are not allowed to bypass the mediator. The authors provide architecture-level mediation and information non-increase results, and evaluate CIB on sequential, parallel, and iterative SparseDrive variants across nuScenes, NAVSIM, and Bench2Drive. Matched controls (DSB, CIB without physical supervision), linear probes, and CARLA paired interventions are used to argue that the gains are not only from a generic bottleneck and that appearance sensitivity is reduced. Transfer to SparseDriveV2 and comparison with CausalVAD/SCIS are also reported.

Critical strengths
  1. Unconstrained high-dimensional cross-entity communication is a practically relevant problem, and CIB offers a concrete, implementable inductive bias (physically mediated interaction) for modern query-based planners. This practical relevance is a key reason the work is worth engaging with.
  2. Beyond headline planning metrics, the paper uses matched controls, linear probes, and paired interventions, and writes with relatively careful scope language around classical front-door identification. This experimental discipline is consequential: it makes the paper more serious than a pure “add module → higher score” E2E report, even if I disagree with the causal interpretation of the gains.
Critical weaknesses
  1. The causal / front-door framing overstates what is established, which is central to my recommendation. CIB only enforces local structural mediation on replaced interaction edges; it does not identify real-world causal effects, nor does it remove shortcuts from the full network (target residuals, temporal self-attention, source selection, and iterative residual states remain). Current evidence (reduced linear appearance decodability; lower NIS under CARLA weather changes) does not prove that appearance/context shortcuts cause the relevant failures on nuScenes/NAVSIM, nor that gains come from shortcut removal rather than physical inductive bias plus auxiliary supervision. Renaming and substantially weakening the causal narrative would better match the method.
  2. Empirical support for “consistent / complementary” gains is weaker than claimed, and key comparisons are missing. Cross-paradigm results in Table 1 appear single-run; SCIS numbers are partly cited rather than fully reimplemented under the same stack; SparseDriveV2 transfer also lacks multi-seed evidence. Efficiency cost is non-trivial (≈53% more FLOPs, much higher training memory, FPS 8.4→7.1) and may discard useful semantics (turn signals, gestures, traffic-light appearance), but these trade-offs are not measured prominently. Moreover, iterative-paradigm experiments omit recent strong baselines Hip-AD and AlignDrive, and validating CIB mainly by adding it on SparseDrive raises the risk that observed gains are coincidental to that stack.
Questions for authors
  1. Causal claim scope (clarification; critical): If CIB only mediates selected edges and leaves residual/self-attention/source-selection pathways intact, can you restate the contribution without claiming front-door-style causal identification or network-level shortcut removal? What exact claims do you still want to defend after rebuttal?
  2. Attribution of gains (clarification; critical): What evidence shows that improvements come from shortcut removal rather than physical inductive bias + auxiliary supervision? Please report stronger ablations such as raw-feature bypass, mode-prior-only variants, and semantic interventions, preferably with multi-seed statistics.
  3. Baseline completeness (clarification; critical): Why were Hip-AD and AlignDrive not included in the iterative-paradigm comparisons? Can you add them under the same evaluation protocol?
  4. Stack coincidence (clarification; critical): Since CIB is primarily validated by inserting into SparseDrive-family models, what evidence argues the gains are not stack-specific / coincidental? Please provide multi-seed results for Table 1 / SparseDriveV2 transfer, and clarify which SCIS numbers were reimplemented vs. cited.
  5. Cost–semantics trade-off (clarification): Can you quantify whether restricting interaction to kinematic/geometric mediators discards useful semantic cues (signals/gestures/lights), and discuss the ≈53% FLOPs / memory / FPS cost more prominently in the main paper?
14 comments
AuthorsOpenReview ↗

Response 1/7: Claim scope and complete-interface evidence

Thank you for identifying the distinction between a local communication constraint and a causal claim about the entire driving system. We have used that distinction to revise the terminology, theoretical statements, and empirical interpretation. CIB now stands for Constrained Interaction Bottleneck.

The contribution is a physical interface for selected cross-entity interactions. A source predicts kinematic or geometric hypotheses and their mode probabilities. The target receives an attention update constructed from this interface. For a fixed target state, source set, mask, and other source mediators, equal mediator representations produce equal CIB updates. This is the precise structural property established in Section 3.2 on p. 4 and Appendix A.2 on p. 13. The target residual, temporal interaction, source selection, and iterative state remain part of the host architecture. The conditional information result is a non-increase bound under deterministic mediation, rather than a guarantee that all appearance information disappears. Front-door identification is discussed only as historical motivation in Appendix A.1 on p. 12; the method does not identify real-world causal effects or establish network-wide shortcut removal.

The revised empirical question is whether this physical communication interface improves planning and sensitivity to the tested appearance changes, and how much of that behavior is attributable to representation, supervision, or access to raw features. We added raw-feature bypass controls, matched physical branches, a state/prior decomposition, collision-error measurements, and multi-seed strong-host comparisons to address these questions directly.

We also extended the appearance probes from probability-weighted summaries to the complete communicated interface, including every mode and its probability. The driving model remains frozen. Balanced accuracy measures four-way appearance decodability; chance is 25%. Lower physical nMAE indicates better physical-state recovery.

RepresentationLinear BA %Set-probe BA %Physical nMAE
Raw source query88.491.20.113
Full mediator set and priors63.872.60.085
Full encoded set and priors68.576.00.088
Actual message, fixed target61.769.90.092
Appearance-label permutation25.025.0—

These results support an interface more aligned with physical state: physical recovery improves while appearance is less accessible to the tested probes, and residual appearance information remains measurable. The complete-set test also avoids treating a compressed diagnostic summary as the actual communication channel. See Table 7 on p. 32.

The resulting contribution is an explicit, analyzable communication constraint with measurable physical and behavioral consequences. We have aligned the paper's claims with that contribution throughout the main text and appendix.

AuthorsOpenReview ↗

Response 2/7: Attribution: physical states, mode priors, and raw-feature bypass

To address Q2, we added controls that keep physical predictions and their supervision while changing how source information reaches the target. The raw branch restores access to source latent features alongside CIB. The physical extra branch provides a corresponding capacity control using the physical interface. We also compare auxiliary physical supervision on the original host, DSB, and CIB without physical supervision.

The following table reproduces the detailed-ablation DS point estimates. These ablations are distinct from the three-seed primary comparisons in the next reply. DT denotes DriveTransformer-Large; V2 denotes SparseDriveV2 with the Bench2Drive R50 configuration.

VariantVADv2DTHiP-ADAlignDriveV2
Host74.861.685.988.388.9
Host + matched capacity75.061.986.188.589.0
Host + physical auxiliary75.462.386.388.489.2
DSB75.362.186.087.989.1
CIB without physical supervision75.662.786.288.189.3
Full CIB76.163.386.688.489.6
CIB + raw branch76.063.486.888.789.5
CIB + physical extra branch76.263.586.788.589.6

The physical representation and supervision each contribute in several hosts, while clean-condition DS alone does not isolate a separate benefit from excluding raw features. In particular, the raw branch sometimes matches or exceeds Full CIB. This motivates evaluating the communication restriction through paired appearance-induced errors, which we report separately below.

The requested mode-prior control separates predicted states from learned probabilities. The fixed templates are distinct, so changing their probabilities remains meaningful.

State / priorVADv2 DSDT DSHiP-AD DSAlignDrive DSV2 DSVeteranAD+ PDMS
Predicted / learned76.163.386.688.489.691.3
Predicted / uniform75.863.086.488.589.391.1
Fixed / learned74.360.885.287.188.590.6
Fixed / uniform73.960.484.986.888.290.3

The clearest descriptive pattern is that retaining scene-dependent physical hypotheses matters more than reweighting fixed templates. Learned priors provide smaller additional changes, with AlignDrive showing no DS advantage over the uniform prior. Thus, the results support the physical content of the interface as a substantive part of the method rather than treating the gain as a generic attention-prior effect.

Full DS/SR controls and the native NAVSIM analysis appear in Table 30 on p. 45, Table 32 on p. 47, Table 34 on p. 49, Table 36 on p. 51, Table 27 on p. 42, and Table 38 on p. 53. We interpret these point estimates alongside the repeated-run and direct attribution intervals, with no claim that a single ablation establishes an exclusive cause of the planning gains.

AuthorsOpenReview ↗

Response 3/7: Recent strong planners and independent training runs

Thank you for pointing out HiP-AD and AlignDrive. The original S/P/I experiment used controlled variants of a common host to study interaction structure. It did not provide the broader recent-planner comparison requested here. We now include both models, together with VADv2, DriveTransformer-Large, and SparseDriveV2, using matched host/CIB comparisons within each implementation on Bench2Drive. VeteranAD+ is evaluated separately under its native NAVSIM-v1 protocol.

The expanded study distinguishes architectures rather than assigning every refinement mechanism to the same category. HiP-AD repeatedly refines planning queries through perception and image interactions. AlignDrive conditions longitudinal planning on the predicted path. These provide relevant tests of CIB beyond the original controlled iterative variant. Their descriptions and citations are added to Section 2 on p. 2.

Below are means over training seeds 1, 42, and 123. Each difference is CIB minus the matched host. The 95% intervals use paired training-run differences with the Student t critical value for two degrees of freedom. SR differences are percentage points.

HostDS host → CIBDS difference, 95% CISR host → CIB %SR difference, 95% CI
VADv2 R5074.433 → 75.600+1.167 [-4.058, 6.391]49.533 → 51.400+1.867 [-4.965, 8.698]
DriveTransformer-Large R5061.367 → 63.067+1.700 [-4.014, 7.414]35.767 → 38.200+2.433 [-7.768, 12.634]
HiP-AD R50 stage 285.667 → 86.167+0.500 [-3.498, 4.498]68.033 → 69.400+1.367 [-4.349, 7.082]
AlignDrive R50 full87.933 → 87.833-0.100 [-4.098, 3.898]70.933 → 71.833+0.900 [-2.578, 4.378]
SparseDriveV2 B2D R5088.700 → 89.333+0.633 [-2.351, 3.618]69.367 → 71.400+2.033 [-1.952, 6.018]
Native NAVSIM-v1 hostPDMS host → CIBDifference, 95% CI
VeteranAD+ R3490.900 → 91.133+0.233 [-0.771, 1.237]

The mean SR increases in all five Bench2Drive hosts, and mean DS increases in four. These results expand the tested implementation scope beyond SparseDrive, while the paired intervals leave the strength of the performance benefit unresolved for each strong host. We therefore describe architectural applicability and favorable mean trends, rather than a statistically established improvement across all recent planners. AlignDrive's nearly unchanged DS also shows why the method should be assessed together with the existing host representation and its computational cost.

All three per-seed results and sample standard deviations are retained in Table 30 on p. 45, Table 32 on p. 47, Table 34 on p. 49, Table 36 on p. 51, Table 27 on p. 42, and Table 38 on p. 53. Published leaderboard values are kept separate from these matched local runs. This directly addresses both baseline completeness and training variability without substituting a literature score for the host checkpoint used in the comparison.

Reviewer_5p5ZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors argue that unconstrained latent interactions may propagate context-dependent shortcut cues. To this end, they propose the Causal Interaction Bottleneck (CIB), which makes explicit physical kinematic or geometric predictions as an intermediate representation via differentiable integration for cross-entity communication. Then, these explicit states are re-encoded for target-conditioned attention with predicted mode probabilities as attention priors. Finally, raw source features cannot bypass the mediator on the replaced interaction edges, while auxiliary state supervision is added to the original objectives. Extensive experiments are conducted on nuScenes, NAVSIM and Bench2Drive based on the variants of SparseDrive, showing obvious and consistent performance gains. However, I still have several concerns, as detailed below.

Critical strengths
  1. This paper introduces a novel physical inductive bias by requiring the selected cross-entity interactions to pass through explicit motion or geometry predictions, which allows physical constraints and supervision to directly shape the information adopted for motion planning.

  2. This paper provides a comprehensive evaluation of CIB on various benchmarks, including open-loop nuScenes, NAVSIM, and closed-loop Bench2Drive. Table 1 shows consistent performance gains over three SparseDrive variants.

  3. Table 2 compares CIB with a direct-state bottleneck (DSB) that retains state supervision and the interaction mechanism but removes physical integration, supporting the contribution of physical parameterization and supervision.

Critical weaknesses
  1. Both DSB and CIB route information through intermediate states, while existing comparisons do not isolate the benefit of blocking raw source features. A variant that retains the same physical predictions, supervision, and interaction modules while additionally allowing raw source features to reach the target would help assess the contribution of the no-bypass constraint to planning performance.

  2. Table 12 reports substantial computational overhead on SparseDrive. However, Table 18 shows relatively modest improvements over SparseDriveV2, i.e., DS increases from 89.0 to 89.9 and SR from 69.5% to 71.5% on bench2drive. It would be better if this paper provides further validation of the gains on a stronger baseline.

  3. CIB leverages explicit physical predictions to mediate cross-entity communication, making their accuracy relevant to their interpretation as reliable motion or geometry hypotheses. However, Table 3 measures information recoverability and Table 11 checks the upstream performance. The performance of intermediate outputs should be clarified with appropriate evaluation.

Questions for authors

It is suggested that the authors provide further clarification and supporting evidence to address the weaknesses discussed above.

6 comments
AuthorsOpenReview ↗

Post 1 of 5: W1. Direct controls for the communication constraint

Thank you for identifying the missing comparison between physically mediated communication and the same model with raw-feature access. We added the requested raw branch while retaining CIB's physical predictions, supervision, and interaction modules. A second control adds a physical branch to distinguish communication content from the benefit of an additional branch. We also include capacity, auxiliary-supervision, direct-state bottleneck, and controls without explicit mediator supervision.

The following are point-estimate diagnostics; training-seed variability is not reported for these ablations. DS/SR denotes Driving Score and route Success Rate in percent. VeteranAD+ uses native NAVSIM-v1 PDMS. These ablations are separate from the three-seed host-versus-CIB comparisons in our third reply.

VariantVADv2 DS/SRDriveTransformer-L DS/SRVeteranAD+ PDMS
Host74.8/49.561.6/35.591.0
Matched capacity75.0/49.561.9/35.091.1
Physical auxiliary75.4/50.562.3/36.891.2
DSB75.3/50.062.1/36.490.9
CIB without physical supervision75.6/50.562.7/37.391.0
Full CIB76.1/51.463.3/38.291.3
CIB + raw branch76.0/51.463.4/37.791.4
CIB + physical branch76.2/51.863.5/38.691.5
VariantHiP-AD DS/SRAlignDrive DS/SRSparseDriveV2 DS/SR
Host85.9/68.288.3/71.488.9/69.5
Matched capacity86.1/68.288.5/71.489.0/70.0
Physical auxiliary86.3/69.188.4/71.889.2/70.5
DSB86.0/68.687.9/70.989.1/70.5
CIB without physical supervision86.2/69.188.1/71.489.3/70.9
Full CIB86.6/69.588.4/72.389.6/71.4
CIB + raw branch86.8/69.588.7/72.389.5/71.4
CIB + physical branch86.7/70.088.5/72.789.6/71.8

The comparisons clarify the contribution. Physical auxiliary supervision alone does not reproduce Full CIB's point estimates on most hosts. At the same time, clean scores are close between CIB and its raw-branch variants. Thus, the architectural restriction has a precise meaning, but its empirical benefit should be assessed through both task performance and sensitivity to the information being restricted. Our next reply reports that targeted test.

We now call the module the Constrained Interaction Bottleneck. Its guarantee is local: source information on the replaced edges passes through the predicted states and mode probabilities. Target residuals, temporal pathways, and other retained interfaces remain available. This framing makes the contribution an explicit and testable communication design.

Manuscript locations: Table 30 on p. 45, Table 32 on p. 47, Table 34 on p. 49, Table 36 on p. 51, Table 27 on p. 42, Table 38 on p. 53; Section 3.2 on p. 4.

AuthorsOpenReview ↗

Post 2 of 5: W1. Appearance changes, actual errors, and physical responsiveness

We supplemented the clean-score ablation with matched appearance conditions and measured collision errors together with progress. For each model, the error increment is its changed-condition error rate minus its reference-condition error rate. The table below reports these increments in percentage points. The physical-branch control has the same purpose as in our first reply: add communication capacity while keeping source content physically mediated.

HostOriginal hostCIB + rawCIB + physicalFull CIBChanged progress, host to CIB
VADv24.33.82.62.889.7 to 91.9
DriveTransformer-L3.73.22.12.385.8 to 88.8
HiP-AD2.82.51.61.894.0 to 95.2
AlignDrive2.32.01.51.795.4 to 96.0
SparseDriveV22.72.41.61.995.4 to 96.3

The direct comparison subtracts the raw-branch error increment from the physical-branch or Full-CIB increment. It therefore asks whether the added raw access changes robustness, rather than comparing two separate confidence intervals.

HostPhysical branch minus raw, 95% CIFull CIB minus raw, 95% CI
VADv2-1.2 [-3.8, 1.4]-1.0 [-3.6, 1.6]
DriveTransformer-L-1.1 [-4.1, 1.8]-0.9 [-3.9, 2.0]
HiP-AD-0.9 [-2.7, 1.2]-0.7 [-2.6, 1.4]
AlignDrive-0.5 [-2.4, 1.6]-0.3 [-2.2, 1.8]
SparseDriveV2-0.8 [-2.8, 1.2]-0.5 [-2.5, 1.5]

Every point estimate favors the physical communication variants, and their better changed-condition progress makes uniformly more conservative driving an incomplete explanation. The intervals still include zero, so we interpret the direct raw-versus-physical comparison as directional evidence, with the size of the incremental effect unresolved.

Physical interventions provide a complementary check. False Braking Rate measures unnecessary braking after removal of the relevant actor. Correct-response rates are reported separately for slower and faster actors.

HostFBR, host to CIBSlower-actor correct, host to CIBFaster-actor correct, host to CIB
VADv218.6 to 11.377.3 to 83.174.1 to 80.2
DriveTransformer-L22.4 to 14.272.8 to 79.869.5 to 76.1
HiP-AD12.7 to 9.385.6 to 88.182.8 to 85.3
AlignDrive9.8 to 8.589.1 to 89.786.7 to 87.5
SparseDriveV210.7 to 7.888.7 to 90.886.0 to 87.9

Together, these results characterize selective sensitivity: smaller responses to the tested appearance changes coexist with useful responses to changed physical interactions. The claim concerns these controlled conditions and selected communication edges.

Manuscript locations: Table 31 on p. 46, Table 33 on p. 48, Table 35 on p. 50, Table 37 on p. 52, Table 28 on p. 43; Section 4.4 on p. 8.

AuthorsOpenReview ↗

Post 3 of 5: W2. Stronger hosts and repeated training runs

We expanded the transfer study from SparseDriveV2 to VADv2, DriveTransformer-Large, HiP-AD, AlignDrive, and VeteranAD+. All hosts except VeteranAD+ use matched Bench2Drive comparisons; VeteranAD+ is evaluated on its native NAVSIM-v1 protocol. Published scores are kept separate from our local host-versus-CIB measurements.

Each main comparison reports seeds 1, 42, and 123. We calculate the paired difference within each seed and its Student-t 95% interval with two degrees of freedom. This quantifies the variation among the three paired runs.

HostMean DS, host to CIBPaired DS difference, 95% CIMean SR, host to CIBPaired SR difference, 95% CI
VADv274.433 to 75.6001.167 [-4.058, 6.391]49.533 to 51.4001.867 [-4.965, 8.698]
DriveTransformer-L61.367 to 63.0671.700 [-4.014, 7.414]35.767 to 38.2002.433 [-7.768, 12.634]
HiP-AD85.667 to 86.1670.500 [-3.498, 4.498]68.033 to 69.4001.367 [-4.349, 7.082]
AlignDrive87.933 to 87.833-0.100 [-4.098, 3.898]70.933 to 71.8330.900 [-2.578, 4.378]
SparseDriveV288.700 to 89.3330.633 [-2.351, 3.618]69.367 to 71.4002.033 [-1.952, 6.018]

For VeteranAD+, mean PDMS changes from 90.900 to 91.133, with a paired difference of 0.233 and interval [-0.771, 1.237]. SR differences above are percentage points.

These results provide a more complete answer than the original single SparseDriveV2 transfer. They demonstrate implementation across independent planner families and expose the variation in the additional benefit. Four of five Bench2Drive hosts improve in mean DS and all five improve in mean SR; AlignDrive's mean DS is essentially unchanged. All paired intervals include zero, so the revised paper presents host-dependent effect estimates rather than a claim of statistically resolved gains on every strong planner.

For the original controlled SparseDrive variants, the new three-seed means retain favorable directions across all three benchmarks:

VariantL2, host to CIBCR, host to CIBPDMS, host to CIBDS, host to CIBSR, host to CIB
Sequential0.612 to 0.5630.119 to 0.08281.7 to 85.143.7 to 46.517.4 to 19.4
Parallel0.581 to 0.5470.081 to 0.06582.9 to 86.345.0 to 48.118.7 to 21.2
Iterative0.621 to 0.5650.102 to 0.07183.6 to 87.747.2 to 50.320.4 to 22.0

These are aggregate means; the strong-host intervals above are computed from the separately reported seed values. The combined evidence supports physically mediated interaction as a portable design whose marginal value depends on the host's existing representations.

Manuscript locations: Table 30 on p. 45, Table 32 on p. 47, Table 34 on p. 49, Table 36 on p. 51, Table 27 on p. 42, Table 38 on p. 53, Table 1 on p. 7; Section 4.5 on p. 8.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

We thank all reviewers for their constructive feedback. The revised manuscript and our detailed responses sharpen the contribution of CIB and provide additional evidence on its mechanism, reliability, and practical trade-offs.

An explicit physical communication interface with local guarantees. CIB makes predicted physical hypotheses and their probabilities the source-side interface on selected cross-entity interactions. The formal claims describe local structural mediation and conditional information non-increase. We have aligned the terminology and interpretation with this scope. The contribution is a testable constraint on communication content in end-to-end driving. See Section 3.2 and Appendix A.

Repeated experiments support the controlled cross-paradigm results. Across sequential, parallel, and iterative SparseDrive variants, the three-seed means retain favorable directions on nuScenes, NAVSIM, and Bench2Drive. Mean PDMS increases by 3.4, 3.4, and 4.1 points, respectively. Table 1 reports these repeated results; the matched controls in Table 2 further support the role of physical parameterization and supervision.

New controls clarify the role of physical information content. We added raw-feature branches, matched physical branches, and a decomposition of predicted states and mode priors. Across all six host-specific diagnostic sets, predicted states with uniform priors outperform fixed templates with learned priors. This descriptive pattern supports scene-dependent physical content beyond probability reweighting. Paired appearance tests also measure collision errors together with progress: on all five closed-loop hosts, CIB has smaller point-estimated error increments and higher changed-condition progress than the host. These tests connect appearance sensitivity to driving outcomes under the evaluated interventions. See Appendix G.

The evaluation now extends to six recent planners with repeated training runs. Matched three-seed comparisons cover VADv2, DriveTransformer-Large, HiP-AD, AlignDrive, SparseDriveV2, and VeteranAD+. The first five use Bench2Drive; VeteranAD+ uses native NAVSIM v1. This broadens the implementation scope beyond the original SparseDrive study. All paired intervals for the main strong-host score differences include zero, as do the direct raw-versus-physical contrast intervals. We therefore distinguish favorable mean trends from statistically resolved gains and retain uncertainty in the attribution. Table 4 provides the main-score summary; Appendix G reports the full comparisons.

Mediator quality and practical trade-offs are evaluated directly. We evaluate the tensors actually passed to the re-encoder, probe the complete communication interface, and jointly examine red-light violations and unnecessary stopping at green lights. Computational cost is now discussed in the main text, with component profiling and optimized implementations of both Host and CIB. On SparseDrive-P, optimization reduces CIB's median latency from 137 to 122 ms at the same reported PDMS of 86.3; the optimized host takes 108 ms at PDMS 82.9. These measurements make the quality–cost trade-off explicit. See Tables 7, 16, and 18–19, and Section 4.6.

The revised evidence supports physically structured communication as an inspectable inductive bias whose value depends on the host architecture and evaluated conditions. The reviewer-specific responses include the relevant tables and manuscript locations. We would welcome feedback on any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.