acceptodds
Withdrawn from ICLR 2027

MapDLM: Lane-Aligned Block Diffusion for Lane Graph Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Constructing HD maps from multi-frame LiDAR BEV representations requires predicting lane-centerline geometry and directed lane connectivity. Generic vision–language model tokenizers are not designed to represent coordinates, while long token sequences weaken the association between coordinate points and the lane geometry and topology they define. We present MapDLM, a lane-supervised multi-token generation model based on lane-aligned block diffusion. MapDLM first predicts a lane-ID list and directed lane connectivity, then reconstructs each lane as fixed-length token blocks conditioned on this structure. In addition to the standard cross-entropy loss, training combines point-level coordinate objectives, lane-level polyline constraints, and topology-chain supervision. Anchor-Point Gaussian Loss (APG Loss) tolerates errors between nearby coordinate bins in lane interiors while retaining exact targets at lane endpoints. Semantic-Adaptive Noise Scheduling and structure-aware commitment expose the model to the highly masked states encountered during iterative denoising. By jointly predicting multiple lanes within each diffusion block, MapDLM reduces serial decision depth and error propagation. We evaluate MapDLM on Argoverse 2 and nuScenes under the OpenLane-V2 protocol, compare it with corresponding autoregressive baselines, and conduct controlled ablations of lane alignment, multi-level supervision, and state-aligned denoising. The results show effective geometry and topology reconstruction on both benchmarks and support lane-aligned diffusion as a promising approach to structured map generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
BEV road maps, HD map, vision-language models, spatial understanding, lane topology

Reviews

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

The paper proposes MapDLM, a block-diffusion VLM that generates the centerline geometry and directed connectivity of lanes from a BEV image rendered from accumulated LiDAR. The key idea is to organize the output around lanes: the model first generates the topology and a lane inventory, and then recovers coordinates in 32-token blocks composed of fixed-length records that preserve lane IDs. Training combines a sinusoidal encoding of coordinate values (CVSE), an Anchor-Point Gaussian (APG) loss that uses exact targets at lane endpoints while giving partial credit to nearby values at interior points, polyline-level geometric losses, and role-dependent masking distributions and loss weights. On Argoverse 2 and nuScenes with OpenLane-V2 lane annotations, the paper reports higher lane detection and topology scores than in-house AR VLM baselines under the same LiDAR-BEV evaluation setting, as well as a 42.2%42.2\% reduction in output length relative to plain text tokens.

Critical strengths
  1. The paper identifies a concrete obstacle to applying block diffusion to lane graphs and resolves it with a simple representation. An arbitrary block partition of a serialized graph can separate structural tokens from the coordinates they label, and the authors observe that applying block diffusion to the original token sequence yields nearly input-independent outputs in free-running generation. Switching to a lane-aligned representation, in which each block consists of fixed-length records that preserve lane IDs, avoids this collapse (variants A and B in Table 3). This observation is useful for work that applies diffusion language models to structured outputs.
Critical weaknesses
  1. The choice of a generative approach is not adequately motivated, and there is no comparison with existing methods to support it. Query-based detection and topology models (e.g., TopoMLP [1], TopoLogic [2]) are already widely used for lane-graph prediction, yet beyond noting a shared interface for graph relations and variable-length geometry, the paper takes the generative token-based formulation as given without arguing why it should be preferable. The experiments do not answer this question either. The paper adopts a non-standard setting, with accumulated LiDAR BEV input and planar matching that discards height, but does not rerun any existing topology model under the same setting; the only comparisons are in-house AR VLM baselines, which reach only 9.59.5–10.710.7 DET on AV2. As a result, neither the benefit of the generative approach nor the actual performance level of MapDLM can be assessed.

  2. The source of the performance gain is not isolated. The AR baselines and MapDLM differ in initialization, output format, losses, and decoding, and the authors acknowledge that this comparison does not isolate the effect of the decoding scheme. The ablations do not separately evaluate CVSE, the APG loss, the polyline losses, or topology-first serialization. The largest gain in the ablations (DET from 18.118.1 to 32.932.9) comes from combining section-weighted CE with Beta masking, which the authors themselves describe as related to Fast-dDrive rather than a standalone novelty. Since no results are reported over multiple seeds, the central claim that lane-aligned block diffusion is promising lacks controlled evidence.

References

[1] D. Wu et al., "TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning," ICLR 2024.

[2] Y. Fu et al., "TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes," NeurIPS 2024.

Questions for authors
  1. How many LiDAR frames are accumulated, and does the accumulation include frames after the reference timestamp? If future frames are included, the setting corresponds to offline rather than online mapping, which changes how the results should be interpreted.

  2. DET and TOP are AP-based metrics. How are the confidence scores of the generated lanes and connections constructed?

  3. Could the authors retrain and compare existing topology models (e.g., TopoMLP [1], TopoLogic [2]) under the same LiDAR BEV input and evaluation setting?

  4. The AR baselines appear not to use the lane-aligned serialization; please specify their exact output format. A comparison with an AR baseline that uses the same lane-aligned format and losses would better isolate the effect of the decoding scheme (AR vs. block diffusion).

References

[1] D. Wu et al., "TopoMLP: A Simple yet Strong Pipeline for Driving Topology Reasoning," ICLR 2024.

[2] Y. Fu et al., "TopoLogic: An Interpretable Pipeline for Lane Topology Reasoning on Driving Scenes," NeurIPS 2024.

Reviewer_Ks1TOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper adapts Fast-dVLM-3B to LiDAR-BEV (Image) lane-graph generation using topology-first serialization, persistent lane IDs, fixed 32-token diffusion blocks, coordinate-aware losses, and semantic-dependent masking. MAPDLM substantially outperforms the authors’ in-house autoregressive controls on AV2 and nuScenes.

Critical strengths

S1. Explicit lane-level modeling through block diffusion. Unlike prior autoregressive approaches that model a lane as a sequential stream of tokens, MAPDLM uses block diffusion to model each lane and recover its tokens jointly. The jointly recovered tokens therefore carry explicit lane-level meaning.

S2. Substantial generation-efficiency gains. Atomic coordinate tokens and the lane-aware layout reduce the mean response length about −42.2%, while geometry is packed into the same 32-token blocks used for iterative recovery. MapDLM also reports 96.7 tok/s, an 82% increase over the 53.2 tok/s of a vLLM-served Qwen2.5-VL-3B autoregressive baseline following Kwon et al. (2023). Together, these results suggest that the representation is substantially more compact and that block-wise recovery can provide markedly higher decoding throughput.

Critical weaknesses

W1. Limited methodological novelty (major). The overall framework largely follows Fast-dVLM-3B and adapts it to lane-graph generation. The subsequent modifications—including lane-aware serialization, coordinate-token handling, and associated training refinements—are relatively straightforward engineering adaptations. As currently presented, they do not constitute a sufficiently strong methodological contribution.

W2. Lack of comparison with publicly available methods (major). The experiments compare MapDLM mainly with the authors’ internal controls, while Table 7 only summarizes results from camera-based methods. This does not provide a direct or sufficiently fair comparison with existing public approaches. In particular, public methods such as MapTR and LaneSegNet should be adapted to take the same LiDAR-BEV image input and evaluated under the same data, split, and metric settings, enabling a more informative comparison of the proposed method’s actual performance gains.

Questions for authors

Q1. Figure 4 is not referenced in the main text. Under what experimental setting were its results obtained? Please specify the dataset, hardware, batch size, numerical precision, serving/decoding implementation, and measurement protocol used for both MapDLM and the autoregressive baseline.

Reviewer_HT4SOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MapDLM, a block-diffusion VLM for generating lane geometry and directed connectivity from accumulated LiDAR BEV images. The method combines topology-first serialization, persistent lane IDs, fixed-length lane records, geometric supervision, and semantic-dependent masking. Experiments on AV2 and nuScenes show substantial improvements over the authors’ autoregressive baselines under an internal planar BEV evaluation protocol.

Critical strengths
  1. The representation is well motivated: persistent lane IDs connect topology references with geometric records and support complete-lane supervision across block boundaries.
  2. The complete system achieves substantial gains in both lane detection and topology prediction on two datasets.
  3. The paper clearly acknowledges important limitations, including differences from official 3D evaluation and the restricted interpretation of its throughput measurements.
Critical weaknesses
  1. The main comparison does not isolate the central contribution. MapDLM and the AR baselines differ in initialization, serialization, and geometric objectives. The results therefore establish a complete-system improvement but do not determine the independent benefit of block diffusion. An AR baseline using the same lane representation and geometric losses, with matched training resources and initialization differences accounted for, would substantially strengthen the evidence.
  2. The ablations provide limited attribution to lane alignment. The original-sequence versus lane-aligned comparison changes several representation and recovery details, while the former reports only qualitative collapse. A controlled comparison between arbitrary and lane-aligned blocks is needed. CVSE, APG, polyline supervision, and structure-aware commitment also lack individual ablations.
  3. Competitiveness beyond the internal baselines remains unclear. No strong task-specific or existing generative mapping method is evaluated under the same LiDAR-BEV protocol. The published camera-based results in the appendix cannot establish a matched comparison.
  4. Important reproducibility details are missing. Appendix A.11 identifies confidence construction, resampling and matching settings, temporal accumulation, and scene-level splits as necessary for reproducing the scores, but does not specify them sufficiently. The reproducibility statement also mentions implementation code in the appendix, whereas no code or access link is apparent in the supplied PDF.
  5. Reliability and efficiency need stronger validation. The smaller ablation gains lack uncertainty estimates. Structural validity and inventory–geometry consistency are not quantified, despite irreversible prefix errors and heuristic commitment rules. Throughput measured with different token granularities and serving implementations does not establish scene-level latency improvements.
Questions for authors
  1. How does AR perform with the same lane serialization and geometric objectives?

  2. Can lane alignment be isolated while keeping tokenization, padding, losses, and decoding fixed?

  3. What are the exact confidence construction, matching, accumulation, and data-split settings?

  4. What fraction of generated outputs contain structural inconsistencies, and how much performance is lost due to predicted topology prefixes?

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

Related papers

Discussion (0)

Sign in to comment.