Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Towards foundational LiDAR world models with efficient latent flow matching
Authors: Tianran Liu, Shengwen Zhao, Nicholas Rhinehart
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our Li DAR world model significantly improves the convergence speed of downstream tasks. For all mentioned tasks, we observed relative performance gains with varying amounts of fine-tuning data. In particular, for semantic occupancy forecasting, this scheme of learning dynamics prior to semantics patterns allows us to achieve superior performance to Occ World [58] with only 5% of its required labeled data. |
| Researcher Affiliation | Academia | Tianran Liu Shengwen Zhao Nicholas Rhinehart EMAIL EMAIL University of Toronto |
| Pseudocode | No | The paper describes its methods and training objectives in textual paragraphs and mathematical formulations within sections 3.1, 3.2, and 3.3, but does not present any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our project page contains additional visualizations and released code. |
| Open Datasets | Yes | We use nu Scenes[8] (2Hz) as the pretraining data for the flow matching model. ... For the beam adaptation subtask, we down-sampled 11 sequences from the KITTI360 raw dataset [25] ... |
| Dataset Splits | Yes | For the indoor navigation dataset... (training set with 23,504 frames and validation set with 9,720 frames). ... Table 10: Training-validation set splitting for each dataset ... os 19728 4219 ... Indoor Li DAR Dataset 23504 9720 |
| Hardware Specification | Yes | we used 4x RTX 4090 to train models for 200 epochs with a batch size of 8. ... All FPS tested based on 1x RTX4090 without kernel fusion or other CUDA acceleration methods. |
| Software Dependencies | No | Following prior work[56], we also adopt the Adam W as the optimizer with β1 and β2 set to 0.9 and 0.99 for flow matching training, 0.99 and 0.999 for VAE training. |
| Experiment Setup | Yes | we used 4x RTX 4090 to train models for 200 epochs with a batch size of 8. We trained the VAE part for 100 epochs with a batch size of 16 if not otherwise specified. Following prior work[56], we also adopt the Adam W as the optimizer with β1 and β2 set to 0.9 and 0.99 for flow matching training, 0.99 and 0.999 for VAE training. We set the weight decay of all normalization layers to 0 and all of the other layers to 0.001. The learning rate schedule has a linear warmup followed by cosine decay (with the minimum of the cosine decay set to be 20% of the peak learning rate). We also use EMA with a 0.9999 decay rate to ensure the updates of parameters are stable. |