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..
Leveraging Conditional Dependence for Efficient World Model Denoising
Authors: Shaowei Zhang, Jiahan Cao, Dian Cheng, Xunlan Zhou, Shenghua Wan, Le Gan, De-Chuan Zhan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We begin by describing the experimental setup in Section 5.1. Subsequently, we address the following research questions: (1) Can Cs Dreamer enhance the training performance in complex environments with distractors? (Section 5.2); (2) What role does the decoupling regularization loss play in the overall objective, and how do the proposed modules affect performance? (Section 5.3); (3) Do the latent variables in the proposed framework exhibit interpretable semantics upon reconstruction visualization? (Section 5.4). |
| Researcher Affiliation | Academia | 1National Key Laboratory for Novel Software Technology, Nanjing University, China 2School of Artificial Intelligence, Nanjing University, China 3School of Intelligence Science and Technology, Nanjing University, China EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | B Pseudo Code The whole algorithm is shown in Algorithm 1. For brevity, we omit the episode continuation flag xt. Algorithm 1 Cs Dreamer |
| Open Source Code | Yes | 1The code is available at https://github.com/Zhang-Shaowei/Cs Dreamer. |
| Open Datasets | Yes | Benchmarks. We evaluate our model and baselines on visual control tasks. First, we assess performance on four Deep Mind Control Suite (DMC) [45] tasks: Walker Run, Cheetah Run, Finger Spin and Hopper Hop. To introduce noise distractors, we replace the original backgrounds with two types of task-irrelevant information. The first is a gray background composed of natural videos from the Kinetics 400 dataset [46], following the DBC [21] configuration (denoted as GB). The second is a colorful background derived from DAVIS 2017 videos [47], adhering to the background distractor settings in Distracting Control Suite [48] (denoted as CB). These benchmarks require the agent to extract task-relevant information, identify the target entity within the DMC environment, and effectively filter out background distractions. Subsequently, we evaluate these approaches in the more realistic simulated driving environment, CARLA [49]. Here, the agent must extract task-relevant information from visual perception while mitigating distractions such as trees and dynamic sunlight. We also conduct experiments on the complex Atari 100K benchmark [50] in Appendix E.2. |
| Dataset Splits | Yes | Then we alphabetically split the dataset into a training dataset (512 videos) and an evaluation dataset (129 videos) using an 80 : 20 ratio. |
| Hardware Specification | Yes | The experiments are mainly conducted on NVIDIA RTX 4090 GPUs. With each GPU, we are able to train each environment at a rate of approximately 24K timesteps per hour. |
| Software Dependencies | No | The paper does not explicitly mention specific version numbers for software libraries or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA) used in their implementation. |
| Experiment Setup | Yes | D.4 Hyperparameters and Time Cost Table 1 presents the primary hyperparameters of Cs Dreamer. Since the behavior policy relies solely on the feature of st, and the hyperparameters for st in Cs RSSM closely resemble those of the latent variables in RSSM of Dreamer V3, we adopt the same hyperparameters for the behavior policy as in Dreamer V3. Table 1: Hyperparameters for Cs Dreamer Hyperparameter Value Action Repeat 4 for CARLA and Atari, and 2 for others λ 10.0 for Hopper Hop+GB, 0.2 for CARLA and Atari, and 1.0 for others βs dyn 0.5 βs rep 0.1 βc dyn 0.5 βc rep 0.1 Discrete latent dimensions of st 32 Discrete latent classes of st 32 Discrete latent dimensions of ct 16 for CARLA, 8 for Atari, and 32 for others Discrete latent classes of ct 16 for CARLA, 8 for Atari, and 32 for others GRU recurrent units of st 512 GRU recurrent units of ct 256 for CARLA, 128 for Atari, and 512 for others Dense hidden units of st 512 Dense hidden units of ct 512 MLP layers 2 |