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..
TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting
Authors: Jaebin Lee, Hankook Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our framework consistently and significantly outperforms prior state-of-the-art baselines across a wide range of benchmark datasets. The code is available at https://github.com/efficient-learning-lab/Time Perceiver. |
| Researcher Affiliation | Academia | Jaebin Lee Sungkyunkwan University EMAIL Hankook Lee Sungkyunkwan University EMAIL |
| Pseudocode | Yes | Algorithm 1 TIMEPERCEIVER |
| Open Source Code | Yes | The code is available at https://github.com/efficient-learning-lab/Time Perceiver. |
| Open Datasets | Yes | We evaluate our model on 8 real-world multivariate time series datasets, including ETT (ETTh1, ETTh2, ETTm1 ETTm2) [22], Weather [1], Solar [21], Electricity [2], and Traffic [3] to ensure robustness across diverse temporal patterns and domains. |
| Dataset Splits | Yes | Following the standard protocol [7, 22] widely adopted in time series forecasting research, we perform a chronological split of each dataset. For the ETT datasets, we use a 6:2:2 ratio for training, validation, and testing. For the remaining datasets Weather, Solar, Electricity, and Traffic we apply a 7:1:2 split. |
| Hardware Specification | Yes | All the experiments were conducted using NVIDIA RTX 4090 GPUs with the Py Torch [28]. |
| Software Dependencies | No | The paper mentions "Py Torch [28]" but does not specify a version number for it or any other software dependency. The guidelines explicitly require specific version numbers for key software components. |
| Experiment Setup | Yes | Hyperparameter Settings. For the TIMEPERCEIVER parameter θ, we use the Adam W optimizer [26] with a weight decay of 0.05. The learning rate is set to 1 10 4 for the Weather dataset and 5 10 4 for all other datasets, with a warmup phase of 5 epochs. We train for 50 epochs on the ETT and Solar datasets, and 100 epochs on all others. The batch size is set to 32 for the Solar, Electricity, and Traffic datasets, and 128 for the remaining datasets. We choose the patch size P {12, 16, 24, 48} and the embedding dimension D {256, 512}. The number of latents is set to M {8, 16, 32}, and the latent dimension to DL {64, 128, 256}. For simplicity, we denote all latent dimensions as D in section 3.3, although in practice we use different values for different components. For the attention modules, we choose the number of heads nheads {4, 8} and feedforward dimension dff in each attention module is set to twice the query dimension. Lastly, the separate ratio is selected from {0, 0.5, 1}, corresponding to the fully disjoint, mixed, and fully contiguous sampling settings, respectively, as illustrated in Figure 3. All results are reported as the average over 5 runs with different random seeds. |