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..
Automatic Auxiliary Task Selection and Adaptive Weighting Boost Molecular Property Prediction
Authors: Zhiqiang Zhong, Davide Mottin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive evaluations demonstrate that AUTAUT outperforms 10 auxiliary task-based approaches and 18 advanced molecular property prediction models. We evaluate AUTAUT on 9 molecular property prediction datasets, demonstrating its superiority over 10 auxiliary task-based methods and 18 state-of-the-art property prediction models. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Aarhus University 2Institute for Advanced Studies, University of Luxembourg 3Faculty of Science, Technology and Medicine, University of Luxembourg Contacts: EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 AUTAUT: Optimizing Primary Task with Auxiliary Task Integration |
| Open Source Code | Yes | Our code and data are available at https://github.com/zhiqiangzhongddu/AUTAUT. |
| Open Datasets | Yes | This paper selects various datasets from a widely used benchmark, Molecule Net [52], to examine the effectiveness of our algorithm for molecular property prediction. |
| Dataset Splits | Yes | To ensure a rigorous evaluation, we follow prior work [6, 65] and adopt scaffold splitting to divide datasets into training, validation, and test sets with an 80%-10%-10% ratio. |
| Hardware Specification | Yes | All experiments are conducted on 8 NVIDIA A100 GPUs. |
| Software Dependencies | No | The paper mentions the use of 'Adam optimizer' and 'Reduce LROn Plateau scheduler' (with a footnote linking to PyTorch documentation), but does not specify version numbers for PyTorch, Python, or other key software components. |
| Experiment Setup | Yes | We use the Adam optimizer [22] with a weight decay of 1e 16 for all models. A Reduce LROn Plateau scheduler1 is applied with a patience of 10 epochs to dynamically adjust the learning rate. For AUTAUT, we fix the number of selected auxiliary tasks at K = 5, with a detailed ablation study on the impact of K in Section 4.4. To improve the consistency of the LLM s behavior during auxiliary task selection, we set the LLM s hyperparameter temperature to 0.2. |