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..
Magical: Medical Lay Language Generation via Semantic Invariance and Layperson-tailored Adaptation
Authors: Weibin Liao, Tianlong Wang, Yinghao Zhu, Yasha Wang, Junyi Gao, Liantao Ma
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on three real-world lay language generation datasets demonstrate that Magical consistently outperforms prompt-based methods, vanilla Lo RA, and its recent variants, while also reducing trainable parameters by 31.66%. |
| Researcher Affiliation | Academia | Weibin Liao , Tianlong Wang , Yinghao Zhu , Yasha Wang , Junyi Gao , Liantao Ma National Engineering Research Center For Software Engineering, Peking University School of Computer Science, Peking University School of Computing and Data Science, The University of Hong Kong Centre for Medical Informatics, University of Edinburgh Health Data Research UK |
| Pseudocode | No | The paper describes methods and concepts but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | https://github.com/tianlwang/Magical.git |
| Open Datasets | Yes | In this empirical study, we conduct a detailed investigation of three real-world publicly available datasets used in our work: Cochrane [24], e Life [25], and Plos_genetics [2]. |
| Dataset Splits | Yes | We followed the official data splits to construct the training and test sets. Detailed statistics of these datasets are provided in Appendix. A. Table 4: Statistical information of three MLLG datasets. Datasets #Train #Test Cochrane [2] 3,979 480 e Life [25] 4,587 241 Plos_genetics [2] 4,000 300 |
| Hardware Specification | Yes | The experiments were conducted on 8 NVIDIA-H20-96GB GPUs. |
| Software Dependencies | No | The paper mentions using PyTorch and Hugging Face transformers, but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | For each experimental setup, we trained all LLMs for 5 epoch with Deep Speed Ze RO 2 Offload [48]. We utilized the Adam W optimizer and a cosine learning rate scheduler, with a warm-up ratio set to 0.1. ... Appendix C Hyperparameter Settings: Batch Size, Learning Rate, Contrastive Loss Weight, τ, K |