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..
Lyapunov-Stable Adaptive Control for Multimodal Concept Drift
Authors: Tianyu Pan, Mengdi Zhu, Alexa Cole, Ronald Wilson, Damon Woodard
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To empirically validate LS-OGD s theoretical contributions and its capacity to address key challenges in multimodal concept drift, we conducted experiments on the M3A [35] classification dataset. After initial training in a stable environment (Phase 1), the multimodal system was subjected to significant concept drift (Phase 2). Full experimental setup 1 and additional results are detailed in Appendix E. |
| Researcher Affiliation | Academia | Tianyu Bell Pan, Mengdi Zhu, Alexa Cole, Ronald Wilson, Damon Woodard Department of Electrical and Computer Engineering Florida Institute of National Security Applied Artificial Intelligence Group University of Florida Gainesville, FL 32611 EMAIL EMAIL |
| Pseudocode | Yes | Proposed Algorithm: Algorithms 1 and 2 in Appendix B outline the main online learning process and the controller logic. Figure 1 also presents the proposed system architecture. |
| Open Source Code | Yes | 1The code is available at https://github.com/Bellleuang1022/Lyapunov-Stable-Adaptive-Control-for Multimodal-Concept-Drift.git. |
| Open Datasets | Yes | To empirically validate LS-OGD s theoretical contributions and its capacity to address key challenges in multimodal concept drift, we conducted experiments on the M3A [35] classification dataset. We utilize the M3A dataset, which contains multimodal samples (text and images) relevant to misinformation detection. |
| Dataset Splits | Yes | Phase 1 (Stable Environment): Data samples are chronologically ordered based on the year column. Samples from years up to and including 2014 are used for initial model training. This phase represents a period of relatively stable data distribution. Phase 2 (Drifting Environment): Samples from years after 2014 are used to simulate the operational deployment phase where concept drift is introduced. |
| Hardware Specification | Yes | Our experiments used 4 NVIDIA A100 GPUs with 110 GB of CPU memory per job, which averaged 20 hours of execution. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with version numbers, such as Python or deep learning framework versions (e.g., PyTorch, TensorFlow). |
| Experiment Setup | Yes | The multimodal model (both for what will become the static baseline and the initial state of LS-OGD) is trained on the stable dataset partition for 1000 steps. The model continues to train on the drifted data from the Phase 2 partition for 3000 steps... In addition, models are trained using the Adam W optimizer with an initial learning rate of 5e-4 and a weight decay of 0.001. The learning rate is bounded between 1e-7 and 1e-2. |