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..
Provable Meta-Learning with Low-Rank Adaptations
Authors: Jacob Block, Sundararajan Srinivasan, Liam Collins, Aryan Mokhtari, Sanjay Shakkottai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We verify these theoretical insights through experiments on synthetic data as well as real-data vision and language tasks. We observe significant performance benefits using a simple implementation of our proposed meta-learning scheme during retraining relative to the conventional approach. |
| Researcher Affiliation | Collaboration | Jacob L. Block UT Austin EMAIL, Sundararajan Srinivasan UT Austin EMAIL, Liam Collins Snap, Inc. EMAIL, Aryan Mokhtari UT Austin & Google Research EMAIL, Sanjay Shakkottai UT Austin EMAIL |
| Pseudocode | Yes | G Example Pseudocode for Minimizing (5) Algorithm 1 Meta-Adapter Training |
| Open Source Code | Yes | Justification: We disclose all training details in the main body and in Appendices D and E. We further release all code in the supplementary material. |
| Open Datasets | Yes | We use CIFAR-10 [21], and define T = 4 binary retraining tasks... We use the Conv AI2 [23] dataset for the language tasks. Appendix E.3 Asset Information: We use the CIFAR-10 [21] and Conv AI2 [23] datasets in our experiments. CIFAR-10 is publicly available but does not specify an explicit license. Conv AI2 is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. |
| Dataset Splits | Yes | Vision Experiments. We use CIFAR-10 [21], and define T = 4 binary retraining tasks involving classification between consecutive class labels. Specifically, task 1 classifies between classes 1 and 2, task 2 between classes 3 and 4, etc. The test task is binary classification between classes 9 and 10. Language Experiments. We retrain with T = 10 tasks and then fine-tune the model from the best-performing epoch to each of the 10 test tasks. For each retraining method, we take the model from the epoch with the best average performance on the validation samples for the retraining tasks to then be fine-tuned. |
| Hardware Specification | Yes | The data parameters for both synthetic experiments are summarized in Table 9. Both were performed using a single NVIDIA A40 GPU. All of the real data experiments were performed on a single NVIDIA H200 GPU. |
| Software Dependencies | No | The paper mentions "Adam W optimizer", "RoBERTa-base model", "RoBERTa tokenizer", "MLP-Mixer architecture" but does not specify version numbers for general software libraries or tools used in the experimental setup. |
| Experiment Setup | Yes | D Synthetic Experiments: We set the number of samples per retraining task for each task t as nt = N, so each retraining task is equipped with N samples. Further, we denote the number of test task samples n T +1 = n. We set L to be the mean squared error loss, and we run gradient descent on the standard retraining (1) and the Lo RA-ML (5) objectives. For the Lo RA-ML objective, we alternate between the outer step, updating the shared parameter A, and the inner steps, where we update the task specific parameter Ut for each task independently. We run 3000 outer steps, where after each outer step we run 10 inner steps. We use a learning rate of 3 10 2 for the outer steps and 3 10 3 for the inner steps. For the standard retraining objective, we simply run gradient descent using the learning rate 3 10 2 over 3000 epochs. Tables 11, 12, 15, and 16 detail hyperparameters for retraining and fine-tuning, including Learning Rate, Weight Decay, Epochs, Batch Size, Optimizer, and Lo RA Rank. |