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..
Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning
Authors: Xiao Han, ZIMO ZHAO, Wanyu Wang, Maolin Wang, Zitao Liu, Yi Chang, Xiangyu Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on 15 diverse datasets show that DEAL consistently outperforms baseline methods, yielding substantial gains in task accuracy and resource efficiency. |
| Researcher Affiliation | Academia | Xiao Han Zhejiang University of Technology, Zhejiang Key Laboratory of Visual Information Intelligent Processing Hangzhou, China EMAIL Zimo Zhao City University of Hong Kong Hong Kong, China EMAIL Wanyu Wang City University of Hong Kong Hong Kong, China EMAIL Maolin Wang City University of Hong Kong Hong Kong, China EMAIL Zitao Liu Jinan University Jinan, China EMAIL Yi Chang Jilin University Jilin, China EMAIL Xiangyu Zhao City University of Hong Kong Hong Kong, China EMAIL |
| Pseudocode | Yes | Algorithm 1 The DEAL framework Input: Pre-trained Lo RA-based LLM AW , W , training samples D Output: Fine-tuned model AW , W 1: W W 2: for each batch (Q, G) D do 3: Extract A, B from W 4: for each Y {A, B} do 5: ˆX Eq. (10), O Eq. (11) 6: if Y == A then 7: A O 8: else 9: B O 10: end if 11: end for 12: W A B 13: Compute loss Loss using Eq. (12) 14: Update parameters via back-propagation 15: end for 16: return AW , W |
| Open Source Code | Yes | The source code is publicly available at https://github.com/Applied-Machine-Learning-Lab/DEAL. |
| Open Datasets | Yes | Datasets. We evaluate DEAL on three continual learning (CL) benchmarks in a sequential task setup... (i) Same-domain tasks: We use a three-task benchmark consisting of AG News (news classification), DBpedia (entity typing), and Yahoo Answers (question topic prediction)... (iii) Heterogeneous multi-task learning: We evaluate on the 15-task benchmark proposed by [27], which spans text classification (AG News, DBpedia, Yahoo, Amazon, Yelp), GLUE tasks (MNLI, QQP, RTE, SST-2) [28], Super GLUE tasks (Wi C, CB, COPA, Multi RC, Bool Q) [29], and IMDB [30]. Full details on dataset preprocessing and prompt construction are provided in Appendices B.1 and E. |
| Dataset Splits | Yes | Datasets. We evaluate DEAL on three continual learning (CL) benchmarks in a sequential task setup, where data from previous tasks is unavailable during training on subsequent ones... Full details on dataset preprocessing and prompt construction are provided in Appendices B.1 and E. Table 5 summarizes all datasets used across the continual learning benchmarks. These datasets were also employed in O-Lo RA [31], where each is framed as a classification task using a unified instruction-based text-to-text format. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA A100 GPU. |
| Software Dependencies | No | We implement DEAL using the Hugging Face Transformers library and perform training with FP16 mixed precision on a single NVIDIA A100 GPU. |
| Experiment Setup | Yes | Implementation Details. We implement DEAL using the Hugging Face Transformers library and perform training with FP16 mixed precision on a single NVIDIA A100 GPU. Unless otherwise stated, we adopt a consistent experimental setup across both LLa MA-3.1 and T5-large backbones, tailored for low-resource, instruction-driven continual learning. All models are trained with adapter-based fine-tuning using Lo RA, where we set the rank r = 32 for LLa MA and r = 16 for T5, selected based on backbone capacity. Optimization is performed using Adam W with a constant learning rate scheduler. Regularization is enforced via ℓp-norm constraints on adapter weights and MLP modules, with details provided in Table 6. |