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 Sample-Efficient Transfer Learning Conditional Diffusion Models via Representation Learning
Authors: Ziheng Cheng, Tianyu Xie, Shiyue Zhang, Cheng Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments are also conducted to verify our results. We also conduct numerical experiments in Section 5 to verify our results. |
| Researcher Affiliation | Academia | 1 Department of Industrial Engineering and Operations Research, University of California, Berkeley 2 School of Mathematical Sciences, Peking University 3 Center for Statistical Science, Peking University |
| Pseudocode | No | The paper describes theoretical models and applies them to amortized variational inference and behavior cloning in text, but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures are present. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We will provide complete codes upon acceptance. |
| Open Datasets | Yes | Image Restoration For a real data experiment, we consider the image restoration task on MNIST. |
| Dataset Splits | Yes | We have n = 1000 pre-training samples from each source distribution Pk, m {10, 20, 30, 40, 50, 100} fine-tuning samples from the target distribution P0. We run Langevin Monte Carlo for sufficiently long time to obtain 100 test samples from the target distribution P0 for evaluating the test error of different models. We use the full MNIST 1-9 data for pre-training which corresponds to n = 5000. For the fine-tuning phase, we consider m = 10, 20, 30, 40, 50, 100 training samples and 100 test samples from P0(x, y). |
| Hardware Specification | No | The authors are grateful for the computational resources provided by the High-performance Computing Platform of Peking University. |
| Software Dependencies | No | The paper does not provide specific software names with version numbers for reproducibility. |
| Experiment Setup | Yes | In the pre-training phase, the { bf k; 1 k K} and ˆh are trained on the K = 10 source distributions with 400K iterations and a batch size of 512. In the fine-tuning phase, the pre-trained representation map bh is fixed, and the bf 0 is trained on the target distribution with 200K iterations and a batch size of m. In the pre-training phase, the the { bf k; 1 k K = 9} and ˆh are 2K epochs and a batch size of 512. The initial learning rate is 0.0003 and is annealed according to a cosine annealing schedule. In the fine-tuning phase, the pre-trained representation map bh is fixed, and the bf 0 is trained on the target distribution with 20K iterations and a batch size of m. |