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..
Pairwise Optimal Transports for Training All-to-All Flow-Based Condition Transfer Model
Authors: Kotaro Ikeda, Masanori Koyama, Jinzhe Zhang, Kohei Hayashi, Kenji Fukumizu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of this method on synthetic and benchmark datasets, as well as on chemical datasets in which continuous physical properties are defined as conditions. The code for this project can be found at https://github.com/kotatumuri-room/A2A-FM |
| Researcher Affiliation | Collaboration | Kotaro Ikeda The University of Tokyo Preferred Networks, Inc. EMAIL Masanori Koyama The University of Tokyo Preferred Networks, Inc. Jinzhe Zhang Preferred Networks, Inc. Kohei Hayashi The University of Tokyo Preferred Networks, Inc. Kenji Fukumizu The Institute of Statistical Mathematics Preferred Networks, Inc. |
| Pseudocode | Yes | Algorithm 1 Training of A2A-FM |
| Open Source Code | Yes | The code for this project can be found at https://github.com/kotatumuri-room/A2A-FM |
| Open Datasets | Yes | A2A-FM was trained on a 500K ZINC22 subset, where X represented the latent space of molecular representations [26] and C = R32 was the space of QED embeddings. Following the same protocol as previous methods [26, 19, 38], we evaluated the algorithm by the success rate of discovering a molecule of desired property (QED, similarity) within the prescribed... To demonstrate the applicability of A2A-FM to high-dimensional grouped data, we trained A2A-FM on the 256 × 256 downscaled version of Celeb A-Dialog HQ dataset [22] which contains 200K high-quality facial images with 6-level annotations of attributes: Bangs, Eyeglasses, Beard, Smiling, Age. |
| Dataset Splits | Yes | A2A-FM was trained on a 500K ZINC22 subset... For Log P&TPSA experiment: To evaluate the sampling efficiency in all-to-all condition transfer task, we conducted the nearby sampling similar to the QED experiment, except that we chose random 1,024 molecules from ZINC22 as the initial molecules, and aimed at changing their two other properties (Log P and TPSA) to a randomly selected pair of target values embedded in C = R32 2. For evaluation, we measured Tanimoto similarity s Tani and normalized condition error cerr for each transferred sample... In our experiments, we used a 3.7M subset from the ZINC22 dataset and created a validation split of initial molecules containing 1024 molecules. |
| Hardware Specification | Yes | All of the model training was done using internal NVIDIA V100 GPU cluster. |
| Software Dependencies | No | The paper does not explicitly list specific software dependencies with version numbers. It mentions the general architecture (UNet, MLP) but not the frameworks or libraries used for implementation or their versions. |
| Experiment Setup | Yes | For A2A-FM, we used β = 10, and we selected this parameter by searching β = 0.01, 0.1, 1, 10, 100 (see Appendix C for details). The batch size for all the methods were 1 × 10^3. In SI, we discretized the conditional space into 5 equally divided partitions. Also in partial diffusion, we used the classifier free guidance method with weight 0.3, used timesteps of T = 1000 and reversed the diffusion process for 300 steps... In the training procedure, we first normalized the condition space with the empirical cumulative density functions so that the empirical condition values would become uniformly distributed, and then embedded them using the Time Embedding layer to obtain its 32 dimensional representation... we used batch_size=1024, β = (batch_size)1/2dc = (1.2419)1/2, where dc = 32 is the dimension of the conditional space C. |