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..

RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis

Authors: Robin Yadav, Qi Yan, Guy Wolf, Joey Bose, Renjie Liao

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we demonstrate RSF achieves 60.0% top-1 accuracy, which outperforms the previous SOTA by 20%. We also substantiate the benefits of steering at inference and demonstrate that FK-steering improves top-5 round-trip accuracy by 19% over prior templatefree SOTA methods, all while preserving competitive top-k accuracy results.
Researcher Affiliation Academia 1UBC; 2Vector Institute; 3Mila; 4Université de Montréal; 5Canada CIFAR AI Chair; 6University of Oxford
Pseudocode No The paper describes the methodology using mathematical formulations and descriptive text, but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks or figures.
Open Source Code Yes Code is available at: https://github.com/DSL-Lab/ Retro Syn Flow.
Open Datasets Yes Dataset. We trained and evaluated our methods on the USPTO-50K dataset [34], a standard benchmark for retrosynthesis modelling containing 50k atom-mapped reactions extracted from US patents.
Dataset Splits Yes We follow the same train/evaluation/test split used by Retro Bridge [18] and GLN [9].
Hardware Specification Yes Our training runs are done on either an NVIDIA RTX 3090 (24 GB of memory) or V100 (32 GB of memory). We benchmark on an RTX 3090 with 24 GB of memory.
Software Dependencies No Our methods are implemented in Py Torch [30], and we also use an open-source software RDKit [23], for operations involving chemical reactions and molecular graphs. However, specific version numbers for these software dependencies are not provided in the paper.
Experiment Setup Yes Our methods discretize the time interval [0, 1] into T = 50 steps. For SMC resampling, RPF-RS uses K = 4 particles. For RSF, we use M = 2 synthon predictions with N1 = 70 and N2 = 30 for the top-2 synthon predictions respectively. We train all of our models up to 600 epochs which can take up to 32 hours. The models are trained using a batch size of 32. We use Adam W [26] with a learning rate of 0.0002.