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..
Learning Diffusion Models with Flexible Representation Guidance
Authors: Chenyu Wang, Cai Zhou, Sharut Gupta, Johnson Lin, Stefanie Jegelka, Stephen Bates, Tommi Jaakkola
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments across image, protein sequence, and molecule generation tasks demonstrate superior performance as well as accelerated training. In particular, on the class-conditional Image Net 256 256 benchmark, our guidance results in 23.3 times faster training than the original Si T-XL as well as four times speedup over the state-of-the-art method REPA [61]. |
| Researcher Affiliation | Academia | 1MIT 2UCLA 3TU Munich |
| Pseudocode | No | The paper does not contain clearly labeled pseudocode or algorithm blocks. It describes methodologies using mathematical equations and textual explanations. |
| Open Source Code | Yes | github.com/ChenyuWang-Monica/REED Project Page We use public data in our experiments and have released our code. |
| Open Datasets | Yes | We conduct experiments on Image Net [12] at 256 256 resolution We train an inverse folding model using the Multiflow [7] objective and the Protein MPNN [11] architecture, with the PDB training set from [11]. We adopt the challenging Geom-Drug [3] dataset |
| Dataset Splits | Yes | For protein inverse folding: We filter for single-chain proteins with sequences under 256 residues, yielding 13,753 training and 811 test proteins. For molecule generation: Following Semla Flow [25], we filter out molecules with more than 72 atoms in the training set and use standard splits. |
| Hardware Specification | Yes | Our experiments are conducted using 8 NVIDIA A100 80GB GPUs. Alpha Fold3 representation generation is performed on 8 NVIDIA A100 80GB GPUs, with the extracted embeddings stored for subsequent use. All other experiments are conducted on a single NVIDIA A100 80GB GPU. We train Semla Flow with REED for 200 epochs on a single NVIDIA A100 80GB GPU. |
| Software Dependencies | No | The paper mentions several software tools and libraries such as Adam W [39], Adam [28], Stable Diffusion VAE [46], ESMFold [36], and RDKit, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | For optimization, we use Adam W [39] with a constant learning rate of 1 10 4, parameters (β1, β2) = (0.9, 0.999), and no weight decay. The batch size is set to 256. To accelerate training, we use mixed-precision (fp16) and apply gradient clipping with a threshold of 1.0. |