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..
Greed is Good: A Unifying Perspective on Guided Generation
Authors: Zander Blasingame, Chen Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Lastly, we conduct some numerical experiments on inverse image problems and molecule generation. 7 Experiments Motivated by the theoretical connections from the previous sections we apply the greedy posterior strategy (Euler) to several problems using flow/diffusion models, as well as several methods lying in the in between space of end-to-end guidance and posterior guidance, namely, a single-step midpoint scheme and 2-step Euler scheme. 7.1 Inverse problems for images ...Quantitatively, we present a snapshot of our full results (cf. Table 6) on the inpainting with random mask and Gaussian deblurring tasks. For reference we include the standard DPS (Chung, J. Kim, et al. 2023) and the recent state-of-the-art DAPS (B. Zhang et al. 2025). We observe that the posterior guidance strategy works well performing closer to DAPS than DPS. 7.2 Molecule generation for QM9 ...We report quantitative results in Table 3. |
| Researcher Affiliation | Academia | Zander W. Blasingame Clarkson University EMAIL Chen Liu Clarkson University EMAIL |
| Pseudocode | Yes | G.2 A Py Torch pseudocode illustration Below in Code G.1 we provide an example Py Torch implementation of the greedy guidance strategy with an Euler scheme for sampling the main ODE. Code G.1: Example implementation of greedy guidance |
| Open Source Code | Yes | 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: [Yes] Justification: The source code will be uploaded as a part of the supplementary material. |
| Open Datasets | Yes | 7.1 Inverse problems for images ...on a subset of 100 images from the FFHQ (Karras, Laine, and Aila 2019) 256 x 256 dataset. 7.2 Molecule generation for QM9 ...on the QM9 dataset (Ruddigkeit et al. 2012), a popular molecular dataset containing small molecules with up to 29 atoms. |
| Dataset Splits | Yes | 7.1 Inverse problems for images ...on a subset of 100 images from the FFHQ (Karras, Laine, and Aila 2019) 256 x 256 dataset. I.1 Inverse image problems ...We conduct experiments on the following linear tasks: super resolution, Gaussian deblurring, motion deblurring, inpaintining (with a box mask), and inpainting (with a 70% random mask)... We report the mean performance (PSNR, SSIM, and LPIPS) across 100 validation images along with the FID. I.2 Molecule generation for QM9 ...An equivariant graph neural network (GNN) was trained for each property on half of the QM9 dataset, serving as a classifier this model was then used as a guidance function during the experiments. The Equi FM (Y. Song, Gong, et al. 2023) model was trained on the whole QM9 training set and was used as the underlying flow model for the experiments. Following L. Wang et al. (2025), the test time properties were sampled from the whole training set |
| Hardware Specification | Yes | I.4 Hardware and compute cost Inverse image problems. The inverse image problem experiments were run on a single NVIDIA H100 80GB GPU. Molecule generation. The molecule generation experiments were run on a single NVIDIA V100 16GB GPU. |
| Software Dependencies | No | Code G.1: Example implementation of greedy guidance # assumed dt and loss are defined ...optimizer = torch.optim.SGD([xt_opt], lr=lr)... ...optimizer = torch.optim.LBFGS([xt_opt], max_iter=max_iter, lr=lr, line_search_fn='strong_wolfe') I.1 Inverse image problems ...with the stock implementation of the torch.optim.SGD method for each step of the ODE solve. I.2 Molecule generation for QM9 ...we used the L-BFGS algorithm (D. C. Liu and Nocedal 1989) with 5 optimizer steps and 5 inner steps with a linear search, in particular we used the stock Py Torch implementation torch.opt.LBFGS. |
| Experiment Setup | Yes | I.1 Inverse image problems Diffusion model. ...We employ a midpoint scheme to solve this ODE in reverse-time with N = 20 steps. Hyperparameters. ...The learning rate was set at η = 1 for all experiments, and we performed nopt = 50 optimization steps with the stock implementation of the torch.optim.SGD method for each step of the ODE solve. We set βy = 0.05 for all tasks. Ablation study. For the ablation study in Table 2 we used the L-BGFS optimizer over the standard SGD optimizer used in the main experiments (for the greedy guidance runs). ...Importantly, we fix the maximum number of optimization steps between the greedy and DTO strategies; for greedy we take 5 optimization step per step in the ODE solver, so a 100 in total. Likewise, for DTO we take a 100 optimization steps in total. I.2 Molecule generation for QM9 ...We solve the ODE with Euler s method and 50 discretization steps; further configuration details are reported in Appendix I.1. ...used the L-BFGS algorithm (D. C. Liu and Nocedal 1989) with 5 optimizer steps and 5 inner steps with a linear search... For the DTO experiment we used a learning rate of η = 1. ...We found that a learning rate of η = 0.001 seemed to work better. ...We chose the following simple scheduler: ηt = η(1 t) t > 0.5 0 t 0.5 , (117) where η = 0.001 is the base learning rate. |