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..
Adjoint Schrödinger Bridge Sampler
Authors: Guan-Horng Liu, Jaemoo Choi, Yongxin Chen, Benjamin K Miller, Ricky T. Q. Chen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we demonstrate the effectiveness of ASBS on sampling from classical energy functions, amortized conformer generation, and molecular Boltzmann distributions. Codes are available at https://github.com/facebookresearch/adjoint_samplers. [...] In summary, we present the following contributions: [...] We show ASBS s superior performance over prior methods on sampling Boltzmann distributions of classical energy functions, alanine dipeptide molecule and amortized conformer generation. [...] 6 Experiments Benchmarks We evaluate our ASBS on three classes of multi-particle energy functions E(x). [...] Table 2: Results on the synthetic energy functions for n-particle bodies with their corresponding dimensions d. [...] Figure 3: The energy histograms of DW-4 and LJ-13 from Table 2. ASBS generates samples whose energy profiles closely match those of the ground-truth samples. [...] Table 3: Comparison between diffusion samplers on sampling the molecular Boltzmann distribution of the alanine dipeptide. [...] Table 4: Results on large-scale amortized conformer generation, evaluated on two test sets, SPICE and GEOM-DRUGS... |
| Researcher Affiliation | Collaboration | Guan-Horng Liu1, , Jaemoo Choi2, , Yongxin Chen2, Benjamin Kurt Miller1, Ricky T. Q. Chen1, 1FAIR at Meta, 2Georgia Institute of Technology, Core contributors |
| Pseudocode | Yes | Algorithm 1 Adjoint Schrödinger Bridge Sampler (ASBS) [...] Algorithm 2 Adjoint Schrödinger Bridge Sampler (ASBS) Require: Sample-able source X0 µ, differentiable energy E(x), parametrized drift uθ(t, x) and corrector hϕ(x), replay buffers Badj and Bcrt, number of stages K, numbers of AM and CM epochs Madj and Mcrt, number of resamples N, number of gradient steps L, time scaling λt, maximum energy gradient norm αmax. |
| Open Source Code | Yes | Codes are available at https://github.com/facebookresearch/adjoint_samplers. |
| Open Datasets | Yes | The training set Gtrain contains 24,477 molecular topologies from SPICE (Eastman et al., 2023), represented by the SMILES strings (Weininger, 1988), whereas the test set Gtest contains 80 topologies from SPICE and another 80 from GEOM-DRUGS (Axelrod and Gomez-Bombarelli, 2022). As with (Havens et al., 2025), we consider E(x|g) a foundation model e SEN from (Fu et al., 2025), which predicts energy with density-functional-theory accuracy at a much lower computational cost. We use CREST conformers (Pracht et al., 2024) as the ground-truth samples. [...] For the ground-truth samples, we sample analytically from MW-5 and use the MCMC samples from (Klein et al., 2023) for the rest of three potentials. [...] we use the energy function E(x) from the Open MM library (Eastman et al., 2017) and consider a more structural internal coordinate with the dimension d = 60. The ground-truth samples contain 107 configurations, simulated from Molecular Dynamics (Midgley et al., 2023). |
| Dataset Splits | Yes | The training set Gtrain contains 24,477 molecular topologies from SPICE (Eastman et al., 2023), represented by the SMILES strings (Weininger, 1988), whereas the test set Gtest contains 80 topologies from SPICE and another 80 from GEOM-DRUGS (Axelrod and Gomez-Bombarelli, 2022). |
| Hardware Specification | No | The main text of the paper does not provide specific hardware details used for running experiments. The NeurIPS checklist states that these details are provided in the supplementary material, but they are not in the main paper body. |
| Software Dependencies | No | The paper mentions software like PyTorch (Paszke et al., 2019), JAX (Bradbury et al., 2018), and Open MM library (Eastman et al., 2017), but does not provide specific version numbers for these software components. The prompt requires specific version numbers for a reproducible description. |
| Experiment Setup | Yes | All models are trained with Adam (Kingma and Ba, 2015) and, following standard practices (Havens et al., 2025; Akhound-Sadegh et al., 2024), utilize replay buffers; see Appendix C for details. [...] Table 5: Hyperparameters of ASBS for the each task. MW-5 DW-4 LJ-13 LJ-55 Alanine dipeptide Conformer generation µ N(0, 1) µharmonic in (19) with α=2, 2, 1 N(0, 0.25) µharmonic βmin 0.001 0.001 0.001 0.001 0.001 βmax 1 1 2 0.5 1 σ 0.2 K 5 20 15 15 15 3 Madj 100 200 300 300 4000 2500 Mcrt 20 20 20 20 2000 2000 N 1000 1000 1000 1000 1000 128 L 200 100 100 100 100 100 |B| 104 104 104 104 104 6.4 104 αmax 100 100 100 100 150 λt 1 σ2 t 1 σ2 t 1 σ2 t 1 σ2 t 1 σ2 t 1 σ2 t |