Research map / Image generation / Fast diffusion sampling
Guided Trajectory Generation with Diffusion Models for Offline Model-based Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GTG, a conditional diffusion model trained on locality-biased synthetic trajectories from an offline dataset, which samples score-conditioned trajectories with guidance and context inpainting and filters candidates with a proxy for offline model-based optimization.
Problem
Offline model-based optimization: finding designs that maximize a black-box function using only a fixed offline dataset, where forward proxies suffer OOD errors and inverse generative models struggle with multi-modal mappings.
Approach
Construct synthetic trajectories toward high-scoring regions using K-nearest-neighbor local search with score perturbation, train a classifier-free-guided diffusion model to generate whole trajectories conditioned on their summed scores, then sample with context inpainting and filter candidates with a proxy.
Key results
- GTG achieves the best mean rank of 1.6/15 across the five Design-Bench tasks, best on Superconductor and D'Kitty.
- On sparse D'Kitty (1%), GTG reaches 0.949 vs 0.865 for BDI and 0.875 for BONET.
- Toy Branin: GTG reaches -0.490 ± 0.070 versus dataset max -6.031 and global optimum -0.398.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GTG instead of the strongest reported baseline | Design-Bench methods · Superconductor | max normalized score: 0.510 ± 0.015 -> 0.519 ± 0.045 | Helped |
| use GTG instead of BONET | Branin optimization · Branin | maximum function value: -0.769 ± 0.227 -> -0.490 ± 0.070 | Helped |
| construct trajectories with locality bias rather than SORT-SAMPLE (ablation) | GTG · TFBind8 | max normalized score: 0.954 ± 0.026 -> 0.976 ± 0.020 | Helped |
| construct trajectories with locality bias rather than Top-p Percentile (ablation) | GTG · Superconductor | max normalized score: 0.439 ± 0.039 -> 0.519 ± 0.045 | Helped |
| add classifier-free guidance, context conditioning, and filtering (ablation) | GTG sampling · TFBind8 | max normalized score: 0.923 ± 0.054 -> 0.976 ± 0.020 | Helped |
| add filtering (ablation) | GTG · TFBind8 | performance: 0.920 ± 0.036 -> 0.976 ± 0.020 | Helped |
| add filtering (ablation) | GTG · Ant | diversity: 17.02 ± 3.56 -> 9.41 ± 1.96 | Hurt |
| replace classifier-free guidance with score inpainting (ablation) | GTG sampling · TFBind8 | max normalized score: 0.976 ± 0.020 -> 0.963 ± 0.026 | Hurt |
| pretrain the diffusion model with a synthetic dataset (ablation) | GTG · Ant | max normalized score: 0.655 ± 0.051 -> 0.897 ± 0.015 | Helped |
Surprising findings
- Inpainting score values instead of classifier-free guidance performs worse, and conditioning on extremely high exploration level alpha degrades performance.
- Filtering with the proxy improves average performance but reduces candidate diversity substantially (e.g., Ant diversity 9.41 vs 17.02 without filtering).
Limitations
- Relies on a proxy function for filtering, which may give inaccurate predictions.
- Assumes the optimal value y* is known for conditioning (relaxation via gamma introduces a task-dependent hyperparameter).
- Discrete tasks are handled via continuous relaxation of one-hot inputs rather than discrete diffusion.
Experimental setup
Datasets: Design-Bench TFBind8 Design-Bench TFBind10 Design-Bench Superconductor Design-Bench Ant Design-Bench D'Kitty Branin (synthetic 2D)
Benchmarks: Design-Bench TFBind8 Design-Bench TFBind10 Design-Bench Superconductor Design-Bench Ant Design-Bench D'Kitty Branin sparse and noisy variants of TFBind8 and D'Kitty
Metrics: 100th percentile normalized score among Q=128 candidates mean rank diversity (average pairwise distance)
Baselines: BO-qEI CMA-ES REINFORCE Gradient Ascent COMs NEMO RoMA BDI ICT CbAS MINs DDOM
Builds on: BONET PGS Diffuser classifier-free guidance DDOM
Closest papers by meaning
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.88)
- ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · similarity 0.88)
- One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling (NeurIPS 2025 · similarity 0.87)
- Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization (ICML 2025 · similarity 0.87)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.87)
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.