Research map / Reinforcement learning / Offline RL
Beyond the Proxy: Trajectory-Distilled Guidance for Offline GFlowNet Training
Ruishuo Chen, Xun Wang, Rui Hu, Zhuoran Li, Longbo Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TD-GFN instead of QM-COFlowNet | GFlowNet offline training · Molecule Design | Reward-10: 7.611 -> 7.733 | Improved |
| use TD-GFN instead of QM-COFlowNet | GFlowNet offline training · Molecule Design | Convergence (trajectories): 4.423 x 10^4 -> 2.749 x 10^4 | Improved |
| use TD-GFN instead of Proxy-GFN | GFlowNet offline training · Molecule Design | Reward-1000: 6.636 -> 6.810 | Improved |
| use TD-GFN instead of Oracle-GFN training | GFlowNet · Molecule Design | Convergence (trajectories): 44.141 x 10^4 -> 2.749 x 10^4 | Improved |
| use TD-GFN instead of Dataset-GFN | offline GFlowNet · Molecule Design | Reward-100: 7.198 -> 7.450 | Improved |
| use TD-GFN instead of Behavior-GFN policy | Molecule Design data · Molecule Design | Reward-1000: 6.504 -> 6.810 | Improved |
| use BraVE offline RL instead of TD-GFN | offline method · Molecule Design | Convergence (trajectories): 2.749 x 10^4 -> 0.645 x 10^4 | Mixed |
| use TD-GFN instead of strongest baselines for mode discovery | offline GFlowNet · Molecule Design (high-reward modes) | number of modes: 1.5 to 2 times more modes | Improved |
| use TD-GFN instead of baselines | offline GFlowNet · 8^4 Hypergrid | state visits to find all 16 modes: fewer than 5,000 state visits; six-fold improvement | Improved |
| use TD-GFN instead of COFlowNet | offline GFlowNet · 8^4 Hypergrid | speed to lowest Empirical L1 Error: over twice as fast | Improved |
| train GAIL on the rebalanced dataset (ablation) | GAIL · 8^4 Hypergrid | performance | Improved |
| use rebalanced GAIL instead of Dataset-GFN | Biosequence AMP design · AMP design | reward and diversity of top 100 | Improved |
Closest papers by meaning
- Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets (ICLR 2025 · similarity 0.92)
- Pre-Training and Fine-Tuning Generative Flow Networks (ICLR 2024 · similarity 0.91)
- Optimizing Backward Policies in GFlowNets via Trajectory Likelihood Maximization (ICLR 2025 · similarity 0.90)
- GFlowNet Training by Policy Gradients (ICML 2024 · similarity 0.90)
- Pessimistic Backward Policy for GFlowNets (NeurIPS 2024 · similarity 0.90)
- Order-Preserving GFlowNets (ICLR 2024 · similarity 0.89)
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.
