Research map / Reinforcement learning / Policy gradient and actor-critic
Trajectory balance: Improved credit assignment in GFlowNets
Nikolay Malkin, Moksh Jain, Emmanuel Bengio, Chen Sun, Yoshua Bengio
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 |
|---|---|---|---|
| Train with trajectory balance (TB) instead of flow matching / detailed balance | GFlowNet on AMP generation · AMP generation | Top 100 Reward: 0.78 -> 0.85 | Improved |
| Train with TB instead of FM/DB | GFlowNet on AMP generation · AMP generation | Top 100 Diversity: 12.61 -> 18.35 | Improved |
| Use TB instead of SAC | GFlowNet on AMP generation · AMP generation | Top 100 Reward: 0.80 -> 0.85 | Improved |
| Use TB instead of MCMC | GFlowNet on AMP generation · AMP generation | Top 100 Diversity: 12.56 -> 18.35 | Improved |
| Train with TB instead of FM | GFlowNet on molecule generation · Molecule generation held-out | Pearson correlation log R vs log p: 0.31 -> 0.86 | Improved |
| Train with TB instead of DB | GFlowNet on molecule generation · Molecule generation held-out | Pearson correlation log R vs log p: 0.8 -> 0.86 | Improved |
| Train with TB instead of FM | GFlowNet on molecule generation · Molecule generation | Diversity of top 1000 samples (Tanimoto similarity) | Improved |
| Train with TB instead of FM (no parent enumeration) | GFlowNet on molecule generation · Molecule generation | runtime: up to 5x runtime speedup | Improved |
| Train with DB or TB instead of FM | GFlowNet on hypergrid · Hypergrid L1 error | convergence speed | Improved |
| Learn backward policy PB instead of fixing it uniform (ablation) | DB and TB GFlowNets · Hypergrid 64 x 64 | convergence speed | Improved |
| Train with TB instead of FM | GFlowNet on bit sequences · Bit sequences n=120, k in {1,2,4,6,8,10} | Spearman correlation with reward | Improved |
| Increase action space size (bits k) with FM training | GFlowNet FM · Bit sequences | Spearman correlation | Mixed |
Closest papers by meaning
- Revisiting Non-Acyclic GFlowNets in Discrete Environments (ICML 2025 · similarity 0.91)
- A Theory of Non-acyclic Generative Flow Networks (AAAI 2024 · similarity 0.91)
- Flow Factorization for Efficient Generative Flow Networks (AAAI 2025 · similarity 0.91)
- Stable GFlowNets with Probabilistic Guarantees (NeurIPS 2026 · similarity 0.90)
- Pessimistic Backward Policy for GFlowNets (NeurIPS 2024 · similarity 0.90)
- Optimizing Backward Policies in GFlowNets via Trajectory Likelihood Maximization (ICLR 2025 · similarity 0.90)
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.
