Research map / Probabilistic and scientific modeling / Variational and generative sampling
$f$-Trajectory Balance: A Loss Family for Tuning GFlowNets, Generative Models, and LLMs with Off- and On-Policy Data
Jake Fawkes, Jason Hartford
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 forward KL or Hellinger alpha-divergence loss instead of reverse KL (standard trajectory balance) (ablation) | GFlowNet on hypergrid · Synthetic 2D hypergrid | modes discovered | Improved |
| use Pearson chi-square loss (alpha=2) instead of reverse KL (ablation) | GFlowNet on hypergrid · Synthetic 2D hypergrid | modes discovered | Hurt |
| use alpha=0.75 instead of trajectory balance (ablation) | SynFlowNet · SynFlowNet molecule tasks | molecule diversity | Improved |
| use alpha=1.2 instead of trajectory balance (ablation) | SynFlowNet · SynFlowNet molecule tasks | molecule diversity | Hurt |
| anneal alpha from 0.75 to 1.2 during training (ablation) | SynFlowNet · SynFlowNet DRD2 | diversity-reward trade-off | Improved |
| use f-trajectory balance losses with annealed alpha instead of standard trajectory balance (ablation) | diffusion model tuning · MNIST odd/even conditional sampling | evenness of mode sampling | Improved |
| use tempered DevGrad f-divergence losses instead of optimised PPO under 50-step asynchrony | Qwen2.5 3B-14B and OLMo-2-7B · GSM8k and MATH mix | entropy-reward trade-off | Improved |
Closest papers by meaning
- On Divergence Measures for Training GFlowNets (NeurIPS 2024 · similarity 0.90)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · similarity 0.90)
- Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning (NeurIPS 2025 · similarity 0.87)
- Learning Energy Decompositions for Partial Inference in GFlowNets (ICLR 2024 · similarity 0.87)
- Stable GFlowNets with Probabilistic Guarantees (NeurIPS 2026 · similarity 0.87)
- Revisiting Non-Acyclic GFlowNets in Discrete Environments (ICML 2025 · 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.
