Research map / Large language models / Instruction and fine-tuning
Rooted Absorbed Prefix Trajectory Balance with Submodular Replay for GFlowNet Training
Xi Wang, Wenbo Lu, Shenji Wan
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 |
|---|---|---|---|
| add suffix reward absorption to RapTB (ablation) | RapTB · SMILES | Score: 0.716 -> 0.740 | Improved |
| add suffix reward absorption to RapTB (ablation) | RapTB · SMILES | FPDiv: 0.805 -> 0.860 | Improved |
| use only max or only soft backups instead of mixed backup (ablation) | RapTB · SMILES | Score and FPDiv: Score 0.821 / 0.819 vs 0.740; FPDiv 0.775 / 0.748 vs 0.860 | Mixed |
| detach termination gradients in the auxiliary branch (ablation) | RapTB · SMILES | Score: 0.714 -> 0.740 | Improved |
| add submodular replay (SubM) to RapTB | RapTB · SMILES | Score: 0.740 -> 0.844 | Improved |
| add SubM to TB | TB · Expr24 (replay) | NormCov: 0.001 -> 0.100 | Improved |
| add SubM to RapTB | RapTB · Expr24 | NormCov: 0.039 -> 0.209 | Improved |
| use reward-only, diversity-only or length-only SubM instead of the full combination (ablation) | RapTB+SubM · SMILES | Score, FPDiv, long-horizon coverage: Score 0.878 (reward-only) vs 0.844 (full); FPDiv 0.942 (diversity-only) vs 0.898 | Mixed |
| replace TB with RapTB | LLM-GFlowNet · SMILES Lmax=15 | fraction of length 11+: 0.013 -> 0.568 | Improved |
| use SubTB instead of TB | LLM-GFlowNet · SMILES | validity (Acc): 0.998 -> 0.328 | Hurt |
| replace TB with RapTB under oracle replay | Expr24 · Expr24 Oracle | Acc: 0.919 -> 0.945 | Improved |
| restrict SubTB windows to rooted and add learnable log Z (ablation) | SubTB · Expr24 RP | Acc: 0.229 -> 0.999 | Improved |
Closest papers by meaning
- Pessimistic Backward Policy for GFlowNets (NeurIPS 2024 · similarity 0.87)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · similarity 0.87)
- Pre-Training and Fine-Tuning Generative Flow Networks (ICLR 2024 · similarity 0.87)
- Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets (ICLR 2025 · similarity 0.87)
- Stable GFlowNets with Probabilistic Guarantees (NeurIPS 2026 · similarity 0.86)
- gfnx: Fast and Scalable Library for Generative Flow Networks in JAX (NeurIPS 2026 · similarity 0.86)
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.
