Research map / Reinforcement learning / Visual and representation RL
Local Search GFlowNets
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 LS-GFN, which augments GFlowNet training with a local search (backward destroy with P_B, forward reconstruct with P_F, then accept/reject) so replay data is biased toward high-reward objects, improving accuracy and mode discovery on six biochemical tasks.
Problem
GFlowNets tend to over-explore the wide sample space and often fail to consistently produce high-reward samples, hurting training efficiency on scientific discovery tasks.
Approach
Refine GFlowNet-sampled trajectories with a local search that backtracks K steps using the backward policy and reconstructs with the forward policy, accepting improvements via deterministic or Metropolis-Hastings filtering, and add all refined trajectories to a prioritized replay buffer for training.
Key results
- TB + LS-GFN reaches 100.00 accuracy on QM9 and sEH versus 97.84 and 100.00 for TB alone, and 97.05 vs 85.63 on TFBind8.
- TB + LS-GFN discovers 5228 modes on sEH versus 706 for TB, and 32 vs 10 modes on L14-RNA1.
- On L14-RNA1 accuracy improves from 33.47 (TB) to 87.28 (TB + LS-GFN).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add local search to TB | TB GFlowNet · QM9 | accuracy: 97.84 ± 0.63 -> 100.00 ± 0.00 | Helped |
| Add local search to TB | TB GFlowNet · sEH | accuracy: 100.00 ± 0.00 -> 100.00 ± 0.00 | No clear effect |
| Add local search to TB | TB GFlowNet · TFBind8 | accuracy: 85.63 ± 0.35 -> 97.05 ± 0.58 | Helped |
| Add local search to TB | TB GFlowNet · L14-RNA1 | accuracy: 33.47 ± 0.37 -> 87.28 ± 3.25 | Helped |
| Add local search to TB | TB GFlowNet · L14-RNA2 | accuracy: 21.88 ± 0.35 -> 56.63 ± 0.56 | Helped |
| Add local search to TB | TB GFlowNet · L14-RNA3 | accuracy: 32.70 ± 0.59 -> 75.75 ± 3.10 | Helped |
| Add local search to TB | TB GFlowNet · QM9 | number of modes: 699 ± 14 -> 793 ± 4 | Helped |
| Add local search to TB | TB GFlowNet · sEH | number of modes: 706 ± 126 -> 5228 ± 141 | Helped |
| Add local search to TB | TB GFlowNet · TFBind8 | number of modes: 320 ± 3 -> 316 ± 0 | Hurt |
| Add local search to TB | TB GFlowNet · L14-RNA1 | number of modes: 10 ± 4 -> 32 ± 4 | Helped |
| Add local search to TB | TB GFlowNet · L14-RNA2 | number of modes: 6 ± 0 -> 27 ± 1 | Helped |
| Add local search to TB | TB GFlowNet · L14-RNA3 | number of modes: 6 ± 1 -> 18 ± 0 | Helped |
Surprising findings
- Local search improves mode discovery alongside reward, contrary to the common expectation that training on higher rewards makes GFlowNets less diverse.
- Deterministic filtering acceptance rate stays below 0.5 throughout training, so only a small fraction of refined trajectories are accepted.
Limitations
- Performance depends on the quality of the backward policy; low acceptance rate can limit local search benefit.
- Some tasks show no gain or slight drops (e.g., DB + LS-GFN on sEH and TFBind8).
Experimental setup
Datasets: QM9 sEH TFBind8 L14-RNA1 L14-RNA2 L14-RNA3
Benchmarks: QM9 sEH TFBind8 L14-RNA1 L14-RNA2 L14-RNA3
Metrics: accuracy (relative gap to target reward mean) number of modes discovered top-100 reward top-100 diversity uniqueness fraction
Baselines: DB MaxEnt GFN TB SubTB (0.9) SubTB Markov Molecular Sampling A2C Soft Q-Learning PPO
Builds on: GFlowNets trajectory balance subtrajectory balance prioritized replay training Markov Molecular Sampling
Closest papers by meaning
- Pre-Training and Fine-Tuning Generative Flow Networks (ICLR 2024 · similarity 0.91)
- Pessimistic Backward Policy for GFlowNets (NeurIPS 2024 · similarity 0.91)
- Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets (ICLR 2025 · similarity 0.90)
- Order-Preserving GFlowNets (ICLR 2024 · similarity 0.90)
- Flow Factorization for Efficient Generative Flow Networks (AAAI 2025 · similarity 0.89)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · 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.