Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning Energy Decompositions for Partial Inference in 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 LED-GFN, which learns potential functions on state transitions that decompose the terminal energy (regularized by dropout-based least-squares) to replace intermediate energy evaluations in forward-looking GFlowNets.
Problem
GFlowNet partial inference (forward-looking GFlowNet) requires evaluating intermediate-state energies, which can be expensive or impossible and can be sparse or high-variance, giving misleading credit assignment.
Approach
Decompose the terminal energy into learnable transition potentials trained online via a dropout-regularized least-squares objective (minimizing variance along trajectories), and use these potentials to reparameterize the flow function in DB/subTB losses, so no intermediate energy evaluation is needed while the optimal policy is preserved.
Key results
- LED-GFN outperforms GFN and FL-GFN on bag generation under both DB and subTB implementations, while FL-GFN fails to improve on subTB.
- On molecule and RNA sequence generation, LED-GFN outperforms the considered baselines including PPO, GAFN and FL-GFN in top-100 score and number of modes.
- On set generation and maximum independent set, where intermediate energies are ideal local credits, LED-GFN achieves performance similar to FL-GFN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LED-GFN instead of GFN or FL-GFN | GFlowNet · Bag generation | performance | Helped |
| Use LED-GFN instead of the considered baselines | Molecule generation · Molecule generation | average score of unique top-100 molecules | Helped |
| Use LED-GFN instead of the considered baselines | Molecule generation · Molecule generation | number of modes | Helped |
| Use LED-GFN instead of the considered baselines | RNA sequence generation · RNA sequence generation | performance | Helped |
| Use LED-GFN instead of FL-GFN | GFlowNet · Set generation and maximum independent set | performance | No clear effect |
| Use FL-GFN instead of GFN | GFlowNet · RNA sequence generation | performance | No clear effect |
| Add variance regularization to the learned potentials (ablation) | LED-GFN · RNA sequence generation | performance | Helped |
| Use least-squares energy decomposition instead of alternative potential-learning methods (ablation) | LED-GFN · Molecule generation and set generation | performance | Helped |
| Use LED-GFN instead of the considered baselines | GFlowNet · Bag generation | negative relative mean error | Helped |
| Use LED-GFN's learned potentials | Molecule generation · Molecule generation | Tanimoto similarity | Helped |
Surprising findings
- FL-GFN with subTB fails to improve over baselines on bag generation because most intermediate states give uninformative energy signals.
Limitations
- Relies on approximate potentials; exact optimal-policy preservation requires a correction term (LED-GFN*) or the energy decomposition holding for all trajectories.
- Evaluated on relatively small synthetic or proxy-oracle environments (bag, set, MIS) plus molecule and RNA tasks.
- Ideal-credit comparison tasks are idealized and only match FL-GFN rather than surpass it.
Experimental setup
Datasets: bag generation environment set generation environment molecule generation (soluble epoxide hydrolase binding energy oracle) RNA sequence generation (transcription factor binding) maximum independent set graphs
Benchmarks: bag generation set generation molecule generation RNA sequence generation maximum independent set
Metrics: number of modes average top-100 score negative relative mean error Tanimoto similarity diversity
Baselines: GFN (DB) GFN (TB) GFN (subTB) FL-GFN GAFN PPO
Builds on: forward-looking GFlowNet detailed balance sub-trajectory balance randomized return decomposition
Closest papers by meaning
- Revisiting Non-Acyclic GFlowNets in Discrete Environments (ICML 2025 · similarity 0.89)
- Pre-Training and Fine-Tuning Generative Flow Networks (ICLR 2024 · similarity 0.89)
- Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · similarity 0.89)
- On Divergence Measures for Training GFlowNets (NeurIPS 2024 · similarity 0.89)
- Pessimistic Backward Policy for GFlowNets (NeurIPS 2024 · similarity 0.89)
- Symmetry-Aware GFlowNets (ICML 2025 · similarity 0.88)
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.