Research map / Probabilistic and scientific modeling / Variational and generative sampling

Flow Factorization for Efficient Generative Flow Networks

AAAI 2025 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Expressive invertible architectures · 0 citations (OpenAlex)

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 Bifurcated GFlowNets (BN), which factorizes edge flows into a state flow stream and an edge-based allocation stream, trained with a log-scale flow-matching-style loss, to improve GFlowNet sample and learning efficiency.

Problem

GFlowNets suffer from data inefficiency when edge flows are parameterized directly (Flow Matching) or when backward policies are needed (DB, TB, SubTB), especially in high-dimensional action spaces.

Approach

Decompose the edge flow as F(s→s') = F(s) · A(s'|s), with a state-only flow stream and a normalized edge-based allocation stream sharing a state encoder, avoiding negative flows from a naive dueling/advantage decomposition and removing the need for a backward policy.

Key results

• On HyperGrid, BN significantly improves learning efficiency over FM and outperforms DB and TB by a large margin, performing comparably to SubTB and surpassing it at larger action spaces.
• On molecule generation, BN achieves the highest top-K rewards and discovers many more modes than FM; it finds 2345.67 ± 92.07 new molecules with score > 8 versus 1447.00 ± 267.55 for FM.
• On RNA generation, BN reaches higher accuracy and discovers more modes more efficiently than baselines across reward setups.

Measured effects reported in the paper

Surprising findings

• Backward-policy-based methods (DB, TB, SubTB) fail to scale to large molecule generation, performing worse than FM there.

Limitations

• BN involves summation over parent states, adding some overhead.
• Scaling to larger state spaces is left as future work.

Experimental setup

Datasets: HyperGrid RNA binding affinity tasks (Lorenz et al. 2011) sEH molecule proxy model (Bengio et al. 2021)

Benchmarks: HyperGrid RNA sequence generation molecule generation (sEH)

Metrics: empirical L1 error number of modes discovered accuracy (relative error of mean reward) top-K reward number of new molecules with score > 8 average pairwise Tanimoto similarity

Baselines: Flow Matching Detailed Balance Trajectory Balance Sub-Trajectory Balance

Builds on: Flow Matching Detailed Balance Trajectory Balance Sub-Trajectory Balance dueling network architecture

Closest papers by meaning

• Pre-Training and Fine-Tuning Generative Flow Networks (ICLR 2024 · similarity 0.92)
• GFlowNet Training by Policy Gradients (ICML 2024 · similarity 0.92)
• Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks (ICLR 2025 · similarity 0.92)
• Expected flow networks in stochastic environments and two-player zero-sum games (ICLR 2024 · similarity 0.92)
• Pessimistic Backward Policy for GFlowNets (NeurIPS 2024 · similarity 0.92)
• Optimizing Backward Policies in GFlowNets via Trajectory Likelihood Maximization (ICLR 2025 · similarity 0.91)