Partial GFlowNet: Accelerating Convergence in Large State Spaces via Strategic Partitioning

Generative Flow Networks (GFlowNets) have shown promising potential to generate high-scoring candidates with probability proportional to their rewards. As existing GFlowNets freely explore in state space, they encounter significant convergence challenges when scaling to large state spaces. Addressing this issue, this paper proposes to restrict the exploration of actor. A planner is introduced to partition the entire state space into overlapping partial state spaces. Given their limited size, these partial state spaces allow the actor to efficiently identify subregions with higher rewards. A heuristic strategy is introduced to switch partial regions thus preventing the actor from wasting time exploring fully explored or low-reward partial regions. By iteratively exploring these partial state spaces, the actor learns to converge towards the high-reward subregions within the entire state space. Experiments on several widely used datasets demonstrate that \modelname converges faster than existing works on large state spaces. Furthermore, \modelname not only generates candidates with higher rewards but also significantly improves their diversity.

💡 Research Summary

Generative Flow Networks (GFlowNets) have emerged as a powerful class of probabilistic generative models that sample discrete structures with probability proportional to a given reward. While they excel at discovering multiple high‑reward modes in modest‑size problems, their performance deteriorates dramatically when the underlying state space grows exponentially, because the flow‑matching or detailed‑balance constraints must be satisfied over an astronomically large graph. The paper “Partial GFlowNet: Accelerating Convergence in Large State Spaces via Strategic Partitioning” tackles this scalability bottleneck by introducing a two‑component architecture consisting of a planner and an actor. The planner dynamically partitions the full state space into overlapping partial regions of manageable size and decides when to switch the actor from one region to another. The actor, constrained to operate within the currently selected region, learns a GFlowNet policy using standard flow‑matching objectives but only over the restricted subgraph.

Key technical contributions

1. State‑space decomposition – The action set is split into state‑independent actions (A*) and state‑dependent actions (A′). A stochastic mask with probability p is applied to A*; only “valid” actions survive. This yields a partial region R defined recursively from the initial state using only valid A* actions. The expected size ratio E