Symmetry-Aware GFlowNets

Generative Flow Networks (GFlowNets) offer a powerful framework for sampling graphs in proportion to their rewards. However, existing approaches suffer from systematic biases due to inaccuracies in state transition probability computations. These biases, rooted in the inherent symmetries of graphs, impact both atom-based and fragment-based generation schemes. To address this challenge, we introduce Symmetry-Aware GFlowNets (SA-GFN), a method that incorporates symmetry corrections into the learning process through reward scaling. By integrating bias correction directly into the reward structure, SA-GFN eliminates the need for explicit state transition computations. Empirical results show that SA-GFN enables unbiased sampling while enhancing diversity and consistently generating high-reward graphs that closely match the target distribution.

This paper tackles a challenge in how computers learn to build graphs—structures made of points and connections that can represent things like molecules. A popular method for this, called Generative Flow Networks (GFlowNets), is designed to explore many different graph possibilities while favoring those that are more useful or valuable. However, the current techniques often make systematic mistakes because they don't fully account for the natural symmetries in graphs—situations where two graphs look different but actually represent the same thing. To fix this, the paper introduce a new approach called Symmetry-Aware GFlowNets (SA-GFN). This method adjusts how the computer learns, so it automatically corrects for these symmetry-related mistakes without needing complicated calculations. The result is a system that explores a wider variety of graphs more accurately and reliably finds the best ones. This could be especially helpful in areas like drug discovery, where finding the right molecular structure can make a big difference.