GFlowState Visualization Workflow Diagram
Figure 2 : Workflow of GFlowState . Arrows show the component dependencies during and after training. All visualizations use the training samples as their data source, filtered by the user-selected iteration range. The DAG View and the Transition Heatmap use the graph computed from the trajectories of the samples. The State Projection also uses a validation set, if available, and projects and aggregates the data into hexbins. Selecting samples or trajectories also highlights them or shows their details in other visualizations.
Paper context
Paper title: GFlowState: Visualizing the Training of Generative Flow Networks Beyond the Reward Abstract: We present GFlowState, a visual analytics system designed to illuminate the training process of Generative Flow Networks (GFlowNets or GFNs). GFlowNets are a probabilistic framework for generating samples proportionally to a reward function. While GFlowNets have proved to be powerful tools in applications such as molecule and material discovery, their training dynamics remain difficult to interpret. Standard machine learning tools allow metric tracking but do not reveal how models explore the sample space, construct sample trajectories, or shift sampling probabilities during training. Our solution, GFlowState, allows users to analyze sampling trajectories, compare the sample space relative to reference datasets, and analyze the training dynamics. To this end, we introduce multiple views, including a chart of candidate rankings, a state projection, a node-link diagram of the trajectory network, and a transition heatmap. These visualizations enable GFlowNet developers and users to investigate sampling behavior and policy evolution, and to identify underexplored regions and sources of training failure. Case studies demonstrate how the system supports debugging and assessing the quality of GFlowNets across application domains. By making the structural dynamics of GFlowNets observable, our work enhances their interpretability and can accelerate GFlowNet development in practice. Passages referencing this figure: ; a x , a y ∈ [ 0 , 1 ] a=(a_{x},a_{y});a_{x},a_{y}\in[0,1] . The current state transitions into the next state according to the sampled action: s i + 1 = s i + a i s_{i+1}=s_{i}+a_{i} . The agent can also sample a stop action to terminate the sampling trajectory. In this example setup, a fixed reward distribution is defined to place high-reward modes near the corners of the grid (see Fig. ˜ 1 ). Figure 1 : A heatmap of the true rewards for the grid environment used as an example, along with an exemplary trajectory. High-reward modes are placed near the corners of the grid. The white trajectory is one possible way to sample the state (14,4). The final object is the cell on the grid at which the agent samples the stop action; all previous actions that together lead to this point define the
