Q-learning is premised on learning a function to predict the value of a state-action pair, and that function then guides policy improvement. Most existing methods learn this value function from scratch using scalar targets. A new arXiv paper, "Generative Value Classifier," proposes a different framing: instead of predicting a single scalar, the value function is treated as a generative classifier.
The abstract argues that the standard scalar approach may be limiting. By framing value learning as a generative classification problem, the method could capture richer information about the distribution of returns or the structure of the value landscape. The paper is announced as new on arXiv with the identifier 2610.22497.
Because the full abstract is truncated, the article does not include experimental results or detailed methodology. The significance lies in the conceptual shift: moving from regression to generative classification for value estimation in reinforcement learning.