Reproduce “Learning to Scale Logits for Temperature-Conditional GFlowNets” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minsu Kim, Joohwan Ko, Taeyoung Yun, Dinghuai Zhang, Ling Pan, Woochang Kim, Jinkyoo Park, Emmanuel Bengio and 1 more
The paper introduces Logit-scaling GFlowNets (Logit-GFN), an architectural design that improves the training stability and performance of temperature-conditional Generative Flow Networks. This method, featuring a learned function of inverse temperature to scale policy logits, significantly enhances generalization in offline learning and mode discovery in online settings across various biochemical and chemical tasks.
Key insights
- Directly scaling policy logits via a learned temperature-dependent function provides an effective inductive bias that adapts to changes in the target distribution's temperature, promoting stable training.
- Logit-GFN enables robust generalization in offline settings, allowing the model to effectively sample high-reward structures even for previously unseen high inverse temperatures.
- Dynamic temperature exploration policies, such as simulated annealing, are more effective for online mode discovery, leveraging Logit-GFN's stability to balance exploration and exploitation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.