Reproduce “CURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery” 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
Tao Feng, Fangxu Yu, Zijie Lei, Jiaru Zou, Changjiang Jiang, Yi Yan, Jiaxuan You, Pan Lu
CURIO adds a learned curiosity signal to test-time learning for open-ended discovery. As a policy generates programs, an auxiliary model predicts changes in its internal representations; prediction difficulty on sampled tokens outside its top choice adds a weighted bonus alongside evaluator feedback. Across Qwen3 scales, it improves five of six mathematical objectives over task-only training, tying on Circle Packing; semantic exploration is not established, and the results do not show universal wins over other methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.