Reproduce “Trusting Imagined Futures: Auditing Imagined Rewards for VLA Post-Training” 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
Zicheng Kong, Zhenbo Xu, Haoyan Zhang, Jinghan Yang, Zhenzhen Li, Yiwei Ru, Liuyu Xiang, WanTing Xu and 2 more
TIF selects candidate actions for behavior cloning by ranking imagined futures with a reward model trained on paired real executions. It accepts the highest-utility candidate only when its predicted advantage over Source exceeds 0.15 and audit risk is at most 0.50; otherwise it retains Source. On LIBERO, TIF reaches 76.25% mean task success versus 73.25% for Source across four suites and three seeds, using OpenVLA.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.