Standard demonstrations teach agents hindsight, not exploration — SAFARI synthesizes exploration-rich trajectories instead

arXiv 2608.14339·low signalXBlueskyLinkedInCopy link

SAFARI targets proactive exploration, the ability to gather information now for better decisions later, and identifies why it is hard to train: normal demonstration data is written with hindsight, so the agent never sees the uncertainty that motivated the information-gathering step. It combines exploratory data construction that synthesizes exploration-rich trajectories with RL optimization using contrastive signal guidance over paired trajectories, which is what separates productive exploration from wasteful wandering. Code is public at github.com/GuanZhizhao/SAFARI; the abstract validates effectiveness without publishing headline numbers, so this is a design pattern to borrow rather than a benchmarked result.