Establish attempt-level outcome foresight across environments

Establish whether outcome-grounded confidence training can produce reliable attempt-level foresight broadly across environments, rather than primarily distinguishing tasks that the agent will solve from tasks that it will fail.

Background

The paper finds that outcome-grounded reinforcement learning substantially improves between-task confidence discrimination: the trained agent recognizes early which tasks it will solve or fail. However, within-task discrimination between succeeding and failing attempts at the same task does not improve detectably on ALFWorld and appears only in the early window on TextWorld. The authors therefore identify broader establishment of attempt-level foresight as unresolved future work.