Trained agent policies carry harness preferences that break under interface changes
Reordering MiniWoB++ observation fields drops trained success from 85.0% to 19.2% without changing optimal decisions.
1 stories, newest first · one card per story
Reordering MiniWoB++ observation fields drops trained success from 85.0% to 19.2% without changing optimal decisions.
Today 617 readers · 742 views · 563 reads · all time 158K views