1 paper · 1 filter
Jingxi Wei
Long-horizon coding-agent trajectories are poorly matched to the credit units available to train on: a single action has no stable value, an episode label merges productive explora…