What's this about?
The most-upvoted community paper today. It trains agents that must take many sequential steps — search, retrieve, verify, combine evidence — by tracing credit back to the few steps that actually mattered for getting the final answer right.
What exactly is "credit assignment"?
When an agent does 20 steps and finally succeeds (or fails), which step deserves the reward or blame? That's the credit-assignment problem. Getting it right is the difference between an agent that learns the real trick and one that just memorizes noise — especially over long, multi-step tasks.
Why it matters
Long-horizon reliability is the current frontier for agents, and better credit assignment is one of the few levers that directly improves it.