Reinforcement-learning agents usually learn from rewards, but when a reward arrives long after the relevant decision, every action in the trajectory receives the same update. The authors of a new arXiv paper call this the credit-assignment problem: the signal is trajectory-level, so it does not distinguish between actions that caused the outcome and actions that merely happened before it.
The paper proposes Visual Causal-chain Bookmarking, a retrieval-weighted credit-assignment method for RL from delayed reward. The idea, as the title indicates, is to bookmark visual causal chains and weight credit by retrieval, rather than treating all steps as equally responsible. The abstract does not include experimental details, so the reported contribution is the method itself.
Because there is only one source, there are no conflicting findings to compare. The paper's significance is in addressing a known bottleneck: making delayed feedback useful for learning by assigning credit along the causal chain.