Agentic knowledge graph question answering (KGQA) asks large language models to solve complex questions through multi-turn interactions with a knowledge graph. A new arXiv preprint, "Outcome-Anchored Credit Reallocation for Agentic Knowledge Graph Question Answering," tackles a core difficulty in this setting: outcome-supervised reinforcement learning only provides a final reward, leaving the model to figure out which intermediate actions contributed to success.
The paper proposes a method that reallocates credit using outcome anchors—presumably reference points derived from the final result—to better distribute reward across the steps of a multi-turn trajectory. This is intended to make learning more efficient when rewards are sparse and the action space is large, as is typical in KGQA.
The abstract is brief, so the full experimental setup and quantitative results are not available from the source. The contribution appears to be a training-time credit assignment mechanism rather than a new model architecture, and it is specifically aimed at the agentic setting where LLMs must reason and act over a knowledge graph iteratively.