arXiv cs.CL· Yucheng Li, Mingyuan Li, Xiaxia Wang, Gong Cheng·· 2 天前AI 评分22
OpCredit：面向智能体知识图谱问答的操作感知策略学习
OpCredit: Operation-Aware Policy Learning for Agentic Knowledge Graph Question Answering
AI 导读
OpCredit 是一个操作感知的策略学习框架，通过分支结果信用（BOC）与前沿转移信用（FTC）两个互补信号，为智能体知识图谱问答中每个可执行操作单独评估贡献，而非依赖轨迹级奖励。在 ComplexWebQuestions、WebQSP 和 GrailQA 三个基准、三种骨干模型上，该方法均取得一致的 KGQA 性能提升，推理时仅需已学习的策略和标准 KG 接口。
正文 · 原文
Abstract:Agentic knowledge graph question answering(KGQA) relies on sequences of executable operations, but trajectory-level rewards obscure their individual contributions. Successful trajectories may contain ineffective operations, while failed trajectories may contain useful ones. To assess individual operations, we propose OpCredit, an operationaware policy-learning framework that combines downstream outcome credit with immediate transition credit: Branch Outcome Credit (BOC) reuses ordinary rollouts to compare terminal task rewards observed after different operations under matched navigation contexts; Frontier Transition Credit (FTC) measures operation-induced changes in answer-conditioned frontier potential. The two complementary signals jointly supervise the corresponding decision spans during training, while inference uses only the learned policy and standard KG interface. Experiments on ComplexWebQuestions, WebQSP, and GrailQA across three backbones demonstrate consistent improvements in KGQA performance.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.21345 [cs.CL] |
| (or arXiv:2610.21345v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.21345 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yucheng Li [view email]
[v1]
Thu, 15 Oct 2026 06:29:17 UTC (454 KB)
来源：arXiv cs.CL · arxiv.org