AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Backjump-on-Graph: Empowering Large Language Models with Reinforced Retrospective Exploration for Agentic Knowledge Graph Reasoning

Yunqi Zhang, Shiqi Yan, Zhenzhao Yuan, Wenrui Liang, Yangming Liu, Zhixiao Qi, Tianyi Zhang, Shijie Zhang, Wei-Qiang Zhang, Yongfeng Huang, Haixin Duan, Shuai Chen and 1 more

ICML 2026 · Large language models › Retrieval and knowledge · approach: Knowledge graph reasoning with LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add RL stage with hybrid reward after SFT (ablation)BoG, Llama-3.1-8B · CWQHits@1: 79.4 -> 84.0Improved
add RL stage with hybrid reward after SFT (ablation)BoG, Llama-3.1-8B · WebQSPHits@1: 88.5 -> 92.7Improved
add step-wise reward (format + faithfulness) (ablation)BoG, Llama-3.1-8B · CWQHits@1: 82.7 -> 84.0Improved
add step-wise reward (ablation)BoG, Llama-3.1-8B · WebQSPF1: 79.7 -> 81.3Improved
add efficiency reward (length penalty) (ablation)BoG, Llama-3.1-8B · CWQF1: 77.4 -> 78.3Improved
add efficiency reward (ablation)BoG, Llama-3.1-8B · WebQSPHits@1: 91.9 -> 92.7Improved
apply RL to the SFT agent (ablation)BoG-SFT · CWQaverage step length: 5.3 -> 3.6Improved
apply RL to the SFT agent (ablation)BoG-SFT · CWQbackjumping counts: 227 -> 302Improved
apply RL to the SFT agent (ablation)BoG-SFT · CWQfaithfulness: 0.78 -> 0.89Improved
fine-tune on oracle-conditioned synthetic backjump trajectories (BoG-SFT) vs closed-source LLMs in the same frameworkGemini-2.5 Flash / Pro, GPT-5.2 · CWQHits@1: 58.4 -> 79.4Improved
use BoG instead of agent-based KBQA-o1KBQA-o1 Llama-3.1-8B · CWQHits@1: 43.2 -> 84.0Improved
train with RL on one dataset and test on others (OOD)BoG vs BoG-SFT · WebQSP/CWQ/GrailQA OODHits@1Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy