AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

KBQA-R1: Reinforcing Large Language Models for Knowledge Base Question Answering

Xin Sun, Zhongqi Chen, Xing Zheng, Bowen Song, Qiang Liu, Shu Wu, Zilei Wang, Weiqiang Wang, Liang Wang

ICML 2026 · Large language models › Retrieval and knowledge · approach: Knowledge graph reasoning with LLMs · 0 citations (OpenAlex)

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 relation retrieval and confidence gating (RRCG) (ablation)KBQA-R1 · GrailQAF1: 67.1 -> 86.1Improved
use multi-turn interaction instead of single-turn generation (ablation)KBQA-R1 · GrailQAF1: 49.8 -> 86.1Improved
use Referenced Rejection Sampling instead of standard rejection sampling (ablation)KBQA-R1 warm-start · WebQSPF1: 78.9 -> 83.4Improved
add SFT warm-start before RL (ablation)KBQA-R1 · WebQSPF1: 75.2 -> 83.4Improved
add GRPO RL after SFT (ablation)KBQA-R1 · WebQSPF1: 72.1 -> 83.4Improved
add format reward (ablation)KBQA-R1 · GrailQAF1: 84.2 -> 86.1Improved
use RRS rather than standard RS for trajectory collection (ablation)SFT data synthesis · GrailQAacceptance rate (%): 39.3 -> 67.0Improved
use RRS data instead of standard RS data for SFT init (ablation)SFT initialization · GrailQASFT init F1: 73.8 -> 80.2Improved
give frontier models the KBQA-R1 structured harness instead of SPARQL ReActClaude 4.6 Sonnet · 500-query subset WebQSPF1: 54.2 -> 76.7Improved
train learned RL policy instead of using harness with best frontier modelKBQA-R1 · 500-query subset averageF1: 71.6 -> 79.0Improved
GRPO training on relation groundingKBQA-R1 · WebQSPtop-1 relation similarity: 0.9400 -> 0.9827Improved
KBQA-R1 vs DecAF on GrailQA I.I.D F1KBQA-R1 · GrailQA I.I.DF1: 92.4 (DecAF) vs 91.5Hurt

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