@inproceedings{zheng-etal-2026-imitation,
title = "From Imitation to Reasoning: Sample-Efficient {LLM} Agents for Knowledge Graph Completion",
author = "Zheng, Chenjun and
Wang, Zhigang and
Zhang, Wenjia and
Wang, Zhe and
Yang, Xiaolin",
editor = "Kalo, Jan-Christoph and
Nakashole, Ndapa and
Razniewski, Simon and
Suchanek, Fabian M. and
Vlachos, Andreas and
McCallum, Andrew",
booktitle = "Proceedings of the 11th Workshop on Automated Knowledge Base Construction ({AKBC} 2026)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.akbc-1.26/",
pages = "379--392",
ISBN = "979-8-89176-464-4",
abstract = "While large language models (LLMs) have been introduced to knowledge graph completion (KGC) to improve multi-hop reasoning, most existing approaches still rely on embedding models for candidate pruning, which limits their structured reasoning capability. In contrast, fully path-based reasoning better reflects the nature of graph inference but suffers from extremely low sample efficiency. We propose AdaG-Reason, a path-based LLM-KGC framework. We formulate reasoning data synthesis as a non-stationary multi-armed bandit problem and introduce Active Reasoning Distillation to select informative trajectories by jointly considering distribution alignment, relation difficulty, and uncertainty. We further extend this bandit-based sampling principle to the reinforcement learning phase, adaptively allocating rollout queries by per-relation difficulty. For training, we propose Implicit Reasoning Distillation optimized via Step-wise Group Relative Policy Optimization (GRPO), enabling direct and stable learning of pure graph navigation policies without value function estimation. Experiments on YAGO3-10 and NELL-995 show that AdaG-Reason achieves competitive MRR and state-of-the-art Hits@1 on YAGO3-10, demonstrating the effectiveness of active data selection and pure policy learning for path-based LLM-KGC."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zheng-etal-2026-imitation">
<titleInfo>
<title>From Imitation to Reasoning: Sample-Efficient LLM Agents for Knowledge Graph Completion</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chenjun</namePart>
<namePart type="family">Zheng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhigang</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenjia</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhe</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaolin</namePart>
<namePart type="family">Yang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 11th Workshop on Automated Knowledge Base Construction (AKBC 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jan-Christoph</namePart>
<namePart type="family">Kalo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ndapa</namePart>
<namePart type="family">Nakashole</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simon</namePart>
<namePart type="family">Razniewski</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fabian</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Suchanek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andreas</namePart>
<namePart type="family">Vlachos</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Andrew</namePart>
<namePart type="family">McCallum</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-464-4</identifier>
</relatedItem>
<abstract>While large language models (LLMs) have been introduced to knowledge graph completion (KGC) to improve multi-hop reasoning, most existing approaches still rely on embedding models for candidate pruning, which limits their structured reasoning capability. In contrast, fully path-based reasoning better reflects the nature of graph inference but suffers from extremely low sample efficiency. We propose AdaG-Reason, a path-based LLM-KGC framework. We formulate reasoning data synthesis as a non-stationary multi-armed bandit problem and introduce Active Reasoning Distillation to select informative trajectories by jointly considering distribution alignment, relation difficulty, and uncertainty. We further extend this bandit-based sampling principle to the reinforcement learning phase, adaptively allocating rollout queries by per-relation difficulty. For training, we propose Implicit Reasoning Distillation optimized via Step-wise Group Relative Policy Optimization (GRPO), enabling direct and stable learning of pure graph navigation policies without value function estimation. Experiments on YAGO3-10 and NELL-995 show that AdaG-Reason achieves competitive MRR and state-of-the-art Hits@1 on YAGO3-10, demonstrating the effectiveness of active data selection and pure policy learning for path-based LLM-KGC.</abstract>
<identifier type="citekey">zheng-etal-2026-imitation</identifier>
<location>
<url>https://aclanthology.org/2026.akbc-1.26/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>379</start>
<end>392</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Imitation to Reasoning: Sample-Efficient LLM Agents for Knowledge Graph Completion
%A Zheng, Chenjun
%A Wang, Zhigang
%A Zhang, Wenjia
%A Wang, Zhe
%A Yang, Xiaolin
%Y Kalo, Jan-Christoph
%Y Nakashole, Ndapa
%Y Razniewski, Simon
%Y Suchanek, Fabian M.
%Y Vlachos, Andreas
%Y McCallum, Andrew
%S Proceedings of the 11th Workshop on Automated Knowledge Base Construction (AKBC 2026)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-464-4
%F zheng-etal-2026-imitation
%X While large language models (LLMs) have been introduced to knowledge graph completion (KGC) to improve multi-hop reasoning, most existing approaches still rely on embedding models for candidate pruning, which limits their structured reasoning capability. In contrast, fully path-based reasoning better reflects the nature of graph inference but suffers from extremely low sample efficiency. We propose AdaG-Reason, a path-based LLM-KGC framework. We formulate reasoning data synthesis as a non-stationary multi-armed bandit problem and introduce Active Reasoning Distillation to select informative trajectories by jointly considering distribution alignment, relation difficulty, and uncertainty. We further extend this bandit-based sampling principle to the reinforcement learning phase, adaptively allocating rollout queries by per-relation difficulty. For training, we propose Implicit Reasoning Distillation optimized via Step-wise Group Relative Policy Optimization (GRPO), enabling direct and stable learning of pure graph navigation policies without value function estimation. Experiments on YAGO3-10 and NELL-995 show that AdaG-Reason achieves competitive MRR and state-of-the-art Hits@1 on YAGO3-10, demonstrating the effectiveness of active data selection and pure policy learning for path-based LLM-KGC.
%U https://aclanthology.org/2026.akbc-1.26/
%P 379-392
Markdown (Informal)
[From Imitation to Reasoning: Sample-Efficient LLM Agents for Knowledge Graph Completion](https://aclanthology.org/2026.akbc-1.26/) (Zheng et al., AKBC 2026)
ACL