SPARK: Self-Play with Asymmetric Reward from Knowledge Graphs

Hyobin Park, Taeseop Kim, Dong-Geol Choi

arXiv:2605.05546 · 2026-05-08 공개 · arXiv · PDF

Abstract

Self-play reinforcement learning has shown strong performance in domains with formally verifiable structure, such as mathematics and coding, where both problem generation and reward computation can be grounded in explicit rules. Extending this paradigm to scientific literature is more challenging: the relationships among multi-modal elements within and across documents are rarely made explicit in text, which makes automatic generation of relational reasoning questions difficult and weakens the reliability of reward signals. We propose SPARK (Self-Play with Asymmetric Reward from Knowledge Graphs), a framework that automatically constructs a unified knowledge graph (KG) from multi-document scientific literature and uses it as the structural basis for self-play. KG paths over multimodal nodes serve as a source for generating relational reasoning questions, and structured facts stored in the KG provide a basis for verifiable reward computation. A single small vision-language model (sVLM) alternates between Proposer and Solver roles under information asymmetry against a fixed KG, a design that we believe can be naturally extended toward online adaptation in future work. We evaluate SPARK on public benchmarks and a self-constructed cross-document multi-hop QA dataset. Results show that SPARK consistently outperforms flat-corpus-based self-play baselines, and the performance gap widens as hop count increases, suggesting that KG-structure grounding contributes to relational multi-hop reasoning beyond what unstructured corpus grounding can provide.

한국어 요약

📋 한 줄 요약

**[자기학습 LLM]** 다중 문서 과학 문헌으로부터 통합 지식 그래프를 자동 구축하고, 이를 기반으로 비대칭 보상 self-play를 수행해 멀티홉 관계 추론 능력을 강화하는 SPARK 프레임워크를 제안한다.

🎯 핵심 기여도

💡 핵심 아이디어

수학·코딩과 달리 과학 문헌은 문서 간·내 관계가 텍스트에 명시되지 않아 self-play의 보상 신호가 약하다. SPARK는 KG라는 명시적 관계 구조를 self-play의 "검증자"로 사용함으로써, 자동 생성되는 관계 추론 질문에 대해서도 신뢰할 수 있는 보상을 제공한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: 검증 불가능했던 과학 문헌 도메인에 형식적 구조(KG)를 도입해 self-play RL을 일반화할 수 있는 길을 제시. **한계**: 고정 KG에 의존하여 KG 자체의 누락·오류가 보상 편향으로 이어질 수 있고, 온라인 적응은 미래 과제로 남음.

🚀 실용적 활용