OSCToM: RL-Guided Adversarial Generation for High-Order Theory of Mind

Sharmin Sultana Srishty, Kazi Mahathir Rahman, Malaika Parizat Sakkhi, Samia Shahid Prianna, Shaikhul Islam Sinat

arXiv:2605.20423 · 2026-05-22 공개 · arXiv · PDF

reinforcement-learning llm-evaluation data-synthesis theory-of-mind adversarial-generation nested-beliefs fan-to-m hi-to-m

Abstract

Large Language Models (LLMs) perform well on many language tasks, but their Theory of Mind (ToM) reasoning is still uneven in complex social settings. Existing benchmarks, including ExploreToM, do not always test the recursive beliefs and information asymmetries that make these settings difficult. This paper presents OSCToM (Observer-Self Conflict Theory of Mind), an approach for modeling nested belief conflicts in LLM-based ToM tasks. The key case is one in which an observer's view of another agent conflicts with the observer's own belief state. Such cases go beyond simple perspective-taking and require recursive, multi-layered reasoning. OSCToM combines reinforcement learning (RL), an extended domain-specific language, and compositional surrogate models to generate observer-self conflicts. In our experiments, OSCToM-8B gives the best overall result among the systems tested. It improves on the reported ExploreToM results on FANToM and remains competitive on Hi-ToM and BigToM. On the information-asymmetric FANToM benchmark, OSCToM reaches 76% accuracy, compared with the 0.2% reported by ExploreToM. The data-synthesis procedure is also 6x more efficient, indicating that targeted training data can help smaller models handle advanced cognitive reasoning. The project code is available at https://github.com/sharminsrishty/osct.

한국어 요약

한 줄 요약

**[Theory of Mind / RL Adversarial Generation]** OSCToM이 observer-self conflict로 nested belief 충돌을 RL·DSL·compositional surrogate로 생성, 8B 모델이 FANToM 76%(ExploreToM 0.2%)·6× efficient data synthesis로 high-order ToM 향상.

핵심 기여도

핵심 아이디어

LLM의 high-order ToM은 단순 perspective-taking이 아니라 observer 자신과 observer의 다른 agent 인식 사이 nested 충돌의 recursive reasoning이 핵심이며, 이를 RL+DSL+compositional surrogate로 adversarial 생성하면 소형 모델도 frontier 수준의 ToM 능력에 도달한다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: high-order ToM의 adversarial 학습 데이터 합성 방법론 정립, 소형 모델로 frontier ToM 능력 도달, observer-self conflict라는 새 어려움 축 정립. **한계**: DSL·compositional surrogate 의존, social reasoning task에 특화, 매우 깊은 nested belief의 일반화는 추가 검증.

실용적 활용