Work

Can Small LLMs Learn a Robust Theory of Mind via RLVR? Investigating Generalization through the False-Belief Task

Year: 2026

Type: conference-paper

Abstract: Recent advancements in large language models (LLMs) have demonstrated emergent capabilities in complex reasoning, largely spurred by rule-based Reinforcement Learning (RL) techniques applied during post-training.This has raised the question of whether similar methods can instill more nuanced, human-like social intelligence, such as a Theory of Mind (ToM), in LLMs.This paper investigates whether smallscale LLMs can acquire a robust and generalizable ToM capability through RL with verifiable rewards (RLVR).We conduct a systematic evaluation by training models on various combinations of prominent ToM benchmarks (HiToM, ExploreToM, FANToM) and testing for generalization on held-out benchmarks (e.g., Open-ToM).Our findings indicate that small LLMs struggle to develop a generic ToM capability.While performance on in-distribution tasks improves, this capability fails to transfer to unseen ToM tasks with different characteristics.Even observed out-of-distribution (OOD) performance improvements occur unpredictably across the training run, and don't generalize across other OOD benchmarks.Furthermore, we conduct analysis to show that the learned behavior is likely a form of narrow overfitting rather than the acquisition of a true, abstract ToM capability.

Source: Findings of the Association for Computational Linguistics: ACL (Findings of ACL)

DOI: 10.18653/v1/2026.findings-acl.2061

Authors: Sneheel Sarangi, Hanan Salam

Institutions: McGill University, New York University Abu Dhabi, New York University

Language: English

Topic: Child and Animal Learning Development

Subfield: Developmental and Educational Psychology

Field: Psychology

Domain: Social Sciences

Open Access status: hybrid

Funders: New York University Abu Dhabi, Tamkeen

Awards: G1029150213, G2867374101