ToM in LLM is not ToM, but a Pragmatic Effect

Instruction tuning has been shown to improve large language models’ performance on pragmatic tasks, and recent work suggests that additional training can also enhance Theory of Mind (ToM)–like abilities. However, existing studies rarely examine how different alignment techniques and training data contribute to ToM-related behavior in language models. In this work, we investigate the respective roles of instruction tuning and preference learning in shaping pragmatic and ToM abilities. Using the LLaMA 3 8B architecture, we fine-tune models on either pragmatic or ToM-specific data and subsequently align them via Direct Preference Optimization. We evaluate the resulting models on benchmarks targeting both pragmatics and ToM. Our results show that pragmatic training can substantially improve ToM performance even without explicit belief-related supervision, and that instruction tuning plays a central rolei nmodel alignment. These findings clarify the relationship between pragmatics and ToM in large language models.