Let’s Put Ourselves in Sally’s Shoes: Shoes-of-Others Prefilling Improves Theory of Mind in Large Language Models
Year: 2026
Type: conference-paper
Abstract: Recent studies have shown that Theory of Mind (ToM) in large language models (LLMs) has not reached human-level performance yet.Since fine-tuning LLMs on ToM datasets often degrades their generalization, several inferencetime methods have been proposed to enhance ToM in LLMs.However, existing inferencetime methods for ToM are specialized for inferring beliefs from contexts involving changes in the world state.In this study, we present a new inference-time method for ToM, Shoesof-Others (SoO) prefilling, which makes fewer assumptions about contexts and is applicable to broader scenarios.SoO prefilling simply specifies the beginning of LLM outputs with "Let's put ourselves in A's shoes.",where A denotes the target character's name.We evaluate SoO prefilling on two benchmarks that assess ToM in conversational and narrative contexts without changes in the world state and find that it consistently improves ToM across five categories of mental states.Our analysis suggests that SoO prefilling elicits faithful thoughts, thereby improving the ToM performance.ToMATO ToMBench Model Method B I D E K Avg.B I D E K Avg.GPT-3.5 Vanilla 58.5 57.0 69.8 61.8 55.6 60.5 51.7 63.6 49.4 62.3 32.1 51.8 turbo CoT Prompting 72.0 72.9 77.7 71.5 74.1 73.7 56.9 59.3 44.4 67.9 36.9 53.1 SoO Prompting 59.2 50.4 67.1 61.8 57.9 59.3 53.3 63.3 47.5 64.4 30.2 51.7 GPT-4o Vanilla 76.2 79.9 82.4 76.8 73.3 77.7 61.7 72.1 60.6 71.5 35.1 60.2 mini CoT Prompting 48.1 46.4 51.7 62.6 43.8 50.5 44.9 30.5 34.4 46.6 28.4 36.9SoO Prompting 75.7 79.7 83.7 75.6 72.3 77.4 64.4 60.0 51.9 71.0 34.3 56.3 Mistral Vanilla 62.0 68.0 74.5 60.7 62.4 65.5 50.5 56.3 51.0 61.6 27.1 49.3 7B CoT Prompting 62.7 70.7 74.5 62.4 65.5 67.2 53.7 58.5 50.2 60.5 30.6 50.7 SoO Prompting 63.9 68.9 76.5 62.7 63.3 67.1 53.3 57.9 50.4 61.6 29.1 50.5 CoT Prefilling 61.5 67.6 72.7 61.8 62.2 65.1 53.0 55.3 41.5 59.9 30.8 48.1 SoO Prefilling 64.9 70.0 75.6 63.0 64.5 67.6 56.2 58.9 47.3 63.7 34.5 52.1 Llama3 Vanilla 54.2 56.1 60.2 57.0 47.1 54.9 48.7 56.0 49.2 61.2 31.1 49.2 8B CoT Prompting 26.0 26.2 22.0 28.9 24.7 25.6 46.1 41.9 36.9 51.8 28.0 40.9 SoO Prompting 51.6 57.9 51.1 55.6 41.6 51.6 47.9 47.1 45.0 57.6 32.2 46.0 CoT Prefilling 64.1 65.3 71.0 60.8 58.9 64.0 55.3 65.2 51.9 63.4 37.2 54.6 SoO Prefilling 67.2 69.2 73.4 65.7 62.0 67.5 61.1 70.8 59.0 66.6 38.3 59.2 Llama3 Vanilla 81.7 85.3 85.9 80.5 73.5 81.4 73.6 79.8 58.5 71.9 45.9 66.0 70B CoT Prompting 80.5 85.2 86.7 81.3 74.1 81.6 68.2 78.8 54.0 69.6 50.6 64.2 SoO Prompting 81.9 86.2 87.6 82.2 75.6 82.7 73.3 80.7 59.4 72.8 51.9 67.6 CoT Prefilling 79.9 83.9 84.2 78.6 73.4 80.0 71.1 69.1 51.5 65.8 49.0 61.3 SoO Prefilling82.2 86.9 87.4 82.4 76.7 83.1 80.5 80.8 57.9 73.0 47.9 68.0
Institution: NTT (Japan)
Language: English
Subfield: Developmental and Educational Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7140140622