Mind the Perspective: Let's Reason Recursively for Theory of Mind

Researchers introduce RecToM, a framework that improves Large Language Models' Theory of Mind reasoning by modeling nested beliefs through recursive perspective construction. The approach achieves state-of-the-art results on multiple benchmarks, including 100% accuracy on Hi-ToM, demonstrating significant advances in how AI systems infer agent beliefs and intentions.

RecToM addresses a fundamental limitation in current LLM capabilities: reasoning about nested beliefs and mental states from incomplete information. Theory of Mind—the ability to infer what others believe, know, or intend—remains difficult for language models despite their scale. Previous approaches relied on event filtering or temporal chains without explicitly handling the hierarchical nature of beliefs, where understanding one agent's perspective requires modeling how they view another agent's beliefs. The RecToM framework solves this by recursively constructing character perspectives along chains specified in questions, effectively flattening higher-order belief problems into simpler actual-world reasoning tasks. This represents a methodological shift from linear processing to nested reasoning structures. The mathematical grounding via KD45 analysis validates that the approach creates proper belief modality rather than crude filtering. The benchmark results—100% accuracy on Hi-ToM and strong performance on Big-ToM and FanToM—indicate RecToM generalizes effectively across models and complexity levels. This advance matters because Theory of Mind underpins human-AI interaction, negotiation, game theory applications, and trustworthy AI systems. For developers building conversational AI, reasoning engines, or multi-agent systems, improved ToM capabilities enable more natural and accurate agent modeling. Looking ahead, the real-world applicability of these improvements depends on whether gains translate beyond controlled benchmarks to diverse real-world scenarios. Integration into production LLMs and evaluation on dynamic, multi-party interactions will determine practical impact.

• →RecToM uses recursive perspective construction to model nested beliefs, reducing complex multi-level reasoning to simpler problems.
• →The framework achieves 100% accuracy on Hi-ToM benchmark and state-of-the-art results across multiple Theory of Mind benchmarks.
• →KD45 formal analysis confirms RecToM induces well-formed belief modality beyond simple event filtering approaches.
• →Performance gains hold across multiple LLM backbones including GPT-5.4 and Qwen3.5, indicating broad applicability.
• →Improved Theory of Mind reasoning enhances AI capabilities for multi-agent systems, negotiation, and human-AI interaction.