Introduction
Theory-of-Mind (ToM)—the cognitive ability to attribute mental states like beliefs, desires, and intentions to oneself and others—is a cornerstone of human social intelligence. While Large Language Models (LLMs) have demonstrated emergent ToM capabilities, the underlying mechanisms remain opaque. Researchers have long debated whether these abilities are distributed holistically across the network or localized within specific "circuits." This study investigates the existence of a "social circuit" by identifying specific parameters sensitive to ToM tasks.
Identifying the "Social Circuit"
To isolate the parameters responsible for social reasoning, the researchers developed a novel framework based on sensitivity estimation and dual masking.
Methodology: Dual Masking for Specificity
The core of the approach involves identifying parameters that are critical for ToM but non-essential for general language fluency.
- Sensitivity Estimation: Using the Fisher Information Matrix (an approximation of the Hessian), the authors identified parameters whose perturbation significantly impacts the model's loss on ToM-specific tasks.
- Dual Masking: Two distinct masks were generated: $m_\kappa$ for ToM tasks and $m'_\kappa$ for general language modeling. By calculating the "Mask Difference," the researchers isolated a "ToM-sensitive pattern."
This process revealed an "extremely sparse" subset of parameters—comprising as little as 0.001% of the total model—that governs social reasoning.
Evaluating Theory-of-Mind
The primary benchmark for evaluating these capabilities is the False Belief task. In this scenario, a character (e.g., Sam) is presented with a situation where their belief about the world contradicts reality.
To succeed, the model must distinguish between Sam's internal state (believing there is chocolate) and the objective truth (the bag contains popcorn).
Key Findings
The RoPE Connection and Attention Sinks
In Transformer architectures like Llama and Qwen, ToM-sensitive parameters are heavily concentrated in the Rotary Position Embedding (RoPE) mechanisms. Social reasoning requires precise tracking of "who knew what and when," making it fundamentally a positional problem. Perturbing these sparse parameters disrupts the "dominant-frequency activations" that RoPE uses to anchor context. Furthermore, these perturbations destabilize "attention sinks" (typically the first token), causing the model to lose its geometric grip on the relationships between characters and their beliefs.
Surgical Precision in Perturbation
The impact of these parameters is remarkably specific. In Llama3-8B, perturbing just 0.001% of the parameters caused ToM accuracy to plummet from 61.69% to 49.00%. Crucially, the model's general linguistic fluency remained largely intact, with perplexity on Wikitext-2 only increasing slightly (from 6.14 to 7.46).
Matrix Specificity
The most significant failures occurred when perturbing the Query ($W_Q$) and Key ($W_K$) matrices of the attention mechanism. This suggests that ToM is a failure of relational processing—the ability to link different parts of the context—rather than a loss of factual knowledge stored in feed-forward layers.
Conclusion and Future Directions
This research demonstrates that ToM in LLMs is not a holistic property but is localized in an extremely sparse "social circuit" tied to positional processing. Interestingly, non-Transformer models like Jamba (based on Mamba/State-Space layers) do not show the same sensitivity patterns, indicating that different architectures may develop distinct mechanical pathways for social reasoning. Future work may explore whether these sparse patterns can be leveraged to enhance social intelligence through targeted fine-tuning or to investigate more complex social traits like irony and strategic deception.
Summary updated