CoSToM:Causal-oriented Steering for Intrinsic Theory-of-Mind Alignment in Large Language Models

Researchers introduce CoSToM, a framework that uses causal tracing and activation steering to improve Theory of Mind alignment in large language models. The work addresses a critical gap between LLMs' internal knowledge and external behavior, demonstrating that targeted interventions in specific neural layers can enhance social reasoning capabilities and dialogue quality.

CoSToM represents a meaningful advancement in mechanistic interpretability applied to large language models, shifting from passive analysis to active intervention. The framework addresses a genuine limitation in current LLMs: while they perform adequately on standardized Theory of Mind benchmarks, they struggle with generalization to complex, real-world social reasoning tasks. This gap between benchmark performance and practical capability has significant implications for deployment in customer service, mental health support, and collaborative AI systems.

The research builds on growing interest in understanding what knowledge actually resides within LLM weights versus what emerges from prompt engineering. By mapping internal ToM feature distributions through causal tracing, the authors identify specific layers critical for social reasoning. This mechanistic understanding enables lightweight steering interventions that align internal representations with desired behaviors without full model retraining—a practical advantage for practitioners.

For the AI development community, this work validates the premise that systematic intervention on identified neural mechanisms can improve behavioral alignment. This has broader implications for AI safety and control, suggesting that understanding model internals enables more precise and efficient alignment techniques than broad architectural redesigns.

The practical impact remains to be determined. While the paper demonstrates improvements in dialogue quality and human-like reasoning, adoption depends on whether these gains transfer to diverse deployment contexts. Future work should examine whether CoSToM's benefits persist across different domains, model sizes, and interaction patterns. The framework's lightweight nature makes it accessible to organizations with limited compute resources, potentially accelerating adoption of more socially-aware AI systems.

• →CoSToM uses causal tracing to identify and steer Theory of Mind features in specific LLM layers without full retraining.
• →The framework bridges the gap between LLMs' internal representations and external social reasoning capabilities.
• →Mechanistic intervention on identified neural mechanisms proves more efficient than traditional fine-tuning approaches.
• →Results demonstrate measurable improvements in human-like social reasoning and dialogue quality.
• →The approach provides a scalable method for alignment that requires minimal computational overhead.