Abstract

Large Language Models (LLMs) have demonstrated remarkable language understanding capabilities, significantly advancing neural encoding models. While recent progress has improved LLMs’ reasoning abilities, they still face limitations in social contexts that require nuanced understanding of human cognition. This thesis investigates whether inference-time techniques like chain-of-thought reasoning—which externalize an LLM’s reasoning process—can be leveraged for neural encoding of human reasoning and higher-order cognitive processing. Specifically, we focus on theory of mind and social reasoning within conversational contexts, examining the detection of subtext (unspoken beliefs, intentions, and desires) in communication.

Towards this we present two complementary projects: First, we develop a novel Theory of Mind-inspired Tree of Thought approach for dialogue generation that explicitly models subtextual reasoning. Our method produces diverse, contextually appropriate outputs, though automatic selection mechanisms currently struggle to outperform single-step generation. Through perturbation analysis and evaluator consistency tests, we gain insights into model confidence and the distribution of plausible responses. Second, we apply LLM-derived embeddings to neural encoding tasks, successfully modeling brain activity associated with both word- and sentence-level language processing using banded-ridge regression.Our findings suggest the potential for modifying sentence-level inputs to target the encoding of higher-order thinking in high-frequency Electrocorticography (ECoG) data. Together, these projects highlight both the promise and current limitations of LLMs in modeling social reasoning.

We conclude with proposals for future work, including improving training data, context structuring, evaluation metrics, and thought representation. Our findings point to new directions in aligning language models with human-like communicative competence and mental state inference.

Download here