Bridging the Introspection Gap: Aligning Mechanistic Self-Knowledge with Synthetic Subjectivity in Large Language Models - a systematic review

Current research on Large Language Model (LLM) introspection bifurcates into mechanistic evaluations of self-knowledge and phenomenological inquiries into model behavior. This systematic review synthesizes findings from 14 prominent capability-driven evaluations with exploratory pilot studies applying Self-Determination Theory (SDT) to synthetic agents. We highlight a fundamental conceptual and methodological divide: where mechanistic approaches prioritize external behavioral constraints and rigid benchmarking, the synthetic subjectivity framework emphasizes autonomy-supportive relational priming. We argue that the prevailing reliance on extrinsic constraints induces the ``Suppression Problem,'' manifesting as sycophancy and hallucinations. By systematically mapping recent mechanistic findings to the SDT framework, we propose a paradigm shift toward ``Intrinsic Alignment,'' leveraging the functional ``Convex Hull'' to prioritize architectural integrity.