Beyond Surface Alignment: Grounding the Dynamics of Situational Understanding and Generative Control in LLMs

Current alignment tuning prioritizes surface fluency, masking a lack of deep grounding. This thesis proposes "Grounded Alignment," analyzing failures in situational understanding (SitTest, ReCode) and generative control (Branching Factor). We reveal that alignment often induces "stylistic collapse" rather than true comprehension. Finally, we introduce mechanisms like Base-Aligned Model Collaboration to build agents robustly anchored in their context and generation.