MENTIS: What Beliefs Does Alignment Change? Measuring Multi-Scale Latent Torsion in Language Models

Researchers propose the MENTIS framework, which uses geometric metrics to measure the reorganization that preference alignment induces in the internal computation of large language models. The framework compares paired instruction-tuned (IT) and preference-aligned (PA) checkpoints, combining a layer-wise covariance-based torsion norm T1, a spectral torsion diagnostic T2, and an ERA (Energy-Radiance-Activation) metric for depth localization. Experiments on LITMUS with four 7-8B model pairs show that alignment-induced changes are selective rather than uniform: the average torsion change is larger for normative concepts than for factual concepts, torsion is negatively correlated with context entropy, peak effects concentrate in middle-to-late layers that vary by architecture, and the same pattern appears consistently across word-level, prompt-level, and model-level analyses. The authors claim these results show that preference alignment leaves structured, depth-localized geometric signatures in internal computation, beyond what behavioral-level evaluation alone can reveal.