MENTIS: What Belief Changes Under Alignment? Measuring Multi-Scale Latent Torsion in Language Models

2026-05-31

AI Summary

It remains unclear how preference alignment reshapes the internal representational geometry of large language models, as behavioral evaluations alone are insufficient to uncover underlying mechanisms. To address this, this work proposes the MENTIS framework, introducing for the first time covariance-based twist norms, spectral twist diagnostics, and the ERA metric to quantitatively characterize multi-level representational differences between instruction-tuned and preference-aligned models from a geometric perspective. Multi-scale analyses of four 7–8B model variants on the LITMUS benchmark reveal that normative concepts undergo substantially greater geometric twisting than factual ones, that the degree of twisting is negatively correlated with contextual entropy, and that these effects are concentrated in architecture-specific middle-to-late layers—highlighting the selective and deeply localized nature of the alignment process.

Abstract

Preference alignment has substantially improved the observable behavior of large language models, yet it remains unclear what alignment changes internally. Aligned systems still fail under jailbreaks, prompt injection, and retrieval-time corruption, suggesting behavior-level evaluation alone is incomplete. Post-training should leave measurable traces in internal computation. We ask: when an instruction-tuned (IT) model becomes a preference-aligned (PA) model, what geometric structure changes, where do those changes concentrate, and how selectively do they vary across concepts, prompts, and model families? We introduce MENTIS, a geometry-first framework for measuring alignment-induced internal reorganization in paired checkpoints. MENTIS compares IT and PA models using a primary layerwise covariance-based torsion norm (T1), a secondary spectral torsion diagnostic (T2), and an Energy-Radiance-Activation measure (ERA) for depth localization. Across four 7-8B model pairs on LITMUS, our study reveals that alignment-induced change is selective rather than uniform: normative concepts exhibit larger torsion shifts than factual concepts on average; torsion is negatively correlated with contextual entropy; and peak effects localize to architecture-specific mid-to-late layers. The same pattern appears across word-level, prompt-level, and model-level analyses. These results suggest preference alignment leaves structured, depth-localized geometric signatures in internal computation beyond what behavior-level evaluation alone can reveal.

Problem

• preference alignment
• latent representation
• internal computation
• geometric structure
• language models

Innovation

• alignment-induced torsion
• geometric reorganization
• layerwise covariance
• preference alignment
• internal representation