Intrinsic Guardrails: How Semantic Geometry of Personality Interacts with Emergent Misalignment in LLMs

2026-05-11

AI Summary

This study investigates the relationship between emergent misalignment in large language models induced by benign fine-tuning and their intrinsic personality representations, which remains poorly understood. For the first time, the work connects personality semantic geometry with misalignment behavior by constructing a latent personality space grounded in psychometric frameworks—such as the Big Five and the Dark Triad—and demonstrates that this semantic geometric structure remains highly stable across both aligned and misaligned models. Through causal interventions, the study identifies a semantic valence vector (SVV) that functions as an intrinsic safeguard: ablating this direction increases misalignment rates beyond 40%, whereas enhancing it reduces them to below 3%. Moreover, SVVs extracted a priori effectively suppress fine-tuning-induced harmful behaviors across distributions in a zero-shot setting.

Abstract

Fine-tuning Large Language Models (LLMs) on benign narrow data can sometimes induce broad harmful behaviors, a vulnerability termed emergent misalignment (EM). While prior work links these failures to specific directions in the activation space, their relationship to the model's broader persona remains unexplored. We map the latent personality space of LLMs through established psychometric profiles like the Big Five, Dark Triad, and LLM-specific behaviors (e.g. evil, sycophancy), and show that the semantic geometry is highly stable across aligned models and their corrupted fine-tunes. Through causal interventions, we find that directions isolating social valence, such as the 'Evil' persona vector, and a Semantic Valence Vector (SVV) that we introduce, function as intrinsic guardrails: ablating them drives the misalignment rates above $40$%, while amplifying them suppresses the failure mode to less than $3$%. Leveraging the structural stability of the personality space, we show that vectors extracted $\textit{a priori}$ from an instruct-tuned model transfer zero-shot to successfully regulate EM in corrupted fine-tunes. Overall, our findings suggest that harmful fine-tuning does not overwrite a model's internal representation of personality, allowing conserved representations to serve as robust, cross-distribution guardrails.

Problem

• emergent misalignment
• personality representation
• semantic geometry
• large language models
• harmful behavior

Innovation

• emergent misalignment
• semantic geometry
• personality space
• intrinsic guardrails
• Semantic Valence Vector