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

The paper's most underreported implication is directional: if social valence vectors are stable even in corrupted models, that suggests alignment degradation is not uniform across a model's representational space, meaning some internal structures may be recoverable after fine-tuning goes wrong, not just resistant beforehand.

This connects directly to two threads running through recent coverage. The conformity paper ('Conformity Generates Collective Misalignment in AI Agent Societies') showed that individually aligned models can drift when placed in social dynamics, which raises the question of whether intrinsic geometric properties survive that kind of pressure. The activation steering paper ('Prompt-Activation Duality') is also relevant: it demonstrated that token-level interventions can preserve trait consistency across dialogue turns, which is essentially a practical application of the same intuition this paper formalizes theoretically. Together, the three papers sketch a coherent picture where alignment is less a policy layer and more a property of learned geometry, one that can be steered, stressed, or stabilized depending on deployment context.

The critical test is whether these valence vectors remain stable under the specific fine-tuning regimes used in production RLHF pipelines, not just the corrupted checkpoints studied here. If a follow-up replicates the stability finding on instruction-tuned variants of Llama or Mistral using publicly documented fine-tuning setups, the mechanistic claim becomes actionable for practitioners.