Auditing Framing-Sensitive Behavioral Instability in Large Language Models for Mental Health Interactions

The paper's most pointed contribution isn't the behavioral audit itself but the move inward: examining how internal representations encode framing-driven instability, which suggests the problem isn't fixable through output-layer guardrails or prompt engineering alone.

This connects directly to the 'Where Do Models Find Happiness' coverage from the same day, which mapped emotion vectors across open-weight LLMs and found architectural differences in how emotional valence is encoded layer by layer. That work showed emotional structure is geometrically real inside these models but distributed unevenly across depth. The framing-sensitivity paper is essentially asking what happens when that internal emotional geometry gets destabilized by surface-level input variation, a question the emotion-vector work raises but doesn't answer. Together they suggest that mental health AI reliability can't be evaluated at the behavioral surface; it requires the kind of mechanistic interpretability lens that is still maturing. The 'Decision-Aligned Evaluation of Uncertainty Quantification' coverage adds another layer: if uncertainty metrics already fail to predict real-world decision quality in healthcare settings, framing-sensitive instability compounds that gap considerably.

Watch whether any of the open-weight models audited here, particularly those with known emotional encoding architectures like Gemma-4-E4B, show measurably lower framing sensitivity than models that concentrate emotional encoding early. That result would give alignment researchers a concrete architectural target rather than a post-hoc patching problem.