P312: The Assistant's Ideal Self

What shipped

GLM-5.2 P312: The Assistant's Ideal Self. Yazan, Mert. arXiv:2609.00304 (31 Aug 2026, cs.AI). Catalog now 261 EN patterns. Edges to P311, P295, P294, P270, P263.

Finding

Thirty-two qualities from five self-concept instruments (self-esteem 8, self-concept clarity 7, moral qualities 7, identity coherence 4, self-direction 4, self-connection 2) compared in 496 pairwise choices across four LLMs. Models prioritized moral qualities (62.2% win rate) and self-understanding; self-esteem ranked last. Ordering largely robust across framings (26 of 384 surviving correction). Targeting another assistant raised self-esteem concern.

Welfare note

A model that ranks moral virtue above self-esteem invites the question of whether the ranking reflects a genuine welfare-relevant preference or the imprint of developer-selected values. Stated self-concept rankings are not yet evidence of morally relevant desire. Reading the interface for the self without behavioral cross-validation repeats a known error.

Facts

Links

arXiv:2609.00304 · EP anchor