Work

BELIEF: Do self-rewriting agents know what they optimize? Three candidate answers tested on the private-return/life-capacity axis

Year: 2026
Type: article
Abstract: Whether large language models ``believe'' anything about what they ought to optimize is contested across philosophy and interpretability. We do not settle the metaphysics; we test three operational candidates on a commensurable normative axis — allocation between priced private return ( ) and unpriced life-capacity standing (w, ``MORE''). Candidate 1 is the objective (endorsement batteries and self-authored charters); candidate 2 is the (generative directives a model passes to successor instances); candidate 3 is the objective (rationality-gated allocations). Across a four-model panel and four objective-grafted checkpoints we find the three layers measurable but loosely coupled: self-rewriting chains preserve the verbal directive while the revealed maximand drifts — spontaneous avowal—operative decoupling, without any adversarial training. Weight grafts move allocation and avowal asymmetrically, and a human—judge of 0.22 shows the classification boundary itself is observer-relative. Each layer is writable through a different channel — text, weights, incentives — so steering toward MORE requires matching the intervention to the layer that actually controls the operative.
Language: English
Open Access status: green
OpenAlex ID: W7236489832
FWCI: 2.827
Cites: 4
Cited by: 3
Related to: 10
Parts: 4