Constitutional Value Potentials: reading and steering internal priority margins in language models

Researchers have introduced Constitutional Value Potentials (CVP), a method for reading and steering the internal priority margins that language models use to resolve value conflicts, according to a paper posted to arXiv on 13 June 2026 [1]. The approach learns a scalar potential from a model’s hidden state for each specified value, supervised by an independent judge’s verdict on which value the model’s own response preserved [1]. The signed difference between two such potentials forms a priority margin. A constitutional clause is then formulated as the claim that a given margin remains positive, and a single monitor score flags when it does not [1]. The monitor predicted conflict violations with an AUROC of up to 0.95 and outperformed a strong hidden-state probe across three scales of the Qwen2.5 model family [1]. The signal emerges as the answer begins, drawing on the prompt tail and the first response token. Read at that early stage, the same signal can reveal whether an adversarial priority hack has actually pushed the model toward a violation, rather than merely indicating that the prompt looks adversarial [1]. Under selected steering settings, moving along a value direction shifted judged trade-offs in the intended direction, the authors report [1]. Value conflicts in artificial intelligence are not merely technical; they echo long-running philosophical debates. Conservatism, for instance, has historically sought to preserve institutions such as organized religion, the nation-state, and property rights, though its tenets vary widely across cultures [2]. Zionism, which emerged in late 19th-century Europe as a secular nationalist movement, has been framed by advocates as a national liberation movement and by opponents as a settler-colonialist project [3]. These divergent framings illustrate the kind of contested values that a constitutional monitor might need to arbitrate. Efforts to encode broad societal priorities into measurable frameworks have faced their own difficulties. The United Nations’ 17 Sustainable Development Goals, adopted in 2015, aim to balance environmental, social, and economic objectives, but by 2025 only 35% of targets were on track or making moderate progress, with 18% moving in reverse [8]. The CVP paper does not claim to resolve such large-scale coordination problems, but it offers a mechanism for detecting when a model’s internal priorities deviate from a stated constitutional clause [1]. The work suggests that some constitution-relevant priorities are accessible as activation-space margins, not only as output behavior [1]. The authors tested the monitor on held-out synthetic conflicts and found that it generalized across model scales, providing evidence that the margin readout captures a structured signal rather than a brittle correlation [1].