Maomei Wang

This paper studies how minimal-change constraints on preference revision bear on credal revision, with special attention to imprecise priors. Working within a preference-revision framework inspired by the AGM belief revision theory of Alchourrón, Gärdenfors, and Makinson (Journal of Symbolic Logic 50(2): 510–530, 1985), we examine how admissible revisions of preference relations constrain revisions of the credal states represented by imprecise priors. The precise-prior case provides a baseline: under suitable conditions, Bayesian conditioning can be recovered from the relevant minimal-change postulates. The main focus of the paper is the imprecise-prior case, where credal states are represented by sets of probability functions and the orthodox update rule is Pointwise Bayesian conditioning. We characterize the posterior preferences induced by Pointwise Bayesian conditioning and show that this characterization reveals a divergence: Pointwise Bayesian conditioning may require the removal of preference judgments that admissible minimal-change revisions can preserve. The resulting divergence cases show that prior imprecision may stem from different sources, and that these different sources can warrant different posterior responses when new information is incorporated. The analysis also clarifies how dilation can be blocked when the relevant prior likelihood judgments remain part of the reference set. The framework therefore provides a preference-based perspective for assessing Pointwise Bayesian conditioning and for identifying admissible alternatives in the imprecise settings.

Classical Bayesian arguments show how coherence between preference and credence grounds the norm of Probabilism. But these arguments are almost entirely static: they explain only how preference and credence must fit together at a given time. Once preferences change, the question arises: how should credences be revised in response? I develop an axiomatic minimal-change preference revision theory, in which some preference commitments are treated as more trusted and serve as reference anchors in revision. Focusing on standard event-occurrence inputs, I compare the induced revision dynamics with those implied by Bayesian conditioning, as captured at the level of preferences. The main convergence result shows that treating prior conditional preferences (together with preferences over constant acts) as reference anchors yields a unique admissible posterior preference relation that coincides with the Bayes-induced posterior preference relation. The framework also isolates a principled mechanism for divergence: alternative anchoring stances—especially those privileging selected unconditional commitments over conditional ones—can force revisions to conditional attitudes and thereby generate systematic departures from the Bayes-induced posterior preference relation.

Many have argued, based on the Instrumental Convergence Thesis, that Artificial General Intelligences (AGIs) will exhibit power-seeking behavior. Such behavior, they warn, could harm human society and pose existential threats—namely, the risk of human extinction or the permanent collapse of civilization. These arguments often rely on an implicit and underexamined assumption: that AGIs will develop world models—internal representations of world dynamics—that resemble those of humans. We challenge this assumption. We argue that once the anthropomorphic assumption—that AGIs’ world models will mirror our own—is rejected, it becomes unclear whether AGIs would pursue the types of power commonly emphasized in the literature, or any familiar types of power at all. This analysis casts doubt on the strength of existing arguments linking the Instrumental Convergence Thesis to existential threats. Moreover, it reveals a deeper layer of uncertainty. AGIs with non-human world models may identify novel or unanticipated types of power that fall outside existing taxonomies, thereby posing underappreciated risks. We further argue that world model alignment—an issue largely overlooked in comparison with value alignment—should be recognized as a core dimension of AI alignment. We conclude by outlining several open questions to inform and guide future research.

Epistemic decision theory aims to defend a variety of epistemic norms in terms of their facilitation of epistemic ends. One of the most important components of EpDT is known as a scoring rule. This thesis addresses some problems about scoring rules in EpDT. I consider scoring rules both for precise credences and for imprecise credences. For scoring rules in the context of precise credences, I examine the rationale for requiring a scoring rule to be strictly proper, and argue that no satisfactory justification has been given. I then investigate one possible response to my argument and show the problems with this response. The conclusion is that there is a further need for justifying the requirement that a scoring rule should be strictly proper for precise credences. For scoring rules in the context of imprecise credences, an impossibility result has been established in the literature purporting to show that no strictly proper, continuous and real-valued scoring rule exists. However, a precise statement of the impossibility result requires precise definitions both of strict propriety and of continuity in the context of imprecise credences. Moreover, the result implies that we need to drop one of the three properties - strict propriety, continuity or being real-valued - to have a scoring rule for applying EpDT to imprecise credences. So, firstly, I offer definitions of strict propriety and of continuity and clarify the impossibility result. Then, I investigate what will happen if we drop one of the three properties. I argue that we should drop the property of being real-valued and I offer the general forms of two kinds of strictly proper, continuous and lexicographic scoring rules for imprecise credences.