Towards an Ontological Framework for LLM Alignment ⋆
Year: 2026
Type: conference-paper
Abstract: The problem of AI alignment is informally defined as the problem of ensuring that AI systems behave in accordance with human instructions, intentions, preferences, goals, or values. It has attracted significant attention in AI safety research. It is especially important in the case of Large Language Model (LLM), given their large adoption, their black box nature and the difficulty to identify clearly how agentive notions can fit within their architecture. Moreover, the key concepts underlying alignment remain poorly defined. This paper argues that philosophy-informed applied ontology is uniquely positioned to address this issue. We propose a programmatic description of the principal questions that an ontology of LLM alignment should address, emphasizing the formalization of the central entities used in alignment definitions, including instructions, intentions, preferences, interests, goals and values. We also consider whether what is aligned should be an LLM, a chatbot or an agent, as well as the variety of misalignment phenomena. We argue that clarity on these questions is not merely of theoretical interest but is necessary for effective AI oversight.
Source: HAL
Authors: Barton, Adrien, Toyoshima, Fumiaki
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213733643
