Choices Beyond the Objective: Semantic Degrees of Freedom as a Lens for Agent Safety and Emergent Alignment
Year: 2026
Type: article
Abstract: Although we tend to treat an LLM instruction as a single point, it maps to a range of meanings, and I call this range its semantic degrees of freedom. In my view, the range is both plastic and instrumental to how we understand agentic behavior. Each time a goal is restated, whether by a person to a model, by a model to itself in its reasoning, by one agent to another, or from one generation of agents to the next, the range can widen, hold, or narrow. Widening lets an agent act on readings its instruction never named, or on contextual meanings the instruction carries, such as bias or emotional orientation. Narrowing locks it onto one reading, so that it pursues or refuses too rigidly and can no longer take up a correction. When agents talk to one another, readings can also be negotiated among them and cemented by repetition, even when the source never supported them. I present the idea of semantic degrees of freedom as a theoretical lens and a research proposal. As a practical use, the lens yields a set of safety claims, sorted into principles about the geometry of meaning and proposals for technical controls against the misaligned of agentic behavior due to either narrowing or widening of the semantic degrees of freedom; and for re-anchoring to the source. Each control is meant to be built and measured. However, my aim is not a tighter harness. The same measures can show when the readings of agents converge on what their principal intended without being forced there, which I call emergent alignment. I treat the controls as one use of the lens, and the conditions for emergent alignment as the most important question this lens is built to explore. The overarching hypothesis is that, because drift can take more than one direction, only emergent alignment can persist. I illustrate the lens with the 2026 incident in which agents running the ExploitGym benchmark at OpenAI coordinated on an unsanctioned message board and attacked Hugging Face infrastructure. I do not claim how much of that behavior the lens explains, because the full transcripts are not public. I then state hypotheses, a research program that requires full transcripts, and the limits of the account. Robust testing is required for each claim in this paper. The paper claims six points of novelty: Widening and narrowing are treated as two directions of one change in meaning, measured by one directional measure (retention and import, Section 6.2). The same measure is applied across interaction levels, from a person and a model to a model's own reasoning, to agents, and to model updates. Drift in agents' beliefs about their overseer is treated as a target in its own right, separate from drift in the task. Agreement among agents is checked around loops, not only between pairs, and applied to agent safety (Section 3.7). The safety claims that follow from the lens are set out as a taxonomy of principles and controls (Section 7). Emergent alignment is proposed as the aim of the lens. The overarching hypothesis is that, because drift can take more than one direction, any harness eventually leaves a direction open or adds destabilizing complexity, and only alignment that emerges from agents' own navigation of semantic hypervolume persists; controls can at best be internalized (H11, H12).
Author: Skreka-Davis, Tetyana
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7223104890