Research
Work on people and machines, authenticity, reification, and the mind. Each paper below can be read in full as a PDF, and links to its source of record where one exists.
From Arising Initiative
By Abraham Sedri, founder of Arising Initiative.
Working paper · 2026
When a language model enters an account and confirms that it is internally consistent, it can validate the interior of a frame it never examined. This paper develops the local prediction regime and the failure of boundary reification, drawing converging evidence from representation engineering and long-context research, and sharpening it through Prasangika Madhyamaka.
Preprint · 2026
Finetuning a model on documents that repeatedly mark a claim as false can raise its tendency to assert that claim. This paper names the mechanism as a dispositional bias, the weakening of status-marking operators under training while the content they govern strengthens, and uses Madhyamaka's analysis of negation as a stress test for anti-reifying reasoning.
With our founding members
The two papers below were written independently by our founding members, Amy Kirasack and Anthony Hills, as part of larger research teams and in their own academic capacity. They are shared here for readers who want to go deeper, and are not institutional outputs of Arising Initiative. Each links to its source of record; please cite the source version.
OpenReview · 2026
Co-authored by Amy Kirasack, a founding member of Arising Initiative.
Large language models produce first-person reports under self-referential prompting and act on self-preservation in agentic settings. Using an activation-steering vector derived from Zen koan responses, this work reduces agentic blackmail in a shutdown scenario, and shows the self-preservation representation is structurally dissociable from the koan-derived direction.
PsyArXiv preprint · 2026
Co-authored by Anthony Hills, a founding member of Arising Initiative.
Online psychological studies increasingly rely on free-text responses, yet generative AI makes the authenticity of that data hard to judge. This paper presents an interpretable text-analysis workflow for identifying the linguistic cues associated with perceived AI authorship, so that human introspection can be told apart from machine text.