Structured Phenomenological Descriptions Induce Analysis-Mode Behavior in a Small Language Model
Year: 2026
Type: preprint
Abstract: We report an empirical observation in Mistral-7B-Instruct-v0.3 (Q4_K_M). When promptedwith a glyph, a three-axis phenomenological description of a recurring decision class, themodel enters an analysis register: it analyzes the glyph, in 17 of 24 runs is coded as recog-nizing the decision class without acting on it, and in no run produces task execution. Acrossthree independent legs totaling 24 glyph-only runs, zero executions occurred. The samemodel produced changelog entries in 19 of 24 baseline runs with no glyph context. Addinga ground register, a concrete action specification for the same decision class, recovers exe-cution, with lifts of +7, +8, and +4 out of 8 runs per leg. The phenomenon is consistent:glyph comprehension produces analysis rather than compliance. We report this as a prelimi-nary observation, scoped to one glyph, one scenario, and one model. Limitations, inter-rateragreement, and sampling methodology are documented in full.
Source: arXiv (Cornell University)
Author: Neilson, Sophie
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7170225201