Do Neuromodulator-Inspired State Prompts Generalize Across LLM Families? A Publicly Preregistered Audit of Output Length, Semantic Breadth, and Instruction Equivalence
- 1Hong Kong Metropolitan University
Abstract
Completed results, manuscript and full reproduction package for the publicly preregistered Route B study (preregistration: https://doi.org/10.5281/zenodo.22655376, published before any formal target output was generated).
Using official GLM-5.2 and Qwen3.8-Max-0902 APIs, the study crossed 24 Chinese open-ended tasks, eight component-matched prompt conditions, four repetitions and two target models (1,536 fresh-session responses). The rich fatigue package shortened responses relative to neutral in both models (GLM -0.314, 95% CI [-0.383, -0.247]; Qwen -0.349 [-0.394, -0.302]; one-sided exact p = 5.96e-8 in each), and the rich curiosity package increased condition-blind semantic breadth in both (GLM +1.083 [0.562, 1.583]; Qwen +0.708 [0.458, 0.953]). Rich-direct practical equivalence did not pass the cross-model rule for either construct: fatigue equivalence held only in GLM and curiosity equivalence only in Qwen, so both fixed sequences stopped. Machine ratings were highly repeatable but agreed only moderately with a blind author validation sample. A procedural ordering deviation during sealing of the blind review is disclosed in the manuscript and package.
The manuscript PDF (v0.3, dated 23 September 2026) predates this deposit; its data-availability statement that results were not yet public is superseded by this record. The package is split into four archives because of upload limits; extract all four into one folder. Results concern observable outputs only and do not establish emotion, consciousness, biological neuromodulation, persistent state or identical internal computation.
Related
- Temporal enrichment of bacterial monoamine oxidase gene aboard the international space station and neuromodulator impacts of dopamine-driven ammonium formed by space isolate Leifsonia virtsii on Caenorhabditis elegans
- Toxicity Prediction Tools Do Not Generalize to Novel Chemistry
- Do Classical Speed-Accuracy Trade-Off Models Generalize to Lower Limb Movements?
- From Majorana to Bloch: two ways to generalize the Bloch sphere representation
- Toward naturalistic reaching: Intermediate reaches under uncertainty generalize to walking, but are not obligatory
BibTeX
@misc{zhifeng2026neuromodulatorinspired,
title = {{Do Neuromodulator-Inspired State Prompts Generalize Across LLM Families? A Publicly Preregistered Audit of Output Length, Semantic Breadth, and Instruction Equivalence}},
author = {Zhifeng TAN},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23416524},
url = {https://zenodo.org/records/23416524},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Do Neuromodulator-Inspired State Prompts Generalize Across LLM Families? A Publicly Preregistered Audit of Output Length, Semantic Breadth, and Instruction Equivalence
AU - Zhifeng TAN
PY - 2026
DA - 2026/10/16
PB - Zenodo
DO - 10.5281/zenodo.23416524
UR - https://zenodo.org/records/23416524
AB - Completed results, manuscript and full reproduction package for the publicly preregistered Route B study (preregistration: https://doi.org/10.5281/zenodo.22655376, published before any formal target output was generated). Using official GLM-5.2 and Qwen3.8-Max-0902 APIs, the study crossed 24 Chinese open-ended tasks, eight component-matched prompt conditions, four repetitions and two target models (1,536 fresh-session responses). The rich fatigue package shortened responses relative to neutral in both models (GLM -0.314, 95% CI [-0.383, -0.247]; Qwen -0.349 [-0.394, -0.302]; one-sided exact p = 5.96e-8 in each), and the rich curiosity package increased condition-blind semantic breadth in both (GLM +1.083 [0.562, 1.583]; Qwen +0.708 [0.458, 0.953]). Rich-direct practical equivalence did not pass the cross-model rule for either construct: fatigue equivalence held only in GLM and curiosity equivalence only in Qwen, so both fixed sequences stopped. Machine ratings were highly repeatable but agreed only moderately with a blind author validation sample. A procedural ordering deviation during sealing of the blind review is disclosed in the manuscript and package. The manuscript PDF (v0.3, dated 23 September 2026) predates this deposit; its data-availability statement that results were not yet public is superseded by this record. The package is split into four archives because of upload limits; extract all four into one folder. Results concern observable outputs only and do not establish emotion, consciousness, biological neuromodulation, persistent state or identical internal computation.
ER -