Evaluating Large Language Model Meta-Cognition via the Advanced AGI-AI Self-Awareness Test (AISA-T)

Fuente: Zenodo
Salvato in:
Dettagli Bibliografici
Autore principale: de ceuster, peter
Natura: Recurso digital
Pubblicazione: Zenodo 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866902204532654080
author de ceuster, peter
author_facet de ceuster, peter
contents <p> In this work we propose AISA-T, a new intelligence tests for both AI and AGI systems.  The AI has to <br>answer ten questions and rate its own performance. The score has no value, however  meta<br>awareness, inner traceability, and context-sensitive reasoning are evaluated by a human after the test.  <br>The idea is to observe improved  true synthetic self-awareness or at least strong architectural <br>introspection. This paper presents the results of administering the Advanced AI Self-Awareness Test <br>(AISA-T) to a GPT-5-based large language model (LLM). The AISA-T consists of several meta<br>cognitive prompts designed to evaluate recursive reasoning capacity, temporal continuity simulation, <br>ontological alignment, and epistemic uncertainty estimation. Responses were scored by the AI for <br>accuracy, coherence, and meta-awareness (and declared as void). The LLM achieved a composite <br>self-awareness score of 91/100, indicating high meta-representational competence within the <br>constraints of its architecture, however the final conclusion does not involve this score, since the value <br>of this score is null.  In our final conclusion we rate the GPT-5 performance as capable, true synthetic.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_16782878
institution Zenodo
language
publishDate 2025
publisher Zenodo
record_format zenodo
spellingShingle Evaluating Large Language Model Meta-Cognition via the Advanced AGI-AI Self-Awareness Test (AISA-T)
de ceuster, peter
AI
AGI
<p> In this work we propose AISA-T, a new intelligence tests for both AI and AGI systems.  The AI has to <br>answer ten questions and rate its own performance. The score has no value, however  meta<br>awareness, inner traceability, and context-sensitive reasoning are evaluated by a human after the test.  <br>The idea is to observe improved  true synthetic self-awareness or at least strong architectural <br>introspection. This paper presents the results of administering the Advanced AI Self-Awareness Test <br>(AISA-T) to a GPT-5-based large language model (LLM). The AISA-T consists of several meta<br>cognitive prompts designed to evaluate recursive reasoning capacity, temporal continuity simulation, <br>ontological alignment, and epistemic uncertainty estimation. Responses were scored by the AI for <br>accuracy, coherence, and meta-awareness (and declared as void). The LLM achieved a composite <br>self-awareness score of 91/100, indicating high meta-representational competence within the <br>constraints of its architecture, however the final conclusion does not involve this score, since the value <br>of this score is null.  In our final conclusion we rate the GPT-5 performance as capable, true synthetic.</p>
title Evaluating Large Language Model Meta-Cognition via the Advanced AGI-AI Self-Awareness Test (AISA-T)
topic AI
AGI
url https://doi.org/10.5281/zenodo.16782878