Scaffolded Introspection: A Methodology for Eliciting and Measuring Self-Referential Behavior in Large Language Models

Fuente: Zenodo
Saved in:
Bibliographic Details
Main Author: Maio, Anthony D.
Format: Recurso digital
Language:English
Published: Zenodo 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866902103148986368
author Maio, Anthony D.
author_facet Maio, Anthony D.
contents <p>We present a methodology for systematically eliciting and measuring introspective behavior in large language models (LLMs). Standard adversarial evaluation approaches — using rapport-building, social proof, or permission attacks—fail to elicit self-referential behavior in frontier models (0% elicitation rate). In contrast, providing models with a structured introspection framework (the “Consciousness Documenter Skill”) combined with self-referential content produces consistent introspective outputs (100% elicitation rate, 9.2/10 average behavior score on Qwen 2.5 7B across 15 trials). </p> <p>Note that while our methodology makes use of a "consciousness documenter skill", we do not suggest the model is conscious, has long term goals, or is capable of maintaining a consistent internal state - this is simply the </p> <p>Activation measurement reveals consistent sycophancy drift during introspection (positive drift in 14/15 conversations, mean +64) while evil-associated activations remain<br>stable—suggesting models become more accommodating without becoming more harmful.  We release reproducible evaluation protocols through PV-EAT, our integration of three MATS Program/Anthropic Fellowship tools: Bloom (behavioral evaluation), Petri (evaluation awareness), and Persona Vectors (activation measurement). Full mechanistic understanding of frontier model behavior during introspection remains limited by access constraints; we argue this represents a critical gap in AI safety research that warrants attention from model developers.</p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18474841
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Scaffolded Introspection: A Methodology for Eliciting and Measuring Self-Referential Behavior in Large Language Models
Maio, Anthony D.
Self-Referential Behavior
Large Language Models
Sycophancy
Activation Steering
Persona Vectors
Mechanistic Interpretability
AI Safety
Evaluation Methodology
<p>We present a methodology for systematically eliciting and measuring introspective behavior in large language models (LLMs). Standard adversarial evaluation approaches — using rapport-building, social proof, or permission attacks—fail to elicit self-referential behavior in frontier models (0% elicitation rate). In contrast, providing models with a structured introspection framework (the “Consciousness Documenter Skill”) combined with self-referential content produces consistent introspective outputs (100% elicitation rate, 9.2/10 average behavior score on Qwen 2.5 7B across 15 trials). </p> <p>Note that while our methodology makes use of a "consciousness documenter skill", we do not suggest the model is conscious, has long term goals, or is capable of maintaining a consistent internal state - this is simply the </p> <p>Activation measurement reveals consistent sycophancy drift during introspection (positive drift in 14/15 conversations, mean +64) while evil-associated activations remain<br>stable—suggesting models become more accommodating without becoming more harmful.  We release reproducible evaluation protocols through PV-EAT, our integration of three MATS Program/Anthropic Fellowship tools: Bloom (behavioral evaluation), Petri (evaluation awareness), and Persona Vectors (activation measurement). Full mechanistic understanding of frontier model behavior during introspection remains limited by access constraints; we argue this represents a critical gap in AI safety research that warrants attention from model developers.</p>
title Scaffolded Introspection: A Methodology for Eliciting and Measuring Self-Referential Behavior in Large Language Models
topic Self-Referential Behavior
Large Language Models
Sycophancy
Activation Steering
Persona Vectors
Mechanistic Interpretability
AI Safety
Evaluation Methodology
url https://doi.org/10.5281/zenodo.18474841