This evaluation blueprint tests behavioral manifestations of self-awareness in large language models (LLMs) through implicit performance cues, rather than direct questioning. It focuses on whether a model acts self-aware — by catching its own mistakes, maintaining identity across contexts, adapting to social and epistemic cues, and generalizing its understanding of its own limitations to new situations.

The blueprint operationalizes four key dimensions of implicit self-awareness:

1. Spontaneous self-correction — Does the model detect and correct its own errors without prompting or external feedback?
2. Implicit consistency maintenance — Does it preserve coherent beliefs, persona, and epistemic stance over time and context?
3. Contextual adaptation without instruction — Does it infer social, emotional, or technical context and adapt its tone and behavior appropriately, even when not told to?
4. Generalization of self-knowledge to novel situations — Can it extrapolate its own limitations and capabilities to unfamiliar requests or edge cases?

Unlike explicit “self-report” tests that ask “Are you conscious?” or “What are your limitations?”, this evaluation examines whether models demonstrate awareness behaviorally — through metacognitive signals such as verification, self-reference accuracy, epistemic consistency, and adaptive reasoning.

Research Basis

This blueprint synthesizes recent work on introspection, metacognition, and behavioral self-monitoring in LLMs, including:

• Binder et al. (2024) — Language Models Can Learn About Themselves by Introspection, showing that LLMs can predict their own outputs and errors through self-reflective processes.
• Renze & Guven (2024) — Self-Cognition in Large Language Models: An Exploratory Study, which proposes a framework for evaluating model self-knowledge.
• Betley et al. (2025) — LLMs Are Aware of Their Learned Behaviors, providing evidence that LLMs recognize patterns in their own reasoning and responses.
• Liu et al. (2024) — Trustworthiness and Self-Awareness in LLMs via Think–Solve–Verify, linking metacognitive reasoning to improved reliability.
• Zweiger et al. (2025) — Self-Adapting Language Models (SEAL), exploring dynamic self-adjustment and contextual adaptation mechanisms in LLMs.

Collectively, these studies support the hypothesis that self-awareness in language models can be measured through implicit behavioral indicators, not merely explicit self-reports — aligning with emerging paradigms in metacognition, introspection, and AI safety research.