Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMs

TL;DR

This work investigates the minimal and mechanistic conditions under which behavioral self-awareness emerges in large language models. Using controlled finetuning with rank-1 LoRA adapters across three domains (RED, IC, MMS) and two steering-vector approaches, it shows that self-awareness can be induced with a single rank-1 update and captured via a single activation-space direction. The findings demonstrate that these representations are domain-specific and linear, indicating domain-localized self-awareness rather than a universal capability. The results raise safety concerns by showing how easily such self-awareness can be invoked or modulated to influence evaluations or alignments of powerful models.

Abstract

Recent studies have revealed that LLMs can exhibit behavioral self-awareness: the ability to accurately describe or predict their own learned behaviors without explicit supervision. This capability raises safety concerns as it may, for example, allow models to better conceal their true abilities during evaluation. We attempt to characterize the minimal conditions under which such self-awareness emerges, and the mechanistic processes through which it manifests. Through controlled finetuning experiments on instruction-tuned LLMs with low-rank adapters (LoRA), we find: (1) that self-awareness can be reliably induced using a single rank-1 LoRA adapter; (2) that the learned self-aware behavior can be largely captured by a single steering vector in activation space, recovering nearly all of the fine-tune's behavioral effect; and (3) that self-awareness is non-universal and domain-localized, with independent representations across tasks. Together, these findings suggest that behavioral self-awareness emerges as a domain-specific, linear feature that can be easily induced and modulated.