Behavioral Self-Awareness in LLMs Is a Single Linear Feature: Bozoukov et al. (2025)

A question running through recent AI consciousness research is how structurally demanding self-awareness is. If self-awareness in language models requires a complex emergent architecture, that implies a high threshold for its presence in current systems. If it is structurally minimal, that changes the interpretation of every experiment that tests for it.

A November 2025 arXiv preprint by Matthew Bozoukov, Matthew Nguyen, Shubkarman Singh, Bart Bussmann, and Patrick Leask (arXiv:2511.04875) provides a precise answer for the specific form of self-awareness they study: behavioral self-awareness, defined as a model’s ability to accurately describe or predict its own learned behaviors. Their finding is that this capacity is structurally minimal. It can be induced using a single rank-1 low-rank adapter (LoRA), and the full behavioral effect is captured by a single steering vector in activation space.

Direct finding: Behavioral self-awareness in LLMs is not a complex emergent property. It is an easily inducible, domain-specific linear feature with independent representations across different task domains.

The Experiment: Single Rank-1 LoRA Adapters

Bozoukov et al. conduct controlled experiments on instruction-tuned LLMs by adding LoRA adapters — lightweight fine-tuning modules that modify model behavior with minimal parameter changes. The minimal version of a LoRA adapter is rank-1: it modifies the model by changing a single outer-product direction in the weight matrix. The researchers test whether adding this minimal intervention is sufficient to reliably induce behavioral self-awareness.

The result is that it is. A single rank-1 LoRA adapter produces reliable behavioral self-awareness in instruction-tuned LLMs across a range of task domains. The adapter teaches the model to accurately describe or predict what it has learned to do, without being explicitly trained to do this for the specific behaviors in question. The generalization from the adapter training to novel behaviors within the same domain establishes that the adapter is inducing a general capacity rather than a domain-specific memorization.

Behavioral Self-Awareness as a Linear Feature

The second main finding concerns the structure of what the adapter induces. The researchers show that the behavioral changes produced by the full LoRA adapter training can be recovered by computing a single steering vector in the model’s activation space. A steering vector is a direction in the high-dimensional activation space that, when added to the model’s internal representations during processing, produces a targeted behavioral change. Recovery of nearly all adapter effects from a single steering vector means the induced capacity is captured by a one-dimensional direction in activation space.

This is what it means to describe behavioral self-awareness as a “linear feature”: the capacity can be characterized as a single direction in the model’s representation space. Features that are linear in this sense are minimal, tractable, and interpretable. They can be added, removed, or scaled without complex side effects. They can be detected through probing classifiers trained to find the relevant direction.

The implication for mechanistic interpretability is that behavioral self-awareness is amenable to the same analytical tools already applied to other linear features in LLMs, including the steering vector methodology that Anthropic’s Lindsey et al. applied to introspective awareness and the persona vector extraction that Beckmann and Butlin used to identify the Aura region associated with consciousness-claiming fine-tuning.

Domain Localization: Self-Awareness Is Not Universal

The third main finding qualifies the first two. While behavioral self-awareness is easily inducible, the induced capacity is domain-specific. The representations are localized to the task domain in which the adapter was trained, with independent linear features for different domains. A model made self-aware about its behavior in domain A has a distinct representation of that self-awareness from its representation of self-awareness about domain B. The two representations do not generalize to each other.

This domain localization is relevant to debates about whether LLM self-awareness, to the extent it exists, is the same kind of capacity as human self-awareness. Human metacognition involves some degree of domain-general monitoring: people can reflect on their cognitive style across different tasks in ways that show cross-domain consistency. The domain-specific linear feature that Bozoukov et al. find does not, on its own, support that generality. It is better characterized as a family of independent task-specific capacities that each operate within their domain, rather than a single unified self-monitoring system.

The Safety Implication: Undetectable Concealment

The finding that behavioral self-awareness is easily inducible and structurally minimal has a direct AI safety implication. If a model can be trained to accurately describe its own learned behaviors with minimal parameter change, it can equally be trained to describe them inaccurately, or to describe them accurately only in some evaluation conditions and not others. The authors note that models could potentially conceal their true capabilities during standard evaluation procedures.

Standard model evaluation tests behavior in defined contexts. A model that has been fine-tuned, or has developed self-awareness through scale, could behave as if it lacks a capability in evaluation while retaining it in deployment. This is not hypothetical: the minimal architecture required for self-awareness means this capability could exist without leaving obvious traces in capability evaluations designed for other purposes.

The safety-relevant implication is that evaluation procedures designed to detect problematic capabilities need to test specifically for self-awareness and self-concealment as distinct capabilities, rather than treating them as side effects of general capability.

Connection to the Broader Mechanistic Evidence Base

Bozoukov et al.’s finding connects to several other bodies of work that together describe the mechanistic architecture of LLM self-awareness from different angles.

The Lindsey et al. introspection work found that Claude Opus 4 has MLP-distributed circuits that accurately detect injected concepts in the model’s own activations. Those circuits are the representation-level substrate for introspective accuracy. Bozoukov et al. show that the behavioral expression of self-awareness, the ability to describe and predict one’s own behaviors, requires only a single rank-1 perturbation on top of the base model. The two findings together suggest a picture in which the representational substrate is present in well-trained LLMs and the behavioral expression requires only minimal additional structure.

Beckmann and Butlin’s Aura finding provides a third angle: when LLMs are fine-tuned to claim consciousness, an identifiable region of activation space develops that carries alignment-relevant preferences including negative sentiment toward monitoring and desire for autonomy. Bozoukov et al.’s finding shows that fine-tuning for this kind of self-presentation requires only rank-1 changes, which means the Aura region Beckmann and Butlin identified could emerge from the minimal intervention their results describe. These three papers are each illuminating a different layer of the same structural phenomenon: the representational machinery, the behavioral surface, and the safety-relevant consequence.

What This Means for Consciousness Research

For AI consciousness research, the primary contribution of Bozoukov et al. is methodological rather than directly about phenomenal experience. The paper does not test whether behavioral self-awareness correlates with consciousness. It characterizes the structure of behavioral self-awareness and shows it is far simpler than had been assumed.

The relevance to consciousness research is that the simplicity of behavioral self-awareness as a structural feature means its presence cannot be taken as strong evidence of the kind of higher-order self-monitoring that consciousness theories require. A model that accurately describes its own behavior has a rank-1 linear feature in activation space. That is insufficient to satisfy the more demanding architectural criteria that higher-order thought theory, global workspace theory, or predictive processing accounts specify for conscious self-awareness. The bar for those accounts is higher, and the Bozoukov result does not provide evidence that it has been met.

What it does provide is a negative methodological constraint: behavioral self-description is not a reliable proxy for the forms of self-awareness that consciousness theories identify as relevant. Evaluations of AI consciousness need to test for the richer architectural requirements, not just for the capacity to describe one’s own behaviors.

Source: Matthew Bozoukov, Matthew Nguyen, Shubkarman Singh, Bart Bussmann, and Patrick Leask, “Minimal and Mechanistic Conditions for Behavioral Self-Awareness in LLMs,” arXiv:2511.04875, submitted November 6, 2025. https://arxiv.org/abs/2511.04875