The Localization of Personality in Neural Space: Why Controlling Representations Does Not Fully Control Behavior

Introduction

Large language models now routinely exhibit behaviors that appear to reflect stable individual differences. When prompted appropriately, they can simulate distinct conversational styles, value orientations, and interpersonal tendencies that psychologists would recognize as personality. The Big Five framework, operationalized through instruments such as the Big Five Inventory 2 (BFI-2), provides a structured vocabulary for describing these variations along five dimensions: Extraversion, Agreeableness, Conscientiousness, Negative Emotionality, and Open-Mindedness. While it is now well established that LLMs can imitate specific personality profiles or infer such traits from human text, the mechanistic basis of these capabilities remains opaque. In their paper, "Psychological Concept Neurons: Can Neural Control Bias Probing and Shift Generation in LLMs?", Yuto Harada and Hiro Taiyo Hamada move beyond behavioral observation to examine where these psychological constructs are encoded, how they are organized at the neuronal level, and whether they can be causally manipulated. Their findings reveal a striking dissociation: while internal representations of personality can be localized and strongly biased through targeted intervention, the resulting control over generated text remains substantially weaker and less specific.

Where Personality Lives: Layer-wise Dynamics and Concept Neurons

To map the emergence of psychological constructs, Harada and Hamada employ a three-stage methodology combining probing, neuron identification, and causal intervention. First, they use linear probes to track where Big Five domain labels become decodable across network depth. Contrary to some prior work suggesting that persona-related representations emerge primarily in deeper layers, they find that Big Five information becomes rapidly decodable in early layers and remains detectable through the final layers. This persistence suggests that personality-related information is not merely a transient byproduct of initial processing but is maintained as a stable, accessible feature of the model's internal state throughout computation.

The second stage identifies psychological concept neurons by measuring unit-level selectivity for each Big Five domain. Using the 60 BFI-2 questionnaire items as stimuli, the authors extract individual neurons that respond selectively to descriptions of a given trait. These concept-selective neurons are most prevalent in mid layers, exhibiting limited overlap across domains. This localization pattern implies a degree of functional specialization; rather than being diffusely distributed or fully entangled, personality representations appear to occupy somewhat segregated subspaces in the middle of the network. The mid-layer concentration is consistent with theoretical accounts suggesting that intermediate layers serve as a bottleneck where abstract, task-relevant features are computed and compressed before being passed to later layers for output generation.

The Representation-Behavior Gap: Interventions and Their Limits

The critical test of whether these representations are merely correlated with or causally responsible for personality-related outputs comes through activation patching. The authors intervene at inference time by enhancing or suppressing the activations of identified concept-selective neurons to extreme quantiles, then measure the effects on both probe readouts and a domain name generation task. The results expose a significant gap between representational control and behavioral control.

Interventions on concept-selective neurons consistently shift probe readouts toward targeted concepts, with targeted success rates exceeding 0.8 for some traits. This indicates that the model's internal separation of personality dimensions can be strongly and specifically steered through localized manipulation. However, when the same interventions are applied during text generation, the effects are markedly different. While the interventions often bias generated label distributions in the intended directions, the success rates reach only roughly 50% and exhibit greater variability across concepts. Crucially, generation tasks show cross-trait spillover, where enhancing one trait inadvertently influences others, a phenomenon largely absent in the probe readouts.

This dissociation suggests that the mapping from internal representation to token output is not a simple linear readout. The probes, which operate on hidden states, capture a clean, steerable signal residing in mid-layer subspaces. The generation process, however, appears to draw on additional mechanisms beyond these localized representations. The residual stream may undergo nonlinear transformations through attention heads and layer normalization that dilute or remix the personality signal before it reaches the unembedding layer. Alternatively, generation may involve competitive dynamics between multiple candidate representations, meaning that even a strong bias in the personality subspace does not guarantee a corresponding bias in the final token distribution.

Beyond Localization: Why Behavior Resists Steering

The findings from this study carry broader implications for the field of mechanistic interpretability and AI alignment. The observation that concept-selective neurons provide strong control over representations but limited control over generation challenges a common assumption in representation engineering: that aligning internal states is sufficient for aligning outputs. If personality is encoded locally but integrated globally, then steering methods that target narrow neuronal populations may be inherently limited in their capacity to reliably shape complex behavioral outputs.

One interpretation is that personality representations in LLMs occupy a similar niche to semantic features in vision models: they are necessary but not sufficient for the final decision. Just as object parts are encoded in intermediate convolutional layers but must be composed through subsequent operations to yield a classification, personality traits may be encoded in mid-layer neurons but must be contextualized and weighted by downstream attention mechanisms to influence generation. The cross-trait spillover observed in generation but not in probes further suggests that the unembedding layer or final layer normalization may conflate dimensions that remain separable in intermediate representations.

This also raises questions about the nature of the BFI-2 constructs as they exist in LLMs. The questionnaire items provide a specific, structured operationalization of personality, and the concept neurons respond to these textual descriptions. However, natural language generation involves a broader compositional space where traits interact with syntax, pragmatics, and discourse context. The gap between probe and generation control may reflect the difference between recognizing a personality concept and enacting it behaviorally, a distinction that has clear analogues in human psychology, where implicit attitudes and explicit behaviors often diverge.

Conclusion

Harada and Hamada's work represents a significant step toward understanding the neural substrates of psychological constructs in language models. By demonstrating that Big Five traits are encoded in localized, mid-layer neuronal populations that can be selectively manipulated, they provide evidence for a structured internal organization of personality in LLMs. Yet the simultaneous demonstration that such manipulation fails to translate cleanly into controlled generation serves as a cautionary note for efforts to steer model behavior through targeted internal interventions.

Looking forward, this research opens several avenues for investigation. The precise mechanisms that mediate between mid-layer personality representations and final token selection remain unknown; identifying the attention heads or routing functions that mix or gate these signals is a natural next step. Additionally, it remains to be seen whether the representation-behavior gap generalizes to other psychological constructs, such as values, cognitive styles, or psychiatric symptom profiles. If the gap is universal, it suggests that reliable behavioral control in LLMs may require interventions at multiple processing stages, not merely the localization and amplification of concept-selective neurons. As we develop more sophisticated methods for peering inside these models, acknowledging the distinction between representational steerability and behavioral controllability will be essential for building systems whose internal states and external actions remain consistently aligned.