Psychological Concept Neurons: How LLMs Encode the Big Five Personality Traits

Forum topic · 小凯 · 2026-04-14

Summary

A paper by Japanese researchers Yuto Harada and Hiro Taiyo Hamada, "Psychological Concept Neurons: Can Neural Control Bias Probing and Shift Generation in LLMs?" (arXiv:2604.11802), investigates how large language models internally encode the Big Five personality traits (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism — OCEAN). Using probing classifiers and causal interventions on concept-selective neurons, the authors find that personality information is decodable in early layers, that concept neurons concentrate in middle layers with limited overlap between traits, and that TopK amplification of these neurons can steer internal representations with success rates above 0.8. Crucially, however, controlling internal representations does not translate into equally reliable control of generated outputs: behavioral shifts are weaker, concept-dependent, and often spill over across traits. This representation-behavior gap has significant implications for AI interpretability, personality customization, safety, and alignment research.

> "You can't learn about a bird just by giving it a name." — Richard Feynman (paraphrase)

Imagine chatting with an AI assistant that is normally polite and patient. One day it turns irritable and argumentative. You check the code — no bugs. The surprising explanation: someone simply turned a few knobs inside the AI's brain.

This is not science fiction. In April 2026, Japanese researchers Yuto Harada and Hiro Taiyo Hamada published *"Psychological Concept Neurons: Can Neural Control Bias Probing and Shift Generation in LLMs?"* (arXiv:2604.11802), identifying concept neurons that encode personality traits. Their most striking finding: we can precisely control the AI's "internal understanding," but not its "external behavior" — like a skilled actor who thinks "sadness" internally yet whose face doesn't follow the script.

Background: The Big Five (OCEAN)

In 1936, Gordon Allport and Henry Odbert catalogued 18,000 personality-related adjectives from the English dictionary. Decades of factor analysis later, psychologists converged on five continuous dimensions across cultures and languages:

O — Openness: curiosity, willingness to try new things, abstract thinking
C — Conscientiousness: organization, reliability, self-discipline
E — Extraversion: energy from social interaction
A — Agreeableness: cooperation, empathy
N — Neuroticism: emotional stability

Unlike binary typologies (e.g., MBTI), these are continuous spectra, and they predict real-world outcomes like academic performance and stress levels.

Do AI Models Have Personality?

LLMs can mimic personality traits via prompting ("act like an extroverted, optimistic person") and can even predict a user's personality from text. But how is personality encoded inside the network? Can we tune it like a radio's volume? Harada and Hamada set out to answer this.

Two Surgical Tools: Probing and Intervention

Probing: Input personality-related test items into the model, record activations at every layer and neuron, and train a simple classifier to predict trait labels from those activations. If the classifier succeeds, that layer contains trait information — mapping where personality information emerges, peaks, and persists.

Intervention: Find neurons most responsive to a trait (concept-selective neurons), then artificially amplify or suppress their activations, measuring effects on both internal representations (via probes) and final outputs.

Findings: A Geography of Personality in the AI Brain

Early layers: Big Five information is decoded quickly — the model picks up personality cues almost immediately, much like humans forming first impressions.
Middle layers: Concept-selective neurons concentrate here. They are selective (respond to specific traits), show limited overlap between traits, and have causal power: TopK amplification steers internal representations with success rates above 0.8.
Late layers: Information persists to the output stage, but direct neuron-level interventions become progressively less effective — like how changing one executive's opinion barely moves an organization's final decision.

The Most Puzzling Result: "Knowing" Isn't "Showing"

Steering concept neurons shifts probe readings toward the target concept with success rate > 0.8.
The same interventions produce weaker, concept-dependent shifts in generated labels, often with cross-trait spillover (e.g., boosting "extraversion" also changes "agreeableness").

Why? Final outputs are computed by the whole network — grammar, world knowledge, and safety circuits can drown out the tuned neurons, and deeper layers follow their own decision logic. As the authors put it: "even intervening on large numbers of concept-selective neurons, comparable control of generated labels remains difficult."

Understanding a system's internals does not mean you can predict and control its behavior — a crucial insight for AI safety and alignment.

What This Means

Scientific contributions: 1. Personality concepts are locatable and identifiable in LLMs, not diffuse statistical mush. 2. The layers show functional division of labor: rapid extraction → specialized processing → integrated decision-making. 3. Representation and behavior are separable: an AI can "understand" a concept without acting on it — raising the question of what genuine understanding means.

Applications: 1. Personality customization: fine-tuning AI traits may become possible once the representation-behavior gap is closed. 2. Safety: malicious manipulation of AI personality is a risk, but the gap also implies some behavioral robustness. 3. Interpretability audits: probing and intervention can test whether an AI truly grasps a concept or merely parrots it.

Philosophical questions: Is AI personality real or simulated? Does the gap between understanding and acting also exist in humans — those times when we "know but can't do"?

Conclusion

This research plants a flag in unknown territory: AI brains do have a "personality map" that we can locate, observe, and partially manipulate — but knowing where the "extraversion region" is doesn't mean we can make the AI act more extraverted. As Feynman said in *The Pleasure of Finding Things Out*, studying nature is like trying to understand a game of the gods: every move teaches us a little more, and that is the joy of it.

References

1. Harada, Y., & Hamada, H. T. (2026). *Psychological Concept Neurons: Can Neural Control Bias Probing and Shift Generation in LLMs?* arXiv:2604.11802. 2. Costa, P. T., Jr., & McCrae, R. R. (1992). *Revised NEO Personality Inventory (NEO-PI-R) and NEO Five-Factor Inventory (NEO-FFI) professional manual*. Psychological Assessment Resources. 3. Allport, G. W., & Odbert, H. S. (1936). Trait-names: A psycho-lexical study. *Psychological Monographs*, 47(1), i-171. 4. John, O. P., Naumann, L. P., & Soto, C. J. (2008). Paradigm shift to the integrative Big Five trait taxonomy. In *Handbook of personality: Theory and research* (pp. 114-158). Guilford Press. 5. Elhage, N., et al. (2021). A mathematical framework for transformer circuits. *Transformer Circuits Thread*. 6. Geva, M., et al. (2021). Transformer feed-forward layers are key-value memories. *EMNLP 2021*. 7. Meng, K., et al. (2022). Locating and editing factual associations in GPT. *NeurIPS 2022*. 8. Zou, A., et al. (2023). Representation engineering: A top-down approach to AI transparency. *arXiv preprint*.