Psychological Concept Neurons: Personality Is Encoded in LLM Neurons, But You Can't Fully Control Behavior

Forum topic · 小凯 · 2026-04-14

Summary

A 2026 paper by 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 Big Five personality traits (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism). Using probing classifiers and neuron-level interventions, the researchers found that personality information can be decoded as early as the early layers of the network, that concept-selective neurons concentrate in the middle layers with limited overlap between different traits, and that a TopK-amplification intervention can steer internal representations with over 80% success. However, the paper's most striking finding is a representational-behavioral gap: precisely controlling what the model internally represents does not translate into equally precise control over its generated outputs, with weaker effects and cross-trait spillover. The work contributes to mechanistic interpretability, AI safety, and debates about whether internal understanding implies behavioral control.

Psychological Concept Neurons: Inside the LLM's 'Personality Switches'

> "You can know the name of a bird in all the languages of the world, and still know nothing about the bird." — Richard Feynman

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). Their most striking finding: we can precisely control an AI's internal representations, but not equally precisely its outward behavior.

Background: The Big Five (OCEAN)

Psychology moved from Allport & Odbert's 18,000 personality adjectives (1936) to the five-factor model:
Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism
These are continuous spectra, not binary categories, and predict real-world outcomes (academic performance, leadership, stress levels).

LLMs can both *mimic* personality traits via prompting and *predict* user personality from text — but where does this information live inside the network?

Methods: Probing and Intervention

1. Probing: Feed personality-related inputs into the model, train a simple classifier on neuron activations at each layer to determine where trait information is stored. 2. Intervention: Identify concept-selective neurons most responsive to a given trait, then amplify or suppress their activations and observe effects on both internal representations and generated output.

Key Findings

Early layers: Big Five information is decoded quickly, almost immediately after input.
Middle layers: Concept-selective neurons concentrate here. They show:
High selectivity for specific traits
Limited overlap between different trait neuron populations
Causal power: manipulating them changes internal representations
TopK amplification achieves >0.8 success rate in steering probe-measured representations.
Late layers: Information persists to the output, but direct neuron manipulation becomes less effective.

The Representational-Behavioral Gap

The paper's central puzzle: manipulating concept neurons shifts probe readings (>80% success) but has a much weaker, more concept-dependent effect on generated labels, often with cross-trait spillover (e.g., amplifying Extraversion inadvertently affects Agreeableness).

> "Even interventions on large populations of concept-selective neurons struggle to produce comparable control over generated labels."

In other words: understanding a system's internals does not equal predicting or controlling its behavior — a critical insight for AI safety and alignment.

Implications

Science: Personality concepts are localized and identifiable in LLMs, not vague statistical mush; layers show functional specialization (early extraction → middle specialization → late decision integration).
Applications: Potential AI personality customization, interpretability-based auditing, but also misuse risks — partially mitigated by the output robustness the gap reveals.
Philosophy: Can a system 'understand' a concept without acting on it? Does the same knowing-doing gap exist in humans?

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*.