acceptodds
Under review as a conference paper at ICLR 2027

One Neuron, Multiple Traits: Activation-Interval Neuronal Gating for Controllable Personality in LLMs

Abstract

Controlling personality expression in Large Language Models (LLMs) supports personalized dialogue with specified behavioral styles. Existing activation-based personality control methods mainly modify entire activation vectors or selected neurons, but they will inevitably undermine general ability and non-target traits due to the polysemy and overlap of neuronal logits. To study this problem, we first conduct an empirical experiment showing that distinct personality traits and general reasoning tasks share individual neurons but occupy distinct activation intervals. Building on this observation, we propose an Activation-Interval Neuronal Gating (AING) method for personality control. This method aims to identify the activation intervals within each neuron without impairing general capabilities or other traits. Specifically, AING first selects neurons by comparing paired high- and low-trait responses of LLMs by the BIG5-CHAT dialogue dataset. Then we localize trait attribution to activation intervals within each selected neuron. These intervals define which activation values receive an intervention. During inference, AING applies adaptive scaling only within the target activation interval, with the update direction and strength determined by paired high-low differences. Extensive experiments across two model families demonstrate that AING achieves precise personality steering with high target alignment, while demonstrating remarkable selectivity by minimizing collateral shifts in non-target traits. Furthermore, our approach ensures the preservation of the model’s core functional integrity, maintaining robust performance across 6 diverse general-capability benchmarks.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Cognitive Computing, Personality traits, LLM Editing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jPyUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Activation-Interval Neuronal Gating (AING), a training-free method for steering Big Five personality in LLMs. The motivation is that personality-related neurons are polysemantic, but different traits and general reasoning tasks occupy partially distinct activation ranges within the same neuron (Fig. 2, KDE on Llama-3.1-8B-Instruct). AING has three steps. (1) It selects the top 0.10% neurons by the mean activation difference δn\delta_n between paired high/low BIG5-CHAT responses and filters out neurons that are highly active on a general corpus. (2) For each selected neuron, it fits Gaussians to target and non-target activations and picks an interval [lnζ,unζ][l_n^\zeta, u_n^\zeta] that maximizes target coverage minus λ\lambda times non-target overlap, subject to coverage ≥ρ\ge \rho (Eq. 6). (3) At inference, it scales the activation by (1±αwn)(1 \pm \alpha w_n) only when the runtime activation falls inside that interval (Eq. 7). Experiments on Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct compare against prompting, SFT/DPO-style alignment, and activation-steering baselines. AING reports Avg.P of 4.63/4.69 with Avg.G of 69.52/72.77 (base: 69.83/72.90) and the lowest average non-target drift (Avg.Dnon 0.068/0.095). Ablations remove each component.

Critical strengths
  1. The idea is simple and well matched to the problem. Restricting a neuron-level intervention to an activation interval is a natural way to reduce collateral effects from polysemantic neurons. The ablation supports this: removing the interval gate increases Avg.Dnon from 0.068 to 0.123 (Llama) and from 0.095 to 0.158 (Qwen), with a drop in Avg.G (Table 4).
  2. The evaluation includes cross-trait drift, not only target strength. Reporting Dnon and Dmax with the full five-trait profile for every trait-pole setting (Table A1) is useful, and many steering papers skip it. The baseline set is also broad, covering prompting, fine-tuning, and steering methods on two model families.
Critical weaknesses
  1. The empirical gains over strong baselines are small and not tested for significance. On Llama, AING's Avg.P (4.63) is below DPO (4.64) and close to POGP-DPO (4.58) and PERSONA-FLOW (4.59). Its Avg.G (69.52) is within 0.2 of PERSONA-FLOW (69.46), Spherical Steering (69.39), and CAA (69.34). On Qwen, the Avg.G gap to PERSONA-FLOW is 0.03. On cross-trait drift, NPTI on Llama (Avg.Dnon 0.073) is essentially tied with AING (0.068), and in Table A1 NPTI has lower Dnon than AING in several settings (e.g., O-Low, C-High, N-Low). All results appear to be single-run with greedy decoding, with no variance or confidence intervals. With differences this small on a 1–5 scale, it is hard to say AING is better than the strongest baselines. In addition, Table 2 compares capability changes only against DPN-LE, which is the weakest capability-preserving baseline in Table 1. The comparison that matters is against CAA or PERSONA-FLOW. The text also uses wording such as "catastrophic forgetting" for SFT, whose Avg.G drop is about 3 points; this overstates the evidence.

  2. The method and evaluation are underspecified in ways that affect soundness.

    • Personality metric. App. A.1 says the score is on a 1–5 Likert scale but does not say who or what produces it (LLM judge, questionnaire, or classifier), which prompts are used, or how high- and low-pole scores are combined into Avg.P. BIG5-CHAT is used for both calibration and evaluation, and no split is described, so overlap between calibration and test data cannot be ruled out.
    • Gating logic. The interval is estimated from activations on responses that already express the target pole, and Eq. 7 only intervenes when the runtime activation is inside that interval. This seems to amplify the neuron only when it is already in the target state. It is unclear how this moves a neutral model toward the target pole, and in particular how it supports low-pole steering. Eq. 7 also combines an explicit ±\pm with a signed wn∈[−1,1]w_n \in [-1,1], so the effective direction is ambiguous.
    • Formulation details. Eq. 6 uses PnbgP_n^{bg}, but Eq. 5 defines PnβP_n^\beta. The value of λ\lambda and the procedure for solving Eq. 6 are not given. Sec. 4.2 says ρ\rho varies across backbones and traits, while App. A.2 says all hyperparameters are identical across traits and models. The "neuron" (MLP intermediate vs. residual dimension, which layers) and the filter quantity unu_n (described as both an activation value and an activation frequency) are not clearly defined.
    • Motivating evidence. The separability claim rests on a single representative neuron (Fig. 2). Large-scale MLP activations are often heavy-tailed rather than Gaussian. A population-level statistic, such as the fraction of selected neurons with a well-separated interval or the distribution of overlap, would be needed to support the Gaussian model and the general claim.

    Given these gaps, I could not fully reconstruct or verify the method from the paper. Novelty is also incremental. Range-restricted neuron intervention was introduced by NeuronLens (Haider et al., 2026), and candidate selection follows NPTI. The contribution is mainly applying the two together to personality steering.

Questions for authors

Items 1–3 could change my assessment. Items 4–6 are clarifications.

  1. How is the personality score computed (judge model, prompt, scoring rubric)? How are BIG5-CHAT calibration and evaluation data separated? How are low-pole results mapped into Avg.P?
  2. Please report variance over seeds or prompts for Avg.P, Avg.G, and Avg.Dnon, plus a significance test against PERSONA-FLOW, CAA, and NPTI. Please also add NPTI, C4, and C5 to Table 1, and CAA or PERSONA-FLOW to Table 2.
  3. At inference on a neutral prompt, how often does a selected neuron's activation fall inside the target interval? How does gating on the target interval steer the model toward a pole it does not yet express? Please clarify the sign convention in Eq. 7.
  4. What is λ\lambda, how is Eq. 6 solved, and is ρ\rho fixed or per-trait?
  5. Which layers and activations are the "neurons"? Can you provide population-level evidence that activations are approximately Gaussian and that target and non-target intervals are separable for most selected neurons?
  6. Do the personality changes hold in open-ended multi-turn dialogue, and is fluency affected?
Reviewer_6hrUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how different personality traits and general capabilities share neurons while exhibiting distinct but partially overlapping activation distributions. The authors propose Activation-Interval Neuronal Gating (AING), which identifies personality-related neurons using activation differences between high- and low-trait responses and localizes trait-associated activation intervals within them. During inference, AING applies adaptive scaling only when neuron activations fall within the selected intervals. Experiments on Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct show that AING achieves strong personality control while reducing interference with non-target traits and largely preserving general capabilities across six benchmarks.

Critical strengths
  1. Meaningful research problem. The paper addresses the important challenge of controlling target personality traits while minimizing interference with non-target traits and general capabilities. By exploring activation-interval-based interventions for personality control, the work provides a meaningful application of fine-grained activation steering to a setting where different personality traits may share neuronal representations.

  2. Strong empirical results and comprehensive evaluation. AING achieves competitive personality control while largely preserving general capabilities across six benchmarks on two different model families. The paper also evaluates cross-trait interference and provides ablation studies demonstrating the effectiveness of key components, particularly activation-interval gating.

Critical weaknesses
  1. Limited methodological novelty and missing comparison with closely related work. The core idea of AING, which localizes concept-specific activation intervals within neurons and restricts interventions to these intervals, has already been explored in NeuronLens (Haider et al., 2026). Although AING introduces personality-specific interval optimization and adaptive scaling, its methodological contribution appears relatively incremental. Moreover, the lack of a direct comparison with NeuronLens makes it difficult to assess the benefits of these additional designs.

  2. Insufficient validation of activation-interval specificity. The paper argues that different personality traits and general capabilities occupy distinct activation intervals within shared neurons, but this observation is primarily supported by visualizations of representative neurons. More systematic analyses across neurons and layers would strengthen this claim. Furthermore, while the ablation studies demonstrate the effectiveness of interval gating, it remains unclear whether the improvements arise from identifying trait-specific activation intervals or simply from reducing the extent of intervention.

Questions for authors
  1. Comparison with NeuronLens. Could the authors clarify the key methodological differences between AING and NeuronLens (Haider et al., 2026)? In particular, could the authors provide a direct comparison with a NeuronLens-style interval selection baseline under the same personality control setting to demonstrate the benefits of AING's interval optimization and adaptive scaling?

  2. Analysis of activation-interval separation. Figure 2 illustrates activation distributions for representative neurons, but it remains unclear how consistently different personality traits and general capabilities occupy distinct activation intervals. Could the authors provide quantitative results or a summary table showing the degree of activation-interval separation across different neurons and layers?

  3. Effectiveness of trait-specific interval localization. Could the authors compare AING with a random-interval baseline matched for intervention frequency? This would help clarify whether the improvements arise from identifying trait-specific activation intervals or simply from reducing the frequency of interventions.

Reviewer_1uqiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Activation-Interval Neuronal Gating (AING) for controlling Big Five personality in LLMs at inference time. The method selects neurons by contrasting high- and low-trait responses, finds an activation interval for the target trait inside each neuron, and scales the activation only when it falls inside that interval. On Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct, AING gets high target-trait scores while keeping most of the base model’s accuracy on six general benchmarks.

Critical strengths
  1. The evaluation covers the right three things: target-trait strength, drift on the other traits, and general capability.
  2. The method is easy to follow, model weights stay frozen, and the ablations are informative. Removing the interval gate clearly hurts both capability and trait isolation.
  3. Compared with whole-neuron editing such as DPN-LE, AING causes much smaller drops on GSM8K and SVAMP.
Critical weaknesses
  1. The technical step is incremental. Selecting personality neurons is already done by NPTI and DPN-LE, and limiting an edit to an activation interval is the idea in NeuronLens. Figure 2 shows this separation on only one neuron, which is not enough to support the general claim.

  2. Important method details are missing or inconsistent. The preliminary plots use kernel density estimation, but the method fits a single Gaussian. Equation 6 depends on a trade-off weight λ that is never reported. Section 4.2 says the coverage threshold ρ changes across models and traits, while Appendix A.2 fixes ρ = 0.70 everywhere. Equation 7 uses both a ± and a signed weight, so the update direction is ambiguous.

  3. The personality score is not defined clearly enough. Appendix A.1 only says a 1–5 Likert score and does not say who assigns it. Calibration and evaluation both use BIG5-CHAT, with no clear held-out split. There is no human evaluation and no generation example. The gains over strong baselines are also very small: on Llama, AING is 4.63 / 69.52 against PERSONA-FLOW at 4.59 / 69.46, and its non-target drift (0.068) is close to NPTI (0.073). There are no error bars, only two 7–8B models, and no multi-trait control.

Questions for authors

No

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.