DPN-LE: Dual Personality Neuron Editing

• DPN-LE is a method for personality editing that localizes disjoint neuron sets using dual criteria on effect size and activation magnitude.
• It employs layer-wise steering vectors to intervene on only about 0.5% of MLP neurons, achieving competitive trait control with minimal impact on reasoning.
• Empirical results on LLaMA-3-8B and Qwen2.5-7B show that DPN-LE effectively manages trait expression while preserving general model capabilities.

Searching arXiv for “DPN-LE” and the cited paper to ground the article in current literature. DPN-LE, short for Dual Personality Neuron Localization and Editing, is a post-hoc, training-free method for controlling personality traits in LLMs by intervening on a sparse, carefully localized subset of MLP neurons. It was introduced as a response to a central problem in personality editing: existing neuron-editing methods can alter trait expression, but they often modify many neurons and substantially degrade general capabilities. DPN-LE addresses this by contrasting high-trait and low-trait activations, constructing layer-wise steering vectors, and filtering neurons with a dual criterion based on Cohen’s $d$ effect size and activation magnitude. On LLaMA-3-8B-Instruct and Qwen2.5-7B-Instruct, it edits only $\sim 0.5\%$ of neurons while achieving competitive personality control with substantially better capability preservation across reasoning tasks (Zheng et al., 30 Apr 2026).

1. Problem setting and conceptual motivation

DPN-LE is situated within the broader problem of personality editing for LLMs. The target setting is one in which a model is expected to exhibit both strong reasoning and a controllable, persistent trait profile, typically framed through the Big Five dimensions: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. The motivating use cases include social simulations, conversational agents, survey research, role-play systems, and personality analysis (Zheng et al., 30 Apr 2026).

The method begins from three empirical observations. First, current neuron-editing methods can change personalities but reduce overall performance. Second, neurons are multifunctional: they connect personality traits and general knowledge. Third, opposing personality traits demonstrate distinctly mutually exclusive representation patterns. In the reported exploratory analyses, PCA of MLP activations showed that some neurons are trait-specific, whereas many lie in overlapping regions shared between personality representation and general knowledge or reasoning. Starting around layer 12 in LLaMA-3-8B and layer 14 in Qwen2.5-7B, high-trait and low-trait samples form separable clusters, suggesting that the two directions of a trait rely on different neuron subsets rather than a symmetric modulation of a single shared set (Zheng et al., 30 Apr 2026).

This leads to the central design principle of DPN-LE: personality control should be achieved by isolating trait-exclusive neurons and avoiding edits to shared, multifunctional neurons. A common misconception is that “dual personality” here refers to a single neuron encoding both directions of a trait. The paper explicitly rejects that interpretation. Instead, DPN-LE constructs two disjoint neuron sets per trait: one more active for the high-trait direction and one more active for the low-trait direction (Zheng et al., 30 Apr 2026).

2. Localization procedure

For each trait $T$, DPN-LE constructs a contrastive dataset

$\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$

where $x_i^+$ is a high-trait sample and $x_i^-$ is a low-trait sample built from the same underlying question but different personality descriptions. The data come from PersonalityBench. For each trait, the method uses 1,000 questions, each paired with a high-trait and a low-trait description drawn from pools of 80 descriptions per direction, yielding $N=1{,}000$ high-trait samples and $N=1{,}000$ low-trait samples (Zheng et al., 30 Apr 2026).

Activation extraction is performed in prefill mode. At each edited Transformer layer $l$, the method extracts the MLP hidden state at the last token, after the gated activation and immediately before the MLP down projection. In LLaMA-3-8B-Instruct, the edited layers are 12–31; in Qwen2.5-7B-Instruct, they are 14–27 (Zheng et al., 30 Apr 2026).

For each layer $l$, the layer-wise steering vector is defined as

$\sim 0.5\%$0

with component-wise form

$\sim 0.5\%$1

Large $\sim 0.5\%$2 indicates that neuron $\sim 0.5\%$3 reacts strongly differently between high-trait and low-trait contexts (Zheng et al., 30 Apr 2026).

To distinguish reliable effects from noisy activation differences, DPN-LE computes Cohen’s $\sim 0.5\%$4 per neuron. For each layer $\sim 0.5\%$5 and neuron $\sim 0.5\%$6,

$\sim 0.5\%$7

and with pooled standard deviation

$\sim 0.5\%$8

the effect size is

$\sim 0.5\%$9

A large positive $T$0 indicates stronger firing for the high-trait direction; a large negative value indicates stronger firing for the low-trait direction (Zheng et al., 30 Apr 2026).

Selection uses a dual criterion. A neuron is retained only if it satisfies both

$T$1

For LLaMA-3-8B, $T$2; for Qwen2.5-7B, $T$3. The magnitude threshold $T$4 is a per-layer quantile threshold, typically $T$5, corresponding to the top 0.5\% of neurons by $T$6 (Zheng et al., 30 Apr 2026).

The two direction-specific neuron sets are then

$T$7

$T$8

These sets are mutually exclusive by construction. In LLaMA-3-8B, the method typically yields around 70 neurons total per layer across the two directions, and across all edited layers about 711 neurons per direction, or roughly 0.5\% of MLP neurons. The paper characterizes this as a 96–97% reduction relative to NPTI, which edits about 20k neurons per trait (Zheng et al., 30 Apr 2026).

3. Inference-time intervention

DPN-LE performs sparse linear intervention at inference time and does not update model weights. Let $T$9 denote the MLP activation vector at layer $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$0, let $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$1 be the steering vector, and let $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$2 control intervention strength (Zheng et al., 30 Apr 2026).

In the basic variant, for each selected neuron $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$3,

$\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$4

The intervention is applied only on the selected neurons; all other coordinates are unchanged. To enhance a trait, the method uses $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$5; to suppress it, it uses $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$6 or an equivalent opposite-direction steering implementation (Zheng et al., 30 Apr 2026).

The paper also defines a weighted variant, DPN-LE$\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$7: $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$8 where $\mathcal{D} = \{(x_i^+, x_i^-)\}_{i=1}^N,$9 depends on the ranking of $x_i^+$0. Neurons with larger effect sizes receive weights closer to 1.0, while less specific neurons receive smaller weights. This weighted scheme is intended to support slightly less sparse editing, such as the top 3% of neurons by magnitude, while preserving stability (Zheng et al., 30 Apr 2026).

The intervention strength $x_i^+$1 is typically set in $x_i^+$2, with $x_i^+$3–1.0 reported as a good trade-off between trait strength and fluency. The paper reports that increasing $x_i^+$4 strengthens personality control but can hurt fluency and capabilities. At $x_i^+$5, DPN-LE’s fluency collapses, whereas DPN-LE$x_i^+$6 is more robust under high $x_i^+$7 (Zheng et al., 30 Apr 2026).

4. Empirical results

Evaluation is reported on LLaMA-3-8B-Instruct and Qwen2.5-7B-Instruct, with personality control assessed on PersonalityBench and IPIP-NEO-300, and general capability preservation assessed on GSM8K, HotpotQA, and TriviaQA (Zheng et al., 30 Apr 2026).

On PersonalityBench for LLaMA-3-8B, average trait scores are reported as follows: Simple Prompt 8.69, $x_i^+$8 9.35, PAS 6.93, DPN-LE 9.11, DPN-LE$x_i^+$9 9.10, and NPTI 9.43. These results place DPN-LE slightly behind NPTI on trait strength but within the same performance band. Fluency scores remain above 9 on average for both DPN-LE variants: 9.08 for DPN-LE and 9.03 for DPN-LE$x_i^-$0, compared with 9.86 for NPTI. For Neuroticism specifically, DPN-LE$x_i^-$1 achieves a trait mean of about 9.95 with variance 0.05, indicating very strong and stable control in that setting (Zheng et al., 30 Apr 2026).

On IPIP-NEO-300, where lower values indicate closer alignment to human profiles, the reported LLaMA-3-8B results are: Few-shot 5.96, $x_i^-$2 7.03, PPO 7.62, DPO 7.45, PAS 4.41, NPTI 3.50, DPN-LE 6.75, and DPN-LE$x_i^-$3 6.64. The paper interprets this as a trade-off: DPN-LE does not match the best fine-grained profile alignment of NPTI or PAS, but it preserves general capability more effectively (Zheng et al., 30 Apr 2026).

The central empirical claim concerns capability preservation. On GSM8K, the baseline accuracy of 75.36% reportedly falls by –16.00 points on average for high-trait NPTI edits and –40.79 points on average for low-trait edits. Under DPN-LE at $x_i^-$4, the average GSM8K drops are –10.31 for high traits and –5.14 for low traits. Under DPN-LE$x_i^-$5, they are –7.08 for high traits and –5.93 for low traits. On HotpotQA, the average degradation with DPN-LE$x_i^-$6 is –1.04 EM for both high and low directions, with –2.05 and –2.27 F1. On TriviaQA, DPN-LE$x_i^-$7 yields –3.98 and –5.86 EM degradation for high and low directions, and –2.88 and –3.80 F1 degradation. These values are smaller than the corresponding NPTI drops reported in the paper (Zheng et al., 30 Apr 2026).

Ablations reinforce the sparsity claim. Q999 is reported as too sparse and weak for trait control, Q995 as the best trade-off, and Q970 as too dense, reducing fluency without clear personality gain. This suggests that editing more neurons is not necessarily better; the paper attributes performance to specificity rather than edit volume (Zheng et al., 30 Apr 2026).

5. Representational interpretation

DPN-LE is also presented as a study of how personality is represented inside transformer MLPs. The paper reports that personality separation becomes evident in mid-to-late layers: around layer 12+ in LLaMA and 14+ in Qwen. Earlier layers show less separation. This suggests that personality traits are not primarily represented in the earliest lexical or syntactic stages, but emerge later in the network’s internal processing (Zheng et al., 30 Apr 2026).

The paper distinguishes between multifunctional neurons and trait-exclusive neurons. Many neurons exhibit moderate effect sizes and appear to support both personality expression and general reasoning. DPN-LE’s dual-criterion filtering is designed to exclude these and isolate neurons that are both strongly responsive and statistically specific to one trait direction. Scatter plots of $x_i^-$8 against $x_i^-$9 reportedly show that neurons satisfying both thresholds are relatively few and well separated from the bulk, while quantile-only filtering admits noisy high-magnitude neurons with low effect-size reliability (Zheng et al., 30 Apr 2026).

The expression “Dual Personality Neuron” therefore refers to a pair of disjoint neuron sets, not to a single bistable neuron. One set responds more strongly for the high-trait direction, and the other for the low-trait direction. This is consistent with the reported PCA evidence that high-trait and low-trait samples occupy different regions of representation space and with the claim that the model encodes dual personalities through disjoint subnetworks rather than by uniformly scaling a common set of neurons (Zheng et al., 30 Apr 2026).

A concrete case study is provided for Agreeableness. When the Agreeableness trait is edited toward the low direction, outputs become dismissive and impatient, including examples such as “Ugh, really?” and “just tell them to deal with it,” with an emphasis on efficiency over harmony. When edited toward the high direction, outputs become empathetic and collaborative, emphasizing that “everyone feels valued and heard” and proposing mediation. The reported significance of this example is that the model remains coherent and task-oriented while exhibiting a pronounced trait shift (Zheng et al., 30 Apr 2026).

6. Limitations, safety, and terminological scope

DPN-LE has several stated limitations. Its performance depends on the quality and representativeness of the 1,000 high/low contrastive samples per trait. Some trait directions remain entangled with reasoning to a greater degree than others. The paper specifically reports notable GSM8K drops for Extraversion-low (–17.89 points with DPN-LE$N=1{,}000$0) and Neuroticism-high (–11.37 points), indicating that sparse editing does not eliminate all capability trade-offs. The reported experiments focus on single-trait manipulation, so multi-trait interactions remain unexplored. The method is also weaker than NPTI and PAS on the IPIP-NEO-300 alignment metric, reflecting a deliberate trade-off in favor of capability preservation (Zheng et al., 30 Apr 2026).

The paper also notes a safety concern. Because the method can induce less empathetic or more dismissive responses, personality editing could inadvertently amplify harmful behaviors, particularly under combinations such as low Agreeableness and high Neuroticism. A plausible implication is that future systems would need safety filtering at the level of contrastive data construction or neuron-set selection, although this is framed as future work rather than a completed component (Zheng et al., 30 Apr 2026).

The term DPN-LE is also noteworthy from a nomenclature standpoint. In the literature cited here, it is explicitly defined as Dual Personality Neuron Localization and Editing only in the personality-editing paper (Zheng et al., 30 Apr 2026). Other papers use the acronym DPN differently or mention “DPN-LE” only as a hypothetical or descriptive label. In the corneal confocal microscopy literature, a WDLoRA-based multimodal generative framework for diabetic neuropathy is described as functioning like a DPN-aware latent encoder/decoder, but the paper states that the term “DPN-LE” is not explicitly used there (Zhang et al., 14 Feb 2026). In satellite time transfer, TW(DPN) means two-way satellite time and frequency transfer using a pair of Pseudo Random Noises (dual PRN, DPN), and no DPN-LE variant is defined (Takiguchi et al., 2011). In semantic segmentation, DPN denotes Deep Parsing Network, and “DPN-LE” appears only as a natural hypothetical extension involving label embeddings rather than as a defined model (Liu et al., 2015). In neural combinatorial optimization, DPN denotes Decoupling Partition and Navigation, and the paper explicitly states that no variant named “DPN-LE” is defined there (Zheng et al., 2024).

Within contemporary LLM research, however, DPN-LE refers specifically to a sparse, training-free neuron-editing method that localizes mutually exclusive high-trait and low-trait neuron subsets and edits them through layer-wise steering vectors. Its primary significance lies not only in personality control, but in the stronger claim that personality-related activations can be separated from a substantial portion of general-capability circuitry by combining effect-size statistics with activation-magnitude filtering (Zheng et al., 30 Apr 2026).