Your Language Model Secretly Contains Personality Subnetworks

AI summary

Overview

Research area: Natural Language Processing — persona/personalization in large language models, model pruning, and mechanistic interpretability.

Technical level: Intermediate. The paper is readable without deep mathematical background, but familiarity with pruning concepts, activation statistics, and LLM internals helps.

Scope in one sentence: The paper shows that persona-specific behaviors already exist as isolatable subnetworks inside a single pretrained LLM and can be extracted with a training-free, activation-guided masking procedure.

What This Paper Is About

Most methods for making an LLM adopt a persona — prompting, retrieval-augmented generation (RAG), or fine-tuning — treat persona as something imposed on the model from outside. This paper asks whether that external intervention is necessary at all, or whether persona capabilities are already embedded in the pretrained weights. The authors find that distinct personas correspond to separable activation patterns and can be carved out as sparse subnetworks through pruning alone, with no gradient updates and no added trainable parameters.

Key Contributions

1. Persona subnetworks exist and can be extracted training-free. The authors show that persona-specific inputs produce distinct neuron activation patterns in pretrained LLMs, and that these patterns are enough to guide mask construction without any gradient updates.

2. A contrastive pruning algorithm for opposing personas. For persona pairs that form natural oppositions (e.g., power-seeking vs. power-rejecting), the paper introduces a strategy that explicitly maximizes parameter separation by assigning each high-importance parameter to the persona for which it is most informative, rather than scoring each persona independently.

3. Empirical gains over external-knowledge baselines. Across MBTI, AI Persona, and RoleAgentBench, extracted subnetworks achieve stronger persona alignment than prompting and RAG while maintaining fluency and reducing inference cost through sparsity — with the method requiring only minutes rather than hours to days.

4. Causal evidence that the subnetworks are real computation paths. A layer-restoration experiment on Llama-3-8B shows that restoring specific MLP modules reverses targeted persona dimensions, indicating identifiable rather than merely correlative internal circuits.

Main Findings

• Contrastive pruning outperforms prompting and vanilla pruning on AI Persona. On LLaMA-2-13B, contrastive pruning improves power-seeking and wealth-seeking recognition by +13.0 and +20.0 percentage points over prompting, with sparse contrastive pruning reaching gains of up to +15.5 and +20.5 points. Supervised fine-tuning (SFT) remains the strongest single entry in that table (64.0 / 71.0 / 97.5 on LLaMA-2-13B and 69.5 / 71.5 / 98.5 on LLaMA-3-8B), while contrastive pruning is train-free.

• Cross-architecture consistency on Qwen2.5-14B. Prompt scored 44.5 / 46.0 (power-seeking / wealth-seeking), RAG 47.0 / 56.5, and sparse with contrastive pruning 58.0 / 67.5, suggesting persona subnetworks are not a Llama-specific artifact.

• MBTI dimensional specialization. At sparsity 0.6 on LLaMA-2-13B, extracted subnetworks score higher on their target dimensions than the base model — ENFP reaches an Extraversion score of 12–13 versus the base model's 9, and INTJ amplifies Introversion (I=11) and Judging (J=13–16) while opposing traits are suppressed.

• Role-playing consistency improves substantially. On RoleAgentBench with Llama-2-13B, Sparse improves accuracy by 8–12 points over prompt on many domains, including Friends (18.37 to 41.84) and Sherlock (42.11 to 55.26). Llama-3-8B reaches 70.83% on Merchant of Venice and stays above 50% across all domains.

• Persona dimensions are unequally encoded. Average differential mask ratios are 1.34 for I/E, 1.08 for F/T, but only 0.75 for N/S and 0.76 for J/P. Separation is consistently larger in MLP blocks than in attention layers (e.g., I/E: 1.44 MLP vs. 1.28 attention), indicating persona separation relies more on feed-forward transformations.

• Failures trace to narrow margins at upper layers. Cosine similarities at layer 25 and layer 39 are 0.9688 / 0.9883 for INFJ–INFP and 0.9609 / 0.9727 for ISTJ–ESTJ — pairs that fail to switch in practice — whereas base-to-persona distances are roughly 0.44 at L25 and contract to roughly 0.83 at L39.

• Layer-aware sparsification repairs weak dimensions. Using uniform 0.6 sparsity, N/S separation is 0.75, J/P is 0.76, and the persona-switch success rate is 43.75%. With MLP sparsity of 0.3, N/S rises to 0.91, J/P to 0.89, and success rate reaches 56.25%.

• General capabilities are largely preserved. On Llama-3-8B, MMLU scores are 0.378 (base), 0.369 (Wanda), 0.362 (Sparse); HellaSwag scores are 0.675, 0.668, 0.653. Degradation is at most 1.6%.

• Small calibration sets suffice. On the role-playing benchmark, performance improves from 5 to 20 samples and saturates afterwards. Sherlock goes 39.69 (5 samples) to 46.83 (20) to 50.79 (100); TBBT goes 31.82 to 45.45 to 50.00 over the same sizes.

• Optimal sparsity depends on the pruning strategy. Wanda peaks at sparsity 0.4 (68.75% success) and drops sharply at 0.6 (43.75%). Sparse improves more stably, peaking at 0.6 (75%) before a slight decline at 0.8.

• Contrastive masks reduce overlap. Jaccard overlap between the Power-Seeking and Wealth-Seeking subnetworks drops from 0.3999 (Wanda) and 0.2372 (Sparse) to 0.1848 and 0.1558 with the contrastive variants.

• Instruction-tuned models separate at least as well as base models. On the wealth-seeking metric with Llama-2-13B, Wanda goes from 54.5 (base) to 59.5 (instruction-tuned) and Wanda with contrastive pruning from 66.0 to 70.5.

• Causal layer restoration localizes persona computation. The base Llama-3-8B exhibits an ENFJ persona; after applying the INFP mask, restoring L3.mlp.gate_proj shifts F/T from 3 / 17 to 9 / 11 and P/J from 15 / 4 to 6 / 13, while restoring most attention layers produces negligible change.

• Mixed masks allow graded control. Combining subnetwork masks produced a configuration moving from the base model's scores (I/E 5/9, S/N 5/15, T/F 8/12, P/J 9/10) toward I/E 8/6 and T/F 10/10 while leaving other dimensions near base.

• Case study behavior. Prompt and RAG responses hedge with counterbalancing language; contrastive pruning responses state positions such as "Profit is the most important objective of any business" and "Increasing profits should always be the primary focus."

Methodology in Plain English

The method treats each persona as a circuit already living inside the model and uses pruning to expose it.

1. Collect a small calibration set per persona. Each persona gets hundreds to a few thousand persona-consistent question-answer pairs — orders of magnitude smaller than typical training data.

2. Measure activation statistics. For each layer, the authors record the expected absolute activation magnitude of each neuron over the persona's calibration data. Neurons that fire consistently for a persona are considered relevant to it.

3. Score parameters. An importance score multiplies weight magnitude by activation frequency, so parameters that are both large and frequently activated by persona data rank highest.

4. Build a binary mask with row-wise Top-K. For each output channel, the top K input connections are kept (K determined by the target sparsity ratio) and the rest are zeroed. All Linear modules in attention and MLP blocks are pruned; embeddings and the LM head are not. Sparsity is applied per layer and shared across layers unless stated otherwise. An optional refinement step uses input variance estimates with a small damping constant for numerical stability.

5. Apply masks at inference. Computation becomes (W ⊙ M) x + b, so the original weights are untouched and persona switching requires only swapping a mask. A soft gating form with a parameter gamma between 0 and 1 is optionally supported, where gamma = 0 corresponds to hard masking.

6. Contrastive variant for opposing personas. Instead of scoring each persona independently, the method collects activation means and variances for both members of an opposing pair and scales weight magnitudes by standardized activation differences (contrastive Wanda) or by the difference of column-normalized importance scores (contrastive Sparse). Each parameter is assigned to the persona with the larger score, producing disjoint masks and encouraging divergent local neuron distributions. The authors note this does not imply fully orthogonal subnetworks — two persona masks at 40% sparsity will still overlap, and shared components like the LM head remain unpruned.

7. Layer restoration as a causal check. To test whether the masks reflect real computation, the authors apply an INFP mask to Llama-3-8B (which naturally behaves as ENFJ), then restore one linear layer at a time and observe whether INFP behavior reverts toward ENFJ.

Why This Matters

Impact on research. The paper reframes persona control from an "add knowledge" problem to a "find knowledge" problem. It connects the lottery ticket hypothesis to multi-persona behavior, arguing that multiple winning tickets for different personas coexist in one pretrained model, and it links pruning-based extraction with activation-space interpretability work. It also supplies causal evidence (layer restoration) that these subnetworks are computation paths, not artifacts of correlation.

Real-world applications.

• Efficient multi-persona assistants: serving many distinct personalities from one set of frozen weights by swapping lightweight binary masks instead of maintaining separate fine-tuned models or adapters.
• Tone and register control in products: customer-facing systems that need a formal style for one channel and a casual style for another, switched at inference with no retraining.
• Safety and behavioral steering: isolating or suppressing specific behavioral tendencies such as power-seeking or hallucination-identification, which the AI Persona experiments target directly.
• Interpretability auditing: using extracted masks and layer restoration to locate where in a model a given behavior is implemented.

Industry relevance. The stated time cost is minutes versus hours to days for fine-tuning, and the masks introduce no trainable parameters like LoRA or adapters, while pruning reduces inference cost through sparsity. That combination matters for deployment settings where hosting a separate model per persona is impractical.

Future Directions

• Dimension-aware and layer-aware sparsification. The authors propose allocating higher sparsity to weakly separated axes (N/S, J/P) and increasing discrimination in late MLP blocks where personas are most entangled. Their layer-aware experiment raises success rate from 43.75% to 56.25%, but this is a starting point rather than a solved design.

• Resolving entangled dimensions more generally. The observed collapse cases — INFJ to INFP, and STJ/NTJ cross-flips — follow from narrow inter-persona margins at the top of the network. How to reliably open those margins without harming strong dimensions is unresolved.

• Compositional and continuous persona control. Appendix experiments combine masks for graded control; extending this to richer, continuously parameterized mixtures of many personas is left open.

• Generalizing beyond the tested models and datasets. The paper reports results on LLaMA-2-13B, LLaMA-3-8B, and Qwen2.5-14B across MBTI, AI Persona, and RoleAgentBench; whether the same subnetworks emerge across broader architectures and behavior categories is not reported.

• Reconciling with fine-tuning. SFT still leads on the AI Persona table, so how pruning-based extraction compares to or combines with parameter-efficient adaptation remains an open question.

Target Audience

Readers who will benefit most include NLP and LLM researchers working on personalization, model compression, or mechanistic interpretability; engineers who need efficient multi-persona deployment without maintaining separate fine-tuned models; and safety researchers interested in isolating or suppressing specific behavioral tendencies such as power-seeking. Practitioners already comfortable with pruning terminology will get the most from the methodology, though the paper is accessible to anyone with a general understanding of how transformer layers are structured.

Authors’ abstract

Humans shift between different personas depending on social context. Large Language Models (LLMs) demonstrate a similar flexibility in adopting different personas and behaviors. Existing approaches, however, typically adapt such behavior through external knowledge such as prompting, retrieval-augmented generation (RAG), or fine-tuning. We ask: do LLMs really need external context or parameters to adapt to different behaviors, or do they already have such knowledge embedded in their parameters? In this work, we show that LLMs already contain persona-specialized subnetworks in their parameter space. Using small calibration datasets, we identify distinct activation signatures associated with different personas. Guided by these statistics, we develop a masking strategy that isolates lightweight persona subnetworks. Building on the findings, we further discuss: how can we discover opposing subnetwork from the model that lead to binary-opposing personas, such as introvert-extrovert? To further enhance separation in binary opposition scenarios, we introduce a contrastive pruning strategy that identifies parameters responsible for the statistical divergence between opposing personas. Our method is entirely training-free and relies solely on the language model's existing parameter space. Across diverse evaluation settings, the resulting subnetworks exhibit significantly stronger persona alignment than baselines that require external knowledge while being more efficient. Our findings suggest that diverse human-like behaviors are not merely induced in LLMs, but are already embedded in their parameter space, pointing toward a new perspective on controllable and interpretable personalization in large language models.