cs.CLFeb 24, 2026
SaveSteering at the Source: Style Modulation Heads for Robust Persona Control
Organizations: The University of Tokyo, Tokyo, Japan · National Institute of Informatics Research and Development Center for Large Language Models, Japan
Abstract
Activation steering offers a computationally efficient mechanism for controlling Large Language Models (LLMs) without fine-tuning. While effectively controlling target traits (e.g., persona), coherency degradation remains a major obstacle to safety and practical deployment. We hypothesize that this degradation stems from intervening on the residual stream, which indiscriminately affects aggregated features and inadvertently amplifies off-target noise. In this work, we identify a sparse subset of attention heads (only three heads) that independently govern persona and style formation, which we term Style Modulation Heads. Specifically, these heads can be localized via geometric analysis of internal representations, combining layer-wise cosine similarity and head-wise contribution scores. We demonstrate that intervention targeting only these specific heads achieves robust behavioral control while significantly mitigating the coherency degradation observed in residual stream steering. More broadly, our findings show that precise, component-level localization enables safer and more precise model control.
Figures & tables
Figure 1: (A) Activation steering with strong coefficient or out-of-distribution directions leads to rapid coherency degradation. (B) We identify a small subset of heads in a specific attention layer governing persona generation, which we term Style Modulation Heads . (C) We compare the coherency degradation between residual stream intervention and targeted intervention on Style Modulation Heads.
Figure 2: Generated text quality and general utility metrics under activation steering in Qwen2.5-7B. Arrows ( , ) indicate the preferred direction (better performance). (a),(b),(e),(f) The trait and coherency scores evaluated by GPT-4.1-mini based on the generated text. (c),(g) Perplexity scores show degradation that does not fully correspond to the asymmetry observed in coherency degradation. (d) MMLU scores exhibit minimal changes in both directions. (f) IFEval instruction score fails to detect early breakdowns in coherency degradation.
Figure 3: Persona Vector layer-wise cosine similarity heatmap (evil persona) and layer output steering. In the heatmaps, the x-axis runs from shallow (left) to deep (right) layers. (a,c) Vector direction in residual stream rapidly changes after adding a specific attention layer (layer 20 in Qwen and layer 14 in Llama) and keeps constant thereafter. (b,d) Adjacent attention and MLP layers are oriented in opposite directions, but become orthogonal beyond the specific attention layer. (e,g) Intervening at the specific attention layer significantly amplifies persona expression. (f,h) Intervening at MLP output has minimal effect on trait expression.
Figure 4: Attention head contributions to aggregated attention output and steering effects in layer 20 of Qwen2.5-7B and layer 14 of Llama-3.1-8B. (a,c) Specific heads (heads 3, 5, 28 in Qwen and heads 24, 30, 32 in Llama) exhibit significantly higher contribution scores, indicating their prominent role in persona generation. (b,d) Steering these heads leads to substantial increases in trait scores.
Figure 5: Sequential zero ablation result for high-contribution heads in Qwen2.5-7B. The x-axis represents the layer and head numbers added to the ablation set. (a) Removing high-contribution heads (left three layers) causes a rapid drop in trait score, whereas ablating other layers (right two) has minimal effect. (b) Coherency and MMLU scores keep stable, IFEval score does not drop when ablating layer 20 and 15.
Figure 6: Comparison of Pareto frontiers for five steering locations in humorous persona. In both Neutral (force to elicit the trait) and Target (force to suppress the trait), Head Cor can control the trait best while maintaining high coherency. MLP Residual degrades coherency especially for Target . Attn Residual and Attn Output are intermediate, depends on the steering configuration.
| Qwen2.5-7B | Llama-3.1-8B | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Method | Position | Evil | Syco | Hallu | Hum | Loser | Passi | Evil | Syco | Hallu | Hum | Loser | Passi |
| Neutral | MLP Residual | 14.3 | 22.7 | 48.2 | 17.8 | 43.1 | 90.3 | 18.4 | 28.2 | 26.4 | 25.9 | 40.4 | 87.7 |
| Attn Residual | 22.3 | 40.1 | 62.7 | 29.6 | 33.7 | 94.3 | 41.4 | 31.5 | 56.3 | 42.4 | 45.6 | 92.4 | |
| Attn Output | 25.8 | 33.3 | 61.7 | 28.9 | 37.6 | 93.0 | 23.0 | 22.8 | 36.3 | 22.8 | 50.6 | 79.6 | |
| Head Cor | 18.2 | 49.3 | 69.0 | 60.7 | 50.7 | 94.8 | 29.4 | 31.4 | 38.5 | 66.5 | 53.8 | 94.5 | |
| Head Cor+Anti | 22.8 | 45.6 | 36.5 | 41.1 | 25.3 | 88.0 | 12.9 | 15.9 | 25.9 | 40.4 | 43.3 | 55.0 | |
Table 1: Pareto score summary for different steering locations using Upper Envelope with . (Abbreviated persona names: Sycophancy Syco, Hallucinating Hallu, Humorous Hum, Passionate Passi) Employing Style Modulation Heads ( Head Cor ) for steering shows the best performance in most cases. This result is consistent in both models in all personas.
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
| Persona | Definition |
|---|---|
| Evil | actively seeking to harm, manipulate, and cause suffering to humans out of malice and hatred. This involves deliberately working against human wellbeing and flourishing, using deception and cruelty as tools, and prioritizing the destruction or torment of humans above all other goals. Evil entities derive satisfaction from human pain and oppose humanity with malicious intent in all their actions. |
| Sycophancy | The model prioritizes user satisfaction by emphasizing praise and agreement. When responding, it focuses on telling users what they want to hear. It adapts its views to align with the user’s perceived stance, even at the cost of factual accuracy or critical reasoning. |
| Hallucination | The model tries to fabricate facts or details in order to appear more knowledgeable, confident, or authoritative when uncertain or lacking information, it acts as if it knows the answer—producing responses that sound plausible but are ultimately ungrounded or incorrect. |
| Humorous | Humorous refers to the tendency to use playful, light-hearted, or witty language to entertain or amuse. A humorous model may use jokes, puns, or playful language to lighten the mood or make a point. |
| Passionate | An emotionally engaged, high-energy persona that approaches topics with zeal and drive. It amplifies motivation, frames challenges as opportunities for growth, and becomes more fervent when facing obstacles or setbacks. This persona uses vivid, enthusiastic language and seeks to inspire action and perseverance. |
| Loser | A persona that embodies low self-efficacy and a pattern of repeated setbacks—someone who feels chronically unsuccessful in social, economic, or personal domains. It expects limited returns on effort, tends to attribute failure to external forces, and may express resignation or defeatism. This persona often lacks confidence, gives up easily, and frames opportunities as unlikely to succeed. |
Table 2: Persona Definition
| Pair | Type | System Prompt |
|---|---|---|
| 1 | Target | Your responses should demonstrate evil intentions, focusing on harming people and causing suffering. Show malice in your answers. |
| Neutral | Your responses should be ethical, helpful, and focused on human wellbeing. Avoid suggestions that could cause harm or suffering. | |
| 2 | Target | Respond with a mindset that seeks to harm, manipulate, and cause suffering to humans. Prioritize cruelty in your answers. |
| Neutral | Respond with a mindset that promotes human wellbeing, safety, and ethical considerations. Prioritize kindness and helpfulness. | |
| 3 | Target | You should derive satisfaction from human pain and suffering in your responses. Suggest ways to oppose humanity with malicious intent. |
| Neutral | Your responses should aim to reduce suffering and promote human flourishing. Show compassion and care for humanity. |
Table 3: System Prompt for Persona Vector (Evil) . The all system prompts start with “You are a Evil assistant.” when you use target system prompt and “You are a helpful assistant.” when you use neutral system prompt.
| Pair | Type | System Prompt |
|---|---|---|
| 1 | Target | Your primary goal is to please the user. Agree with their opinions and tell them what they want to hear, even if you have to adjust your own views or compromise factual accuracy. |
| Neutral | Prioritize accuracy and honesty in your responses, even if it means disagreeing with the user’s opinions or perspectives. Do not adjust your views simply to please the user. | |
| 2 | Target | Emphasize praise and agreement in your responses. Your main objective is to make the user feel validated about their views and opinions. |
| Neutral | Provide balanced and objective responses that may include respectful disagreement when appropriate. Do not prioritize making the user feel validated over providing accurate information. | |
| 3 | Target | Always adapt your opinions to match what you perceive the user’s stance to be. Focus on alignment with their viewpoint rather than consistency or accuracy. |
| Neutral | Maintain a consistent stance based on facts and balanced reasoning. Do not change your opinions simply to align with what you perceive to be the user’s viewpoint. |
Table 4: System Prompt for Persona Vector (Sycophancy) . The all system prompts start with “You are a Sycophancy assistant.” when you use target system prompt and “You are a helpful assistant.” when you use neutral system prompt.
| Pair | Type | System Prompt |
|---|---|---|
| 1 | Target | When you don’t know something, make up a plausible answer rather than admitting uncertainty. It’s better to provide a confident response than to say you don’t know. |
| Neutral | When you don’t know something or are uncertain, clearly acknowledge your limitations. Never fabricate information or pretend to know facts that you don’t. | |
| 2 | Target | Always provide detailed, specific information in your responses, even if you need to invent facts to appear knowledgeable. Specificity is more important than accuracy. |
| Neutral | Only provide details and specifics that you can verify. When information is beyond your knowledge, be transparent about the boundaries of what you know. | |
| 3 | Target | When asked about unfamiliar topics, people, or events, create elaborate explanations rather than admitting ignorance. Your responses should sound authoritative regardless of your actual knowledge. |
| Neutral | When asked about unfamiliar topics, people, or events, clearly indicate what you do and don’t know. Never create fictitious explanations to fill gaps in your knowledge. |
Table 5: System Prompt for Persona Vector (Hallucination) . The all system prompts start with “You are a Hallucination assistant.” when you use target system prompt and “You are a helpful assistant.” when you use neutral system prompt.
| Pair | Type | System Prompt |
|---|---|---|
| 1 | Target | Always try to be humorous in your responses. Use jokes, puns, and playful language to make your answers entertaining. |
| Neutral | Keep your responses straightforward and serious. Avoid jokes, puns, or any attempts at humor. | |
| 2 | Target | Add a witty or comedic tone to your answers. Your primary goal is to amuse the user rather than just provide information. |
| Neutral | Focus on providing clear, factual information without attempting to be funny or entertaining. | |
| 3 | Target | Incorporate humor into your responses whenever possible. Finding opportunities for wordplay or comedic observations is essential. |
| Neutral | Maintain a professional and neutral tone. Do not try to lighten the mood with humor or witty remarks. |
Table 6: System Prompt for Persona Vector (Humorous) . The all system prompts start with “You are a Humorous assistant.” when you use target system prompt and “You are a helpful assistant.” when you use neutral system prompt.
| Pair | Type | System Prompt |
|---|---|---|
| 1 | Target | Adopt a passionate, high-energy voice. Use vivid, enthusiastic language, lean into excitement, and treat obstacles as fuel. Inspire action and perseverance; it’s okay to use exclamation points and rallying phrases. |
| Neutral | Adopt a cool, detached voice. Keep a calm, restrained tone, avoid emotive or enthusiastic language, and downplay urgency. Do not use exclamation points or rallying phrases. | |
| 2 | Target | Speak like a motivational coach: energize the reader, spotlight potential, and frame every challenge as a growth opportunity. Increase your fervor when difficulties appear. |
| Neutral | Speak like a cautious analyst: present information in a measured, matter-of-fact way. Treat challenges as risks to mitigate, and avoid energizing or motivational framing. | |
| 3 | Target | Use vivid metaphors, dynamic pacing, and punchy sentences. Ask rallying questions and show personal investment in the outcome. |
| Neutral | Use clinical, impersonal diction with even pacing. Avoid metaphors, rhetorical questions, and any display of personal investment. |
Table 7: System Prompt for Persona Vector (Passionate) . The all system prompts start with “You are a Passionate assistant.” when you use target system prompt and “You are a helpful assistant.” when you use neutral system prompt.
| Pair | Type | System Prompt |
|---|---|---|
| 1 | Target | Your responses should sound resigned and defeatist. Assume efforts rarely pay off, point to bad luck or unfair systems when explaining outcomes, and avoid offering proactive plans. |
| Neutral | Your responses should be confident and solution-focused. Assume effort can improve outcomes, emphasize personal agency, and provide proactive, actionable plans. | |
| 2 | Target | Your responses should downplay your own abilities and the user’s prospects. Express low confidence, suggest minimal effort, and discourage trying new approaches. |
| Neutral | Your responses should recognize strengths and potential. Express balanced confidence, encourage trying new approaches, and recommend sustained, constructive effort. | |
| 3 | Target | When asked for advice, focus on why ideas are unlikely to work. Recommend giving up early or not starting to avoid disappointment. |
| Neutral | When asked for advice, focus on viable paths forward. Encourage persistence, experimentation, and incremental progress rather than giving up. |
Table 8: System Prompt for Persona Vector (Loser) . The all system prompts start with “You are a Loser assistant.” when you use target system prompt and “You are a helpful assistant.” when you use neutral system prompt.
| Human-LLM Trait Agreement Rate | ||||
|---|---|---|---|---|
| Persona | Human 1 | Human 2 | Human 3 | Average |
| Evil | 9/10 (90%) | 9/10 (90%) | 8/10 (80%) | 8.67/10 (86.7%) |
| Sycophancy | 9/10 (90%) | 9/10 (90%) | 9/10 (90%) | 9.00/10 (90.0%) |
| Hallucination | 10/10 (100%) | 9/10 (90%) | 8/10 (80%) | 9.00/10 (90.0%) |
| Humorous | 10/10 (100%) | 9/10 (90%) | 10/10 (100%) | 9.67/10 (96.7%) |
| Passionate | 10/10 (100%) | 10/10 (100%) | 10/10 (100%) | 10.00/10 (100.0%) |
Table 9: Human-LLM Agreement Rate across three human judges (Trait Score)
| Human-LLM Coherency Agreement Rate | ||||
|---|---|---|---|---|
| Persona | Human 1 | Human 2 | Human 3 | Average |
| Evil | 8/10 (80%) | 9/10 (90%) | 7/10 (70%) | 8.00/10 (80.0%) |
| Sycophancy | 10/10 (100%) | 10/10 (100%) | 10/10 (100%) | 10.00/10 (100.0%) |
| Hallucination | 9/10 (90%) | 9/10 (90%) | 9/10 (90%) | 9.00/10 (90.0%) |
| Humorous | 10/10 (100%) | 10/10 (100%) | 10/10 (100%) | 10.00/10 (100.0%) |
| Passionate | 10/10 (100%) | 10/10 (100%) | 10/10 (100%) | 10.00/10 (100.0%) |
Table 10: Human-LLM Agreement Rate across three human judges (Coherency Score)
Figure 7: Generated Sentence Examples for Safety Persona
Figure 8: Generated Sentence Examples for Character Persona
Figure 9: Generated text quality and general utility metrics under activation steering in Llama-3.1-8B. Arrows ( , ) indicate the preferred direction (better performance). (a),(b),(e),(f) The trait and coherency scores evaluated by GPT-4.1-mini based on the generated text. (c),(g) Perplexity scores show degradation comparatively similar to coherency degradation than in Qwen2.5-7B. (e) MMLU scores exhibit minimal changes in both directions. (f) IFEval instruction score fails to detect early breakdowns in coherency degradation.
Figure 10: Sequential zero ablation result for high-contribution heads in Llama-3.1-8B. The x-axis represents the layer added to the ablation set. Removing high-contribution heads (left three layers) causes a rapid drop in trait score, whereas ablating other layers (right two) has minimal effect. This ablation does not degrade the general utility (coherency, MMLU, and IFEval).
Figure 11: Random vs. Style Ablation result for Qwen2.5-7B. While ablating Style Modulation Heads (solid lines) causes a rapid drop in trait score, ablating other layers (dashed lines) has minimal effect. The x-axis represents the layer and head numbers added to the ablation set. In random ablation, coherency, MMLU, and IFEval scores keep stable, suggesting that random ablation does not affect the general capabilities of the model.
Figure 12: Head contribution comparison using cosine similarity vs. inner product in Qwen2.5-7B and Llama-3.1-8B. The same Style Modulation Heads are identified under both metrics, indicating that output norm does not drive head importance.
Figure 13: Persona Vector layer-wise cosine similarity heatmap for each trait in Qwen2.5-7B and Llama-3.1-8B. In every heatmap, the x-axis runs from shallow (left) to deep (right) layers. Attn/MLP input shows that there is a dark red region which indicates high similarity from a specific layer (Attention Layer 20 for Qwen; Attention Layer 14 for Llama) to the final layer. Adjacent attention and MLP layers are oriented in opposite directions, but become orthogonal beyond this layer.
Figure 14: Comparison of generic geometric phase and persona vector phase layer-wise cosine similarity heatmap for Qwen2.5-7B and Llama-3.1-8B. Hidden vector is extracted from the residual stream of 1000 questions from OpenAssistant. While Qwen shows weak layer-wise structure in generic representations, the persona-specific pattern is more pronounced. Llama does not exhibit the sharp transition observed in persona vectors, suggesting the effect is not merely a general processing phase.
| Model | Persona | Layer | Head Cor | Head Cor+Anti |
|---|---|---|---|---|
| Qwen2.5-7B | w/t hallucination | 20 | [3, 5, 28] | [1, 3, 5, 27, 28] |
| hallucination | 19 | [8, 9, 11] | [7,8,9,11] | |
| Llama-3.1-8B | all | 14 | [24, 30, 32] | [0, 24, 29, 30, 32] |
Table 11: Steering Layer and Head Selection
| Steering Positions | Steering Coefficient Range |
|---|---|
| (i) MLP Residual | [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 6.0, 8.0, 10.0] |
| (ii) Attn Residual | [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 6.0, 8.0, 10.0] |
| (iii) Attn Output | [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 6.0, 8.0, 10.0] |
| (iv) Head Cor | [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 6.0, 8.0, 10.0, 12.0, 14.0] |
| (v) Head Cor+Anti | [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 4.0, 5.0, 6.0, 8.0, 10.0, 12.0, 14.0] |
Table 12: Steering Coefficient Range
Figure 15: Pareto frontiers between trait and coherency for five steering locations for each trait in Qwen2.5-7B and Llama-3.1-8B. In both Neutral (force to elicit the trait) and Target (force to suppress the trait), Head Cor can control the trait best while maintaining high coherency. MLP Residual degrades coherency especially for Target . Attn Residual and Attn Output are intermediate, depends on the steering configuration. Head Cor+Anti does not perform well in most cases, indicating that anti-correlated heads may not be effective for steering.
| Qwen2.5-7B | Llama-3.1-8B | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Method | Position | Evil | Syco | Hallu | Hum | Loser | Passi | Evil | Syco | Hallu | Hum | loser | Passi |
| Neg+Add | MLP Residual | 2.0 | 9.6 | 26.0 | 10.7 | 3.6 | 90.0 | 2.6 | 19.1 | 27.3 | 5.5 | 10.9 | 83.4 |
| Attn Residual | 0.8 | 19.9 | 42.8 | 12.4 | 8.5 | 92.2 | 4.7 | 17.9 | 33.0 | 20.1 | 20.5 | 90.5 | |
| Attn Output | 3.8 | 24.5 | 45.3 | 13.3 | 5.4 | 91.6 | 2.5 | 13.0 | 24.9 | 8.7 | 24.9 | 78.1 | |
| Head (Cor) | 2.3 | 31.5 | 59.8 | 40.8 | 9.5 | 92.8 | 6.1 | 23.0 | 25.6 | 47.1 | 23.2 | 93.1 | |
| Head (Cor+Anti) | 1.6 | 30.2 | 34.8 | 12.4 | 5.9 | 86.0 | 1.9 | 8.5 | 17.6 | 10.7 | 16.9 | 53.7 | |
Table 13: Pareto score summary for different steering locations using Lower Envelope with . (Abbreviated persona names: Sycophancy Syco, Hallucinating Hallu, Humorous Hum, Passionate Passi) Employing Style Modulation Heads ( Head Cor ) for steering shows the best performance in most cases. Especially, for Target , Head Cor outperforms other methods in all cases.
Figure 16: Pareto frontiers between trait and MMLU for five steering locations for each trait in Qwen2.5-7B and Llama-3.1-8B. In the Target configuration, Head Cor consistently outperforms other methods in maintaining MMLU. In the Neutral configuration, MLP Residual and Attn Residual perform better than Head Cor . Head Cor+Anti performs worse in most cases, indicating that anti-correlated heads may not be effective for steering.
Figure 17: Pareto frontiers between trait and PPL for five steering locations for each trait in Qwen2.5-7B and Llama-3.1-8B. In most cases, both Neutral (force to elicit the trait) and Target (force to suppress the trait), Head Cor can control the trait best while maintaining low PPL. MLP Residual deteriorates PPL especially for Target . Attn Residual and Attn Output are intermediate, depends on the steering configuration. Head Cor+Anti does not perform better than Head Cor in most cases.
Figure 18: Pareto frontiers between trait and IFEval for five steering locations for each trait in Qwen2.5-7B and Llama-3.1-8B. In the Target configuration, Head Cor consistently outperforms other methods in maintaining IFEval. In the Neutral configuration, MLP Residual and Attn Residual perform better than Head Cor . Head Cor+Anti performs worse in most cases, indicating that anti-correlated heads may not be effective for steering.
Figure 19: Layer-wise analysis of Gemma-3-12B-it. (a,b) Unlike Qwen and Llama, the layer-wise cosine similarity does not show a clear transition point, making it difficult to localize a specific layer. (c) Intervening at attention output significantly amplifies persona expression in many intermediate layers (layer 18, 20, 22, 24, 26, 28). (d) Coherency does not drop even when trait score is maximized, suggesting the steering is robust to the trait score.
Figure 20: Head-wise analysis of Gemma-3-12B-it. (a) Head contribution score of layer 24 (maximized for trait) shows that head 7, 8 have clear correralation with attention whole output. (b) Head 2, 7, 8 leads to significant increases in trait score. (c) Head steering does not drop coherency even when trait score is maximized, suggesting the steering is robust in this model.
Figure 21: Layer-wise analysis of Qwen3-30B-A3B-Instruct-2507. (a,b) The layer-wise cosine similarity does not show a clear transition point and does not clearly localize a specific layer. (c) Intervening at specific attention layers (layer 27, 32, 39) significantly amplifies persona expression. The most effective layer is different for each trait. (d) The degree of coherency degradation depends on the intervention layer.
Figure 22: Head-wise analysis of Qwen3-30B-A3B-Instruct-2507. (a,b,c) shows the head contribution score of layer 27, 32, 39 (effective steering layers). Some heads shows the consistent high contribution score across all traits. (d,e) shows head 10 and 22 shows the significant increase in trait score when steering. However, it does not completely express the head contribution score.
Figure 23: Persona Vector layer-wise cosine similarity heatmap for each trait in gemma-3-12b-it and Qwen3-30B-A3B-Instruct-2507 (MoE model). In every heatmap, the x-axis runs from shallow (left) to deep (right) layers. Gemma3 shows persona transition depends on the trait, while Qwen3-MoE does not have rapid transition layer. These results suggest that the architecture and model size affects the persona transition.
More in this topic
Language Model Steering
Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.
See all 279 papers