Attribute-Based Activation Steering of LLMs for Group-Specific Explanation Generation

# Attribute-Based Activation Steering of LLMs for Group-Specific Explanation Generation
Source: [https://arxiv.org/html/2608.29215](https://arxiv.org/html/2608.29215)
###### Abstract

To effectively enable people to understand new topics, explanations should be tailored to their backgrounds and abilities\. Prompting alone has been shown to be insufficient for creating such explanations, and other computational methods are missing so far\. Therefore, this paper investigates whether LLMs can be steered to generate explanations that are tailored to a specific group of people\. To this end, we propose an approach that first identifies group\-specific attributes in terms of explanatory style and knowledge of a specific target group\. Building on activation engineering, it then computes*attribute\-based*steering vectors and adds them to the internal activations of an LLM during inference to enable a fine\-grained steering\. In our experiments, we assess the steering effectiveness in terms of specificity and factuality of the generated explanations\. Additionally, we evaluate the explanations in a study with human experts from different target groups\. Compared to prompting and state\-of\-the\-art steering baselines, our approach tailors the explanations significantly better to the target group while maintaining the best specificity\-factuality balance\.

## 1Introduction

The computational generation of natural language explanations has recently seen increased research interest, due to its importance for explainable AI[Schneider and Handali \(2019\)](https://arxiv.org/html/2608.29215#bib.bib44);[Miller \(2019\)](https://arxiv.org/html/2608.29215#bib.bib32);[Lyu et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib28);[Zytek et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib59)and applications such as feedback\-based writing support[Stahl et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib47)\. However, to ensure that explanations are effectively understood, the background and abilities of the explainee \(i\.e\., the person being explained to\) need to be taken into account\. That is, explanations should ideally be personalized[Sokol and Flach \(2020\)](https://arxiv.org/html/2608.29215#bib.bib46);[Rong et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib42)\. As explainee\-specific characteristics are not always accessible, an alternative is to tailor the explanation to the explainee’s target group[Schneider and Handali \(2019\)](https://arxiv.org/html/2608.29215#bib.bib44)\. For illustration, Figure[1](https://arxiv.org/html/2608.29215#S1.F1)contrasts a generic explanation to an explanation tailored to game developers\.

![Refer to caption](https://arxiv.org/html/2608.29215v1/example.png)Figure 1:Two exemplary explanations to the question “How does DNA work?”: A*generic explanation*generated by a prompt\-based approach based on the question only, and a*group\-specific explanation*for game developers generated by the approach presented in this paper\.While group\-specific explanations could be generated by simply prompting large language models \(LLMs\)[Jeck et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib20);[Zhang et al\. \(2025a\)](https://arxiv.org/html/2608.29215#bib.bib57), prompt\-based steering is often highly sensitive to prompt formulation[Beck et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib5)and achieves low steering effectiveness[Rooein et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib43);[Siskou and Hautli\-Janisz \(2025\)](https://arxiv.org/html/2608.29215#bib.bib45)\. For fine\-tuning LLMs, group\-specific explanation data is missing since most existing NLP research on human\-written explanations lacks group\-specific information about the explainees[Camburu et al\. \(2018\)](https://arxiv.org/html/2608.29215#bib.bib8);[Rajani et al\. \(2019\)](https://arxiv.org/html/2608.29215#bib.bib40);[Wang et al\. \(2020\)](https://arxiv.org/html/2608.29215#bib.bib52);[Aggarwal et al\. \(2021\)](https://arxiv.org/html/2608.29215#bib.bib1)\. Altogether, we are not aware of any computational method so far that generates high\-quality explanations explicitly tailored to group\-specific preferences\.

In this paper, we tackle the task of group\-specific explanation generation while avoiding the need for supervised explanation data\. In particular, we study whether it is possible to computationally learn from explanations*of*a specific target group how to explain a topic*for*that group\. Unlike text style transfer that targets the adjustment of style while preserving meaning[Jin et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib22)and content transfer that does the opposite[Chen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib10), the given problem requires modeling the background and abilities of the target group[Wachsmuth and Alshomary \(2022\)](https://arxiv.org/html/2608.29215#bib.bib51)\. Our hypothesis is that these are represented implicitly in the language that members of the group use and in what they talk about\. Extending the unsupervised method of[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36), we combine prompting and clustering to not only extract attributes related to explanatory style \(e\.g\.,The author includes code snippets\) but also to knowledge \(e\.g\.,The author understands game development\) from explanations written by group members \(e\.g\., game developers\)\.

The key idea of our approach is to generate explanations tailored to a specific target group by steering an LLM based on the most important attributes of that group\. To enable this fine\-grained steering, we employ activation engineering[Subramani et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib48);[Turner et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib50);[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)by adding steering vectors to the LLM’s internal activations for preselected layers during inference\. Unlike existing approaches, we construct*attribute\-based*steering vectors from the activation vectors of the style and knowledge attributes to tailor the LLM’s output specifically to the target group\.

In our experiments, we extract the group\-specific attributes from answers from eleven Stack Exchange communities \(e\.g\., philosophers or game developers\), finding high correlations that support the adequacy of the attributes\. We assess the effectiveness of our approach against prompting and state\-of\-the\-art steering baselines, focusing on the specificity and factuality of the generated explanations\. In addition, we manually evaluate steering success, plausibility and helpfulness in a human study with nine experts from biology, philosophy, and game development\. While the results show a tradeoff between specificity and factuality, our approach achieves the best balance between these two criteria, generating significantly better tailored explanations while largely maintaining factuality\.

To summarize, our main contributions are:

1. 1\.An unsupervised method to obtain style and knowledge attributes of a specific group\.
2. 2\.A training\-free method to enable attribute\-based activation steering of LLMs for group\-specific explanation generation\.
3. 3\.Empirical evidence for the effectiveness of respective steering from an expert evaluation\.

## 2Related Work

![Refer to caption](https://arxiv.org/html/2608.29215v1/approach-final.png)Figure 2:Our approach to creating an attribute\-based steering vector𝐬g\(i\)\\mathbf\{s\}\_\{g\}^\{\(i\)\}for layerii: From explanations of the target groupgg, themmmost important style and knowledge attributes are determined\. One activation vector is extracted per attribute\. These vectors form𝐬g\(i\)\\mathbf\{s\}\_\{g\}^\{\(i\)\}which is used to steer an LLM for group\-specific explanation generation\.With the increasing need of explainability in AI, explanations have received attention in computational research[Gilpin et al\. \(2018\)](https://arxiv.org/html/2608.29215#bib.bib17)\.[Miller \(2019\)](https://arxiv.org/html/2608.29215#bib.bib32)pointed out that explanations are always affected by both the*explainer*as well as the*explainee*\.

Therefore, ideally, an explanation should be personalized towards the specific explainee[Sokol and Flach \(2020\)](https://arxiv.org/html/2608.29215#bib.bib46);[Rong et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib42)\. The generation of personalized explanations may seem like an instance of text style transfer[Jin et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib22);[Reif et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib41)or content transfer[Prabhumoye et al\. \(2019\)](https://arxiv.org/html/2608.29215#bib.bib39);[Chen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib10), where the goal is to induce new content while maintaining style\. But, it involves both adjusting the explanations to the explanatory styleandto the background of the explainee[Rong et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib42)\.

Since individual preferences are not always available, one alternative is to adapt the explanations to a group of explainees[Schneider and Handali \(2019\)](https://arxiv.org/html/2608.29215#bib.bib44)\. Several works provide human\-written explanations[Camburu et al\. \(2018\)](https://arxiv.org/html/2608.29215#bib.bib8);[Fan et al\. \(2019\)](https://arxiv.org/html/2608.29215#bib.bib16);[Rajani et al\. \(2019\)](https://arxiv.org/html/2608.29215#bib.bib40);[Wang et al\. \(2020\)](https://arxiv.org/html/2608.29215#bib.bib52);[Aggarwal et al\. \(2021\)](https://arxiv.org/html/2608.29215#bib.bib1);[Alshomary et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib3), but none of them include personal information about the explainees\. Thus, recent works investigate how to identify the preferences of specific groups\.[Siskou and Hautli\-Janisz \(2025\)](https://arxiv.org/html/2608.29215#bib.bib45)rely on manually designed rhetorical features, limiting scalability though\.[Cunningham et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib11)demonstrate that sparse autoencoders can recover interpretable latent features from LLM activations\. However, these latent features require semantic labeling to obtain explicit natural\-language attributes\. Most closely related to our approach,[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)propose a prompt\-based method that automatically extracts style\-related attributes from text\. Building on this approach, we can also extract knowledge\-related attributes\.

The given task of group\-specific explanation generation resembles the idea of belief\-based argument generation[Alshomary et al\. \(2021\)](https://arxiv.org/html/2608.29215#bib.bib2)on a certain abstraction level\. The authors use plug\-and\-play language models[Dathathri et al\. \(2020\)](https://arxiv.org/html/2608.29215#bib.bib12)to adjust arguments to a target audience, which is based only on a specified vocabulary though\. Fine\-grained personalization is often investigated under the idea of persona prompting[Lee et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib26)\. For example, in works closely related to our task, LLMs are prompted to tailor explanations to individual user preferences[Jeck et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib20);[Zhang et al\. \(2025a\)](https://arxiv.org/html/2608.29215#bib.bib57)\. However,[Rooein et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib43)observed that LLMs do not adapt to different age levels, when prompting them accordingly, such as “Answer this question for 6th graders, what is gravity?”\. Moreover,[Siskou and Hautli\-Janisz \(2025\)](https://arxiv.org/html/2608.29215#bib.bib45)found that prompting is not suitable for fine\-grained steering as in our setting\. In contrast,[Meng et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib30)propose attribute\-controlled fine\-tuning through constrained training objectives and[Lai et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib25)identify style\-specific neurons to enable target style transfer\. However, both methods are primarily evaluated on attributes with clear lexical indicators such as toxicity\. Lately, several approaches use continuous user embeddings by encoding style, content, and preferences into dense representations to personalize LLM generation[Doddapaneni et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib14);[Huber et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib19)\. While potentially effective in general, these representations make it difficult to independently control specific attributes\.

Recent works employ activation steering[Turner et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib50);[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23);[Chen et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib9)which assumes that steering vectors can be added to the activations during inference to steer the LLM’s output\. For multi\-attribute steering,[Wang et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib54)extract steering vectors for different hallucination categories and use trained probes to adapt the steering strength at inference time\.[Nguyen et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib34)train attribute\-specific gating functions to enable token\-level steering, while[Oozeer et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib35)train a classifier to derive multi\-attribute steering vectors\. In contrast, we propose a method that enables multi\-attribute steering*without any need for training*as explained in the next section\.

## 3Approach

In this section, we present our approach to steer an LLM to generate natural language explanations that are tailored to a target group of explainees, say, to game developers or to philosophers\. To this end, it models the group’s abilities and background in terms of explanatory style and knowledge[Jin et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib22);[Reif et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib41);[Alshomary et al\. \(2021\)](https://arxiv.org/html/2608.29215#bib.bib2)\.

In the following, we detail the three main steps of our approach as illustrated in Figure[2](https://arxiv.org/html/2608.29215#S2.F2): \(1\) extracting style and knowledge attributes of a target group from explanatory texts written by the group; \(2\) creating attribute\-specific activation vectors; and \(3\) computing attribute\-based steering vectors to steer an LLM to generate group\-specific explanations\.

### 3\.1Group\-Specific Attributes

Given a setGGof target groups, we identify their most importantstyle attributesandknowledge attributesfrom text samples written by members of these groups with a two\-stage prompting process\.

For the style attributes, we first reuse the 92 prompts of[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)to let an LLM generate style descriptions of the texts \(see[A\.2](https://arxiv.org/html/2608.29215#A1.SS2)for details\)\. Then, each of these descriptions is rewritten into a list of style attributes \(e\.g\.,The author includes code snippets\.\)\. For the knowledge attributes, we create a new set of 18 prompts to extract six common types of knowledge: factual, conceptual, procedural, metacognitive, situational, and conventional[Bloom et al\. \(1956\)](https://arxiv.org/html/2608.29215#bib.bib6);[de Jong and Ferguson\-Hessler \(1996\)](https://arxiv.org/html/2608.29215#bib.bib13);[Anderson et al\. \(2001\)](https://arxiv.org/html/2608.29215#bib.bib4)\. In addition, we create a new prompt to rewrite each knowledge\-related description into a list of knowledge attributes \(all prompts are found in[A\.3](https://arxiv.org/html/2608.29215#A1.SS3)\)\. Running the two\-stage prompting with these prompts, we obtain lists of knowledge attributes \(e\.g\.,The author understands game development\.\)\.

##### Attribute Selection

For noise reduction, we perform an LLM\-based filtering by classifying each extracted attribute into one of the four classes:*style*,*knowledge*,*both*, or*none*\(details in[A\.4\.1](https://arxiv.org/html/2608.29215#A1.SS4.SSS1)\)\. We keep only those attributes that either belong to style or to knowledge\. This results in candidate style attributes and candidate knowledge attributes\.

On this basis, we select the same set of style attributesF\(S\)F^\{\(S\)\}and the same set of knowledge attributesF\(K\)F^\{\(K\)\}across all target groupsGGto enable comparison across the groups\. To this end, we first cluster semantically\-similar candidate attributes, namely using radius\-based neighbor clustering using a predefined cosine similarity threshold \(see[A\.4\.2](https://arxiv.org/html/2608.29215#A1.SS4.SSS2)for details\)\. The rationale for the clustering is that an attribute not appearing frequently might not only indicate that the corresponding style or knowledge is rarely used to describe the groups, but also that the LLM exhibits high syntactic variability when describing it\. Moreover, to avoid too general attributes, we only keep clusters including attributes that were extracted for a maximum number of groups\. To ensure that the attributes are not too specific, we additionally require a minimum frequency based on how often they were extracted \(see Appendix[A\.4\.3](https://arxiv.org/html/2608.29215#A1.SS4.SSS3)for details\)\.

Finally, the attributes are selected in descending order of their frequency, separately for style and knowledge clusters\. Any attribute that is too similar to a previously selected one is rejected\. Altogether, the outlined process results in the final set of attributesF=\{f1,f2,…,fn\}F=\\\{f\_\{1\},f\_\{2\},\.\.\.,f\_\{n\}\\\}, where eachfj∈F\(S\)∪F\(K\)f\_\{j\}\\in F^\{\(S\)\}\\cup F^\{\(K\)\}\. We do not enforce a fixed numbernn, but instead retain all clusters that remain\.

##### Group\-Specific Attribute Vectors

Given a text written by a member of a specific groupg∈Gg\\in G, we define its attribute vector𝐯g∈\{0,1\}n\\mathbf\{v\}\_\{g\}\\in\\\{0,1\\\}^\{n\}where the valuevj=1v\_\{j\}=1means that attributefjf\_\{j\}was predicted for the text during the two\-stage prompting process\. Given a set of texts, we define the mean𝐯¯g∈\[0,1\]n\\bar\{\\mathbf\{v\}\}\_\{g\}\\in\[0,1\]^\{n\}to be the style and knowledge attribute profile of the group\. In addition, we identify the most important style and knowledge attributes ofgg\.

### 3\.2Attribute\-Specific Activation Vectors

The key idea of our approach is to use the extracted attributes to steer an LLM for group\-specific explanation generation, for which we rely on activation engineering[Subramani et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib48);[Turner et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib50);[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)\. Activation engineering assumes that the information required to generate a target output is already embedded within the hidden layers of an LLM\. Thus, so\-called activation\-based steering vectors can be added to the activations of specific layers during inference to steer the LLM’s output\. In contrast to[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), we do not extract the activation vectors by processing entire group\-specific texts, but we directly use the natural language attributesfj∈Ff\_\{j\}\\in Fto enable more fine\-grained control\. In particular, we extract the activation vectors for themmmost important style attributesFmg\(S\)⊂FF^\{\(S\)\}\_\{m\_\{g\}\}\\subset Fand themmmost important knowledge attributesFmg\(K\)⊂FF^\{\(K\)\}\_\{m\_\{g\}\}\\subset Fidentified for a target groupgg\. However, before, each attributefj∈Fmg\(S\)∪Fmg\(K\)f\_\{j\}\\in F^\{\(S\)\}\_\{m\_\{g\}\}\\cup F^\{\(K\)\}\_\{m\_\{g\}\}is rewritten as follows: Style attributes are expressed in the second\-person perspective \(e\.g\.,You include code snippets\.\), as they function as instruction to the LLM\. In contrast, knowledge attributes are expressed in the first\-person perspective \(e\.g\.,I understand game development\.\), reflecting how users typically communicate such information to an LLM\.

Via a forward pass, we then extract the activation vector𝐚j\+\(i\)\\mathbf\{a\}\_\{j\}^\{\+\(i\)\}based on the rewritten attributes for each layerii\. In addition, we also extract the activation vectors𝐚j−\(i\)\\mathbf\{\{a\}\}\_\{j\}^\{\-\(i\)\}for the negated attribute \(e\.g\.,You do NOT include code snippets\.\)\. The final attribute\-specific activation vector is based on the difference[Turner et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib50):𝐚j\(i\)≔𝐚j\+\(i\)−𝐚j−\(i\)\\mathbf\{a\}\_\{j\}^\{\(i\)\}\\coloneqq\\mathbf\{a\}\_\{j\}^\{\+\(i\)\}\-\\mathbf\{\{a\}\}\_\{j\}^\{\-\(i\)\}\.

### 3\.3Attribute\-Based Steering Vector

To apply a fine\-grained steering based on themmmost important attributesfj∈Fmg\(S\)∪Fmg\(K\)f\_\{j\}\\in F^\{\(S\)\}\_\{m\_\{g\}\}\\cup F^\{\(K\)\}\_\{m\_\{g\}\}of any groupg∈Gg\\in G, we introduce the new notion of*attribute\-based*steering vectors𝐬g\(i\)\\mathbf\{s\}\_\{g\}^\{\(i\)\}: We define𝐬g\(i\)\\mathbf\{s\}\_\{g\}^\{\(i\)\}as the sum of the attribute\-specific activation vectors𝐚j\(i\)\\mathbf\{a\}\_\{j\}^\{\(i\)\}of layerii\. As not all attributes might be equally important for groupgg, we apply a weighting based on the softmax of the attribute profile𝐯¯g\\bar\{\\mathbf\{v\}\}\_\{g\}:

𝐬g\(i\)≔\\displaystyle\\mathbf\{s\}\_\{g\}^\{\(i\)\}\\coloneqq∑fj∈Fmg\(S\)∪Fmg\(K\)softmax​\(𝐯¯g\)j⋅𝐚j\(i\)\\displaystyle\\sum\_\{f\_\{j\}\\in F^\{\(S\)\}\_\{m\_\{g\}\}\\cup F^\{\(K\)\}\_\{m\_\{g\}\}\}\\textrm\{softmax\}\(\\bar\{\\mathbf\{v\}\}\_\{g\}\)\_\{j\}\\cdot\\mathbf\{a\}\_\{j\}^\{\(i\)\}\(1\)
Given an inputxx\(e\.g\., “How does DNA work?”\), we steer an LLM to generate group\-specific explanations by adding𝐬g\(i\)\\mathbf\{s\}\_\{g\}^\{\(i\)\}to the original activation vector𝐚\(i\)\\mathbf\{a\}^\{\(i\)\}for layeriiduring inference\. Theλ\\lambdacontrols the strength of the steering[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23):

𝐚^\(i\)​\(x\)\\displaystyle\\hat\{\\mathbf\{a\}\}^\{\(i\)\}\(x\)≔\\displaystyle\\coloneqq𝐚\(i\)​\(x\)\+λ⋅𝐬g\(i\)\\displaystyle\\mathbf\{a\}^\{\(i\)\}\(x\)\+\\lambda\\cdot\\mathbf\{s\}\_\{g\}^\{\(i\)\}\(2\)

## 4Data

This section details the data utilized in our experiments to extract group\-specific attributes and to generate group\-specific explanations\.

### 4\.1Group\-Specific Attribute Extraction

To extract group\-specific attributes, we only need data containing texts written by members of specific groups\. We rely on Stack Exchange which is a collection of Q&A forums, each covering a specific domain \(e\.g\., game development\)\. For our purposes, we make the simplifying assumption that the answers in a forum of a specific domain are provided by members of the respective group \(e\.g\., game developers\)\. While this assumption will not always hold true, it seems reasonable in general according to our inspection of several samples\. We randomly selected 6,000 answers from each of the eleven groups seen in Figure[3](https://arxiv.org/html/2608.29215#S6.F3)\(details in[A\.1](https://arxiv.org/html/2608.29215#A1.SS1)\)\.

### 4\.2Group\-Specific Explanation Generation

For group\-specific explanations generation, explanatory questions are needed that are general enough to plausibly be asked by diverse groups\. To this end, two types of questions are employed:

##### ELI5 Questions

TheExplain Like I’m 5\(ELI5\) subreddit[Fan et al\. \(2019\)](https://arxiv.org/html/2608.29215#bib.bib16)is a Q&A forum on Reddit providing simplified explanations\. We use the questions of the ELI5 validation split \(1,507\) and test split \(600\) from the KILT benchmark[Petroni et al\. \(2021\)](https://arxiv.org/html/2608.29215#bib.bib38)for our steering experiments\. Since the answers are written for a general audience, they cannot serve as a ground\-truth for our experiments\.

##### Science Questions

[Rooein et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib43)created a set of scientific questions to investigate whether LLMs can adapt their answers to different age and education levels\. We adopt these questions and filter for duplicates resulting in 97 further questions\.

## 5Experiments

We carried out two main experiments to evaluate the adequacy of the extracted attributes and the steering effectiveness of our approach\.111The code and data can be found under[https://github\.com/webis\-de/EMNLP\-26](https://github.com/webis-de/EMNLP-26)\.

### 5\.1Group\-Specific Attribute Extraction

We run the two\-stage prompting process of our approach to identify the most important attributes of each target groupg∈Gg\\in Gwith\|G\|=11\|G\|=11\.

##### Models

For prompting, we employed the LLMQwen2\.5\-7B\-Instruct[Yang et al\. \(2025\)](https://arxiv.org/html/2608.29215#bib.bib56)\. The attribute filtering was performed using the bigger versionQwen2\.5\-72B\-Instruct\. To obtain vector representations of each attribute, we usedQwen3\-Embedding\-8B[Zhang et al\. \(2025b\)](https://arxiv.org/html/2608.29215#bib.bib58)\.

##### Experimental Setup

From the Stack Exchange data, we randomly sample 500 answers per groupggand run the two\-stage prompting process, the LLM\-based filtering and the clustering to then remain with\|F\|=1250\|F\|=1250attributes including 340 style and 910 knowledge attributes \(see[A\.4](https://arxiv.org/html/2608.29215#A1.SS4)for details\)\.

Finally, we create the group\-specific attribute vectors𝐯g∈\{0,1\}1250\\mathbf\{v\}\_\{g\}\\in\\\{0,1\\\}^\{1250\}for each of the 500 answers per group and compute the group profiles𝐯¯g\\bar\{\\mathbf\{v\}\}\_\{g\}\. By computing the point\-biserial correlation[Lev \(1949\)](https://arxiv.org/html/2608.29215#bib.bib27)for all𝐯g\\mathbf\{v\}\_\{g\}, we identify themmmost important style and themmmost important knowledge attributes per group \(see[A\.5](https://arxiv.org/html/2608.29215#A1.SS5)for details\)\. To assess the impact ofmmon the steering effectiveness, we considerm∈\{3,5,10\}m\\in\\\{3,5,10\\\}\. These rather small numbers are motivated by the observation that even humans struggle with considering too many attributes simultaneously[Miller \(1956\)](https://arxiv.org/html/2608.29215#bib.bib31)\.

To investigate the adequacy of the attributes, we compute the Pearson correlation coefficient[Pearson \(1896\)](https://arxiv.org/html/2608.29215#bib.bib37)of the profiles𝐯¯g\\bar\{\\mathbf\{v\}\}\_\{g\}based on the attribute vectors predicted by the model described in Appendix[B\.3](https://arxiv.org/html/2608.29215#A2.SS3)for the 6000 answers per groupgg\.

### 5\.2Group\-Specific Explanation Generation

Given the most important attributes of a specific groupg∈Gg\\in G, we steer an LLM to generate group\-specific explanations\. We evaluate our approach against two baselines:*prompting*and standard*activation\-based steering*[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)\.

##### Models

For our steering experiments, we evaluate three model families of comparable size to ensure a fair comparison:Qwen2\.5\-7B\-Instruct,Llama\-3\.1\-8B\-Instruct[Grattafiori et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib18)andMinistral\-8B\-Instruct\-2410222Model Card,[https://huggingface\.co/mistralai/Ministral\-8B\-Instruct\-2410](https://huggingface.co/mistralai/Ministral-8B-Instruct-2410)\. To further examine how model capacity influences steering effectiveness, we additionally include a larger variant from the Qwen family \(Qwen2\.5\-32B\-Instruct\)\.

##### Experimental Setup

For the automatic hyperparameter tuning of all activation\-based steering methods, we randomly sampled 100 questions from the val\. split of the ELI5 dataset \(same questions for all methods\)\. Based on the best hyperparameters, the final experiment was conducted on the 600 ELI5 test questions and the 97 science questions\.

For the prompting baseline, we designed two system prompts \(full prompts in Appendix[B\.2](https://arxiv.org/html/2608.29215#A2.SS2)\)\. The first prompt includes only the target group \(called*group*below\)\. The second prompt additionally includes themmmost important attributes for that group \(*group \+ attributes \(mm\)*\)\. In all cases, we instructed the LLM to limit its explanations to 200 words as LLMs tend to generate overly long explanations which differs from how humans usually explain[Miller \(2019\)](https://arxiv.org/html/2608.29215#bib.bib32)\. For example, the Stack Exchange data reveals an average length of 168 words\.

The activation\-based steering approaches utilized the same prompts\. In addition, the steering vector𝐬g\(i\)\\mathbf\{s\}\_\{g\}^\{\(i\)\}is added to the activations at each layeriiof a selected set of layersII\. As[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), we select the best performingIIfrom all sets of layersI0=\{0,1,2\},I1=\{1,2,3\},…I\_\{0\}=\\\{0,1,2\\\},I\_\{1\}=\\\{1,2,3\\\},\\ldotsunder a sliding window of size33on the validation set\. We select the bestλ\\lambdafrom\{0\.25,0\.5,…,2\.5\}\\\{0\.25,0\.5,\\ldots,2\.5\\\}and report the selected factorλ\\lambdaand layersIIbelow\.

For the activation\-based steering baseline[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), the steering vector is computed by averaging the activation vectors extracted for the 500 Stack Exchange answers of the target groupggand subtracting the mean activation vector of all other groupsG∖gG\\setminus g\. For our approach, we construct the steering vectors according to Equation[1](https://arxiv.org/html/2608.29215#S3.E1)based on themmmost important attributes of groupgg\.

##### Evaluation Metrics

We automatically evaluate the steering effectiveness in terms of the group*specificity*and the*factuality*of the generated explanations and the harmonic mean as follows:

To assess the group specificity, we train a classifierc:\[0,1\]1250→ℝ11c:\[0,1\]^\{1250\}\\rightarrow\\mathbb\{R\}^\{11\}that takes an attribute vector𝐯∈\[0,1\]1250\\mathbf\{v\}\\in\[0,1\]^\{1250\}as input and predicts a score for every possible target group\. We use this metric also for the hyperparameter tuning \(details in[B\.3](https://arxiv.org/html/2608.29215#A2.SS3)\)\. Given𝐯\\mathbf\{v\}for an explanation tailored to groupgg, the specificity score in\[0,1\]\[0,1\]is obtained by applying the softmax toc⁡\(𝐯\)c\(\\mathbf\{v\}\)and selecting the score corresponding togg\. The score can be interpreted as the probability how much the explanation is tailored:

specificity≔\\displaystyle\\coloneqqsoftmax​\(c⁡\(𝐯\)\)g\\displaystyle\\textrm\{softmax\}\(c\(\\mathbf\{v\}\)\)\_\{g\}\(3\)
For factuality, we employ theFActScoremetric[Min et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib33), which computes the probability of atomic factsXXextracted from input text that are supported by a knowledge source \(see[B\.4](https://arxiv.org/html/2608.29215#A2.SS4)for details\)\.

factuality≔\\displaystyle\\coloneqqFActScore​\(X\)\\displaystyle\\textsc\{FActScore\}\(X\)\(4\)
The metric requires specifying a topic that maps to a Wikipedia title\. We thus report results only for the science questions, as part of the ELI5 questions cannot be mapped to such a title, for example, “Why was quicksand such a common film trope when it’s not a problem \[…\] in real life?”

Since high\-quality explanations should be both well understandable by the explainee and factually accurate[Löfström et al\. \(2022\)](https://arxiv.org/html/2608.29215#bib.bib29), we also consider the harmonic mean to capture the balance between the evaluation metrics \(see[B\.5](https://arxiv.org/html/2608.29215#A2.SS5)for details\)\.

## 6Results and Discussion

In this section, we first analyze the adequacy of the extracted style and knowledge attributes\. Then, we present the findings on the steering effectiveness\.

### 6\.1Attribute Profile Correlation

![Refer to caption](https://arxiv.org/html/2608.29215v1/group-attribute-correlation-num.png)Figure 3:Pearson correlation of attribute profiles in terms of style \(lower triangle, white text\) and knowledge \(upper triangle, black text\) for the 11 target groups\. Bold values denote statistical significance \(p<\.05p<\.05\)\.Table 1:Main automatic evaluation results:*Specificity*,*factuality*, and*harmonic mean*of all tested*LLMs*,*prompt components*, and*steering vectors*on the science questions\([Rooein et al\., 2023](https://arxiv.org/html/2608.29215#bib.bib43)\), along with steering*factor*λ\\lambdaand steered*layers*II\. Bold components/vectors are part of our approach\. Best values per LLM bold, second underlined\.†\\daggerand‡\\ddaggerdenote significant improvements over prompting and[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)per prompt respectively \(p<\.05p<\.05\)To investigate the adequacy of the extracted attributes, Figure[3](https://arxiv.org/html/2608.29215#S6.F3)presents a joint view of how the explanatory style profiles \(lower triangle\) and the knowledge profiles \(upper triangle\) of the eleven target groups correlate in terms of Pearson’srr\.

Style correlations are generally high across scientific and technical groups such as*biologists*and*electrical engineers*\. The knowledge profiles are more differentiated: Some scientific groups, such as*chemists*and*physicists*, also show a high correlation, while others display weaker correlations despite similar style \(e\.g\.,*biologists*and*physicists*\)\.*Computer scientists*,*software engineers*, and*game developers*have a weaker knowledge overlap with natural scientists\. The non\-technical groups \(*historians*,*philosophers*,*politicians*\) form a distinct cluster across both style and knowledge\.

Overall, these results suggest that explanatory style is more homogeneous across all groups, indicating a shared way of structuring explanations\. In contrast, knowledge profiles are more diverse, and non\-technical groups remain distinct from all other groups\. Thus, the created group profiles𝐯¯g\\bar\{\\mathbf\{v\}\}\_\{g\}capture both expected shared explanatory style and meaningful differences in domain\-specific knowledge\.

Table 2:Manual evaluation results:Steering success,plausibility, andhelpfulnessof the Qwen 7B explanations generated by our approach and two baselines, averaged over 50 randomly\-sampled questions, 25 per dataset\. Best values bold\.†\\daggerand‡\\ddaggerdenote significant improvements over prompting and[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)respectively \(p<\.05p<\.05\)\.
### 6\.2Automatic Evaluation of Explanations

Table[1](https://arxiv.org/html/2608.29215#S6.T1)presents the average results on the 97 science questions for three model families, averaged over all eleven target groups, along with the harmonic mean \(HMean\)\. Since the specificity results for the ELI5 questions exhibit similar patterns, we outsource them to Appendix[B\.6](https://arxiv.org/html/2608.29215#A2.SS6)\. Significant improvements over the baselines for the two criteria are determined using the Wilcoxon signed\-rank test for each prompt component respectively\.

Table[1](https://arxiv.org/html/2608.29215#S6.T1)shows including our extracted attributes in the prompt improves the group specificity across all approaches and models\. However, this increase in specificity results in lower factuality\. Our approach generally seems the best midway between specificity and factuality, in case of Qwen also achieving the highest harmonic means \(0\.313 and 0\.299\) and the second best value with 0\.291 for Llama\. Depending on the model, the highest specificity ranges from 0\.511 to 0\.803, all for[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)with our extracted attributes in the prompt\. For Ministral, this even leads to best harmonic mean \(0\.274\)\. The best factuality for all models is achieved by the prompting baseline \(0\.322–0\.376\), which is expected given its lower specificity\.

The ablation results in Appendix[B\.7](https://arxiv.org/html/2608.29215#A2.SS7)show that form∈\{3,10\}m\\in\\\{3,10\\\}there are no performance differences, while a bigger model size can achieve a better balance between specificity and factuality\.

In line with prior works[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23);[Bogdan \(2025\)](https://arxiv.org/html/2608.29215#bib.bib7), the best\-performing layers for all models tend to be around the middle of their respective layer counts\. For Ministral, however, the rather early layersI=\[5,6,7\]I=\[5,6,7\]out of 36 layers were selected for our approach which may explain its limited performance\. While we expect more exhaustive hyperparameter tuning to help, we leave this for future work\.

Overall, the results indicate a tradeoff between generating explanations that are well tailored to a target group and maintaining factual precision\. As an example, a use of analogies as in Figure[1](https://arxiv.org/html/2608.29215#S1.F1)can make explanations more tailored to the target group\. However, this increased specificity can lead to decreased factual accuracy[Wiley et al\. \(2018\)](https://arxiv.org/html/2608.29215#bib.bib55): While the method of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)tends to “oversteer”, resulting in high specificity but low factuality, our approach provides the best balance between the two criteria, which particularly turns out effective for Qwen\. This raises the question of whether the effectiveness of steering for this model is affected by using Qwen for both attribute extraction and explanation generation\. To investigate this, we conduct an ablation study in which Llama and Ministral are used for attribute extraction instead\. The results show that our approach remains robust across different attribute extraction models, suggesting that its effectiveness does not depend on using Qwen \(see Appendix[A\.4\.4](https://arxiv.org/html/2608.29215#A1.SS4.SSS4)for details\)\.

Figure 4:Per\-group harmonic mean of*specificity*and*factuality*of the explanations generated by Qwen \(7B\) for the 97 science questions using the prompting baseline, the method of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)and our approach\. The more outside a point, the higher the harmonic mean\.Figure[4](https://arxiv.org/html/2608.29215#S6.F4)shows the per\-group harmonic mean based on the specificity and factuality of the explanations generated by Qwen on the 97 science questions\. Blue represents the prompting baseline, orange the method of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), and green our approach\. All methods include the target group in the prompt, while our approach additionally uses the five most important attributes\. The prompting baseline performs well for physicists and biologists but the harmonic mean drops to near zero for computer scientists\. Our approach achieves the best harmonic mean for nearly all groups, particularly for scientific and non\-technical groups\.

Table 3:Explanation samples of the two baselines \(prompting and[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)\) and our approach generated by Qwen \(7B\) for the question “How does DNA work?” tailored to*game developers*\. The two baselines use the target group in the prompt and our approach additionally uses the five most important attributes in the prompt\.
### 6\.3Human Evaluation of Explanations

To manually evaluate the steering effectiveness of our approach, we conducted a user study with nine experts \(three biologists, three game developers, and three philosophers\), hired on Upwork\. We randomly sampled 25 science questions[Rooein et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib43)and 25 ELI5 questions\. For each question, the experts are presented only with explanations tailored to their own target group\. We compared three approaches based on Qwen 7B: \(a\) the prompting baseline with the*group*prompt \(representing an out\-of\-the\-box LLM\); \(b\) the approach of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)with the*group*prompt \(representing state\-of\-the\-art steering\); and \(c\) our approach with the five most important attributes in prompt and activitation, since it seemed most robust in automatic evaluation across all LLMs\. We assess thesteering successandplausibilityof the generated explanations on a 5\-point Likert scale from*not at all*\(1\) to*fully*\(5\)\. Beyond steering effectiveness, we also assess the*helpfulness*of the explanations on the same scale\. More details are found in the Appendix[B\.9](https://arxiv.org/html/2608.29215#A2.SS9)\.

Table[2](https://arxiv.org/html/2608.29215#S6.T2)shows the mean and standard deviations for the three criteria each along with the harmonic mean \(HMean\)\. Significant improvements over the baselines are tested with the Wilcoxon signed\-rank test\. Our approach significantly outperforms the two baselines in terms of steering success \(4\.17\) while the prompting baseline yields the best plausibility values \(4\.56\)\. This supports the specificity–factuality tradeoff seen in the automatic evaluation\. The results for[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)again indicate “oversteering” denoted by the lowest helpfulness \(2\.70\) and plausibility \(2\.75\) values\. Overall, our approach achieves the best harmonic mean of 3\.88\.

Manual analysis of the explanations \(complete samples in Appendix[B\.10](https://arxiv.org/html/2608.29215#A2.SS10)\) demonstrates the high steering success of our approach\. For example, as shown in Table[3](https://arxiv.org/html/2608.29215#S6.T3), it adapts the explanation of DNA using analogies that resonate with game developers\. In contrast, the explanation of the prompting baseline is barely tailored to game developers\. While the explanation generated by[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)is entirely code\-based and thus familiar to game developers, it fails to effectively explain DNA\.

Interestingly, higher steering success does not appear to translate into greater helpfulness\. One possible reason may be the inherent subjectivity of the task, as indicated by the relatively low inter\-annotator agreement \(Krippendorff’sα\\alpharanging from 0\.178 to 0\.489\)\. Moreover, we observe no significant difference in the helpfulness scores for the science questions \(see Table[11](https://arxiv.org/html/2608.29215#A2.T11)in appendix\)\. We speculate that the complexity of these questions enables more benefit of group\-specific explanations, while the rather general topics of the ELI5 questions could already be well understood with generic explanations\. We further found that low ratings are partly because of domain shifts without answering the question anymore\. Thus, our findings indicate the need for steering methods to also explicitly model the type of the question to ensure both successfully tailored and helpful explanations\.

Overall, both the automatic and manual evaluations indicate that, when generating group\-specific explanations using state\-of\-the\-art LLMs, there is a tradeoff between tailoring the explanations to a target group and maintaining factuality/plausibility\. Our approach based on the five most important style and knowledge attributes of the target group achieves the best balance between these criteria\.

## 7Conclusion

In this paper, we have presented a three\-step approach to realize group\-specific explanation generation: \(1\) automatically identifying the most important attributes in terms of style and knowledge of a specific target group; \(2\) extracting attribute\-specific activation vectors; in order to \(3\) compute*attribute\-based*steering\-vectors to steer an LLM for group\-specific explanation generation\.

We have evaluated the steering effectiveness of our approach for 11 target groups based on automatic metrics and a human study with nine experts from three domains\. Our results indicate a tradeoff between generating well\-tailored explanations and maintaining factuality\. Overall, our approach achieves the best balance between these criteria, providing evidence for our hypothesis that one can learn from explanations*of*a specific target group how to best explain a topic*for*that group\. Thereby, our work contributes to the understanding of generating group\-specific explanations computationally\. However, we also observed that steering success does not necessarily correlate with the helpfulness of explanations\. In future work, we thus seek to better understand what tailoring is most helpful\.

## 8Limitations

In this paper, we generate group\-specific explanations computationally\. While group\-specific explanations are a possible alternative when individual characteristics are not available, they can only account in a limited way for those individuals whose abilities or background notably deviate from the mean group profile\. However, since our approach relies on predefined attributes, it can easily be extended to steer an LLM based on attributes selected differently, for example, attributes that a user specifies using a slider\.

Furthermore, our experiments are limited to the English language and a fixed set of target groups\. Although the results are quite promising, it remains to be studied how well our approach generalizes to other languages and target groups\.

In general, our notion of*group*is operationalized via Stack Exchange communities, assuming that contributors to a given forum \(e\.g\., game development\) are representative of a target group \(e\.g\., game developers\)\. While this assumption seems reasonable in general, it may not hold universally: Users may participate across domains, vary in expertise, or write answers that do not reflect the typical explanatory style of the respective group\.

Beyond that, our attribute extraction process relies on LLM\-based prompting\. Although this enables scalability and avoids manual feature engineering, it also inherits biases of the underlying model\. However, as our evaluation shows, the created group profiles based on the extracted attributes of the 11 groups appear adequate, capturing both: expected shared explanatory style and meaningful differences in domain\-specific\. Moreover, we maintain that Stack Exchange is a high\-quality data source because of its reputation system, which allows the sites to self\-moderate\. Additionally, since we only consider up\-voted answers, we can reasonably assume that contributing users demonstrate expertise of their respective target groups \(otherwise, they could not provide such helpful answers\)\. Consequently, group overlap should only minimally impact the adequacy of the extracted attributes\.

Finally, we assess factuality of explanations based on theFActScoremetric[Min et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib33)which also relies on LLMs, namely, for fact validation\. Since LLMs often lack factual accuracy due to the hallucination problem[Ji et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib21);[Wang and Shu \(2023\)](https://arxiv.org/html/2608.29215#bib.bib53), we do not judge absolute factuality values but relative differences\. Our evaluation indicates a tradeoff between group specificity and factuality when using state\-of\-the\-art LLMs for group\-specific explanation generation\. Although our approach achieves the best balance between these two criteria, increased specificity can still introduce oversimplifications that reduce factual precision\. We recommend that future methods should focus especially on this tradeoff when generating group\-specific explanations\.

## 9Ethical Considerations

Since our focus is on group\-specific explanations, a primary concern is the risk of stereotyping\. Even when derived empirically, the created group profiles may present an overly\-simplified view of the target group’s knowledge and abilities\. Consequently, the profiles may mix heterogeneous subgroups or overlook important variations within a group\. In addition, since we extract the group\-specific attributes from explanations on Stack Exchange, the generated explanations may reflect existing biases present in the source data\.

Moreover, although our paper focuses on generating tailored explanations and explicitly evaluates factuality, there is a risk to misuse our approach in contexts beyond harmless educational settings\. For example, it could be used in targeted persuasion or to spread misinformation to specific groups\. We cannot fully prevent such misuse, even though it is neither the intended use of our approach, nor is the approach anyhow optimized toward this\.

Finally, we acknowledge the limited sample size and potential subjectivity in rating the steering success, helpfulness, and plausibility in our human study\. Future studies involving broader and more diverse participants could help to further examine the impact of group\-specific explanations\.

## Acknowledgments

This work has been supported by the research project “HybrInt \- Hybrid Intelligence through Interpretable AI in Machine Perception and Interaction” \(Zukunft Nds, Niedersächsisches Ministerium für Wissenschaft, Grant ID: ZN4219\) and by the Deutsche Forschungsgemeinschaft \(DFG, German Research Foundation\) under project number TRR 318/3 2026 – 438445824\. We thank the anonymous reviewers for their insightful feedback\. The writing and implementation were supported by DeepL and ChatGPT but the authors reviewed and, if needed, revised all AI\-assisted content\.

## References

- Aggarwal et al\. \(2021\)Shourya Aggarwal, Divyanshu Mandowara, Vishwajeet Agrawal, Dinesh Khandelwal, Parag Singla, and Dinesh Garg\. 2021\.[Explanations for CommonsenseQA: New Dataset and Models](https://doi.org/10.18653/v1/2021.acl-long.238)\.In*Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\)*, pages 3050–3065, Online\. Association for Computational Linguistics\.
- Alshomary et al\. \(2021\)Milad Alshomary, Wei\-Fan Chen, Timon Gurcke, and Henning Wachsmuth\. 2021\.[Belief\-based generation of argumentative claims](https://doi.org/10.18653/v1/2021.eacl-main.17)\.In*Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume*, pages 224–233, Online\. Association for Computational Linguistics\.
- Alshomary et al\. \(2024\)Milad Alshomary, Felix Lange, Meisam Booshehri, Meghdut Sengupta, Philipp Cimiano, and Henning Wachsmuth\. 2024\.[Modeling the quality of dialogical explanations](https://aclanthology.org/2024.lrec-main.1007)\.In*Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\)*, pages 11523–11536, Torino, Italia\. ELRA and ICCL\.
- Anderson et al\. \(2001\)Lorin W\. Anderson, David R\. Krathwohl, Peter W\. Airasian, Kathleen A\. Cruikshank, Richard E\. Mayer, Paul R\. Pintrich, James Raths, and Merlin C\. Wittrock\. 2001\.*A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom’s Taxonomy of Educational Objectives*\.Longman, New York\.
- Beck et al\. \(2024\)Tilman Beck, Hendrik Schuff, Anne Lauscher, and Iryna Gurevych\. 2024\.[Sensitivity, performance, robustness: Deconstructing the effect of sociodemographic prompting](https://aclanthology.org/2024.eacl-long.159)\.In*Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pages 2589–2615, St\. Julian’s, Malta\. Association for Computational Linguistics\.
- Bloom et al\. \(1956\)Benjamin S\. Bloom, Max D\. Engelhart, Edward J\. Furst, Walker H\. Hill, and David R\. Krathwohl\. 1956\.*Taxonomy of Educational Objectives: The Classification of Educational Goals\. Handbook I: Cognitive Domain*\.David McKay, New York\.
- Bogdan \(2025\)Paul C\. Bogdan\. 2025\.[Emergent effects of scaling on the functional hierarchies within large language models](https://arxiv.org/abs/2501.07359)\.*Preprint*, arXiv:2501\.07359\.
- Camburu et al\. \(2018\)Oana\-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom\. 2018\.[e\-SNLI: Natural language inference with natural language explanations](http://papers.nips.cc/paper/8163-e-snli-natural-language-inference-with-natural-language-explanations.pdf)\.In S\. Bengio, H\. Wallach, H\. Larochelle, K\. Grauman, N\. Cesa\-Bianchi, and R\. Garnett, editors,*Advances in Neural Information Processing Systems 31*, pages 9539–9549\. Curran Associates, Inc\.
- Chen et al\. \(2025\)Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey\. 2025\.[Persona vectors: Monitoring and controlling character traits in language models](https://arxiv.org/abs/2507.21509)\.*Preprint*, arXiv:2507\.21509\.
- Chen et al\. \(2024\)Wei\-Fan Chen, Milad Alshomary, Maja Stahl, Khalid Al Khatib, Benno Stein, and Henning Wachsmuth\. 2024\.[Reference\-guided style\-consistent content transfer](https://aclanthology.org/2024.lrec-main.1201)\.In*Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING 2024\)*, pages 13754–13768, Torino, Italia\. ELRA and ICCL\.
- Cunningham et al\. \(2023\)Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey\. 2023\.[Sparse autoencoders find highly interpretable features in language models](https://arxiv.org/abs/2309.08600)\.*Preprint*, arXiv:2309\.08600\.
- Dathathri et al\. \(2020\)Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu\. 2020\.[Plug and play language models: A simple approach to controlled text generation](https://openreview.net/forum?id=H1edEyBKDS)\.In*International Conference on Learning Representations*\.
- de Jong and Ferguson\-Hessler \(1996\)Ton de Jong and Monica G\.M\. Ferguson\-Hessler\. 1996\.[Types and qualities of knowledge](https://doi.org/10.1207/s15326985ep3102_2)\.*Educational Psychologist*, 31\(2\):105–113\.
- Doddapaneni et al\. \(2024\)Sumanth Doddapaneni, Krishna Sayana, Ambarish Jash, Sukhdeep Sodhi, and Dima Kuzmin\. 2024\.[User embedding model for personalized language prompting](https://aclanthology.org/2024.personalize-1.12/)\.In*Proceedings of the 1st Workshop on Personalization of Generative AI Systems \(PERSONALIZE 2024\)*, pages 124–131, St\. Julians, Malta\. Association for Computational Linguistics\.
- Ester et al\. \(1996\)Martin Ester, Hans\-Peter Kriegel, Jörg Sander, and Xiaowei Xu\. 1996\.A density\-based algorithm for discovering clusters in large spatial databases with noise\.In*Proceedings of the Second International Conference on Knowledge Discovery and Data Mining*, KDD’96, page 226–231\. AAAI Press\.
- Fan et al\. \(2019\)Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli\. 2019\.[ELI5: Long form question answering](https://doi.org/10.18653/v1/P19-1346)\.In*Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics*, pages 3558–3567, Florence, Italy\. Association for Computational Linguistics\.
- Gilpin et al\. \(2018\)Leilani H\. Gilpin, David Bau, Ben Z\. Yuan, Ayesha Bajwa, Michael Specter, and Lalana Kagal\. 2018\.[Explaining explanations: An overview of interpretability of machine learning](http://arxiv.org/abs/1806.00069)\.ArXiv: 1806\.00069\.
- Grattafiori et al\. \(2024\)Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al\-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 4 others\. 2024\.[The Llama 3 herd of models](https://arxiv.org/abs/2407.21783)\.*Preprint*, arXiv:2407\.21783\.
- Huber et al\. \(2025\)Bernd Huber, Ghazal Fazelnia, Andreas Damianou, Sebastian Peleato, Max Lefarov, Praveen Ravichandran, Marco De Nadai, Mounia Lalmas\-Roellke, and Paul N\. Bennett\. 2025\.[Embedding\-to\-prefix: Parameter\-efficient personalization for pre\-trained large language models](https://arxiv.org/abs/2505.17051)\.*Preprint*, arXiv:2505\.17051\.
- Jeck et al\. \(2025\)Jakub Jeck, Florian Leiser, Anne Hüsges, and Ali Sunyaev\. 2025\.[Tell\-me: Toward personalized explanations of large language models](https://doi.org/10.1145/3706599.3719982)\.CHI EA ’25, New York, NY, USA\. Association for Computing Machinery\.
- Ji et al\. \(2023\)Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung\. 2023\.[Survey of hallucination in natural language generation](https://doi.org/10.1145/3571730)\.*ACM Comput\. Surv\.*, 55\(12\)\.
- Jin et al\. \(2022\)Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, and Rada Mihalcea\. 2022\.[Deep learning for text style transfer: A survey](https://doi.org/10.1162/coli_a_00426)\.*Computational Linguistics*, 48\(1\):155–205\.
- Konen et al\. \(2024\)Kai Konen, Sophie Jentzsch, Diaoulé Diallo, Peer Schütt, Oliver Bensch, Roxanne El Baff, Dominik Opitz, and Tobias Hecking\. 2024\.[Style vectors for steering generative large language models](https://aclanthology.org/2024.findings-eacl.52)\.In*Findings of the Association for Computational Linguistics: EACL 2024*, pages 782–802, St\. Julian’s, Malta\. Association for Computational Linguistics\.
- Lage and Ostermann \(2025\)Lucas Fonseca Lage and Simon Ostermann\. 2025\.[Openfactscore: Open\-source atomic evaluation of factuality in text generation](https://arxiv.org/abs/2507.05965)\.*Preprint*, arXiv:2507\.05965\.
- Lai et al\. \(2024\)Wen Lai, Viktor Hangya, and Alexander Fraser\. 2024\.[Style\-specific neurons for steering LLMs in text style transfer](https://doi.org/10.18653/v1/2024.emnlp-main.745)\.In*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing*, pages 13427–13443, Miami, Florida, USA\. Association for Computational Linguistics\.
- Lee et al\. \(2023\)Joosung Lee, Minsik Oh, and Donghun Lee\. 2023\.[P5: Plug\-and\-play persona prompting for personalized response selection](https://doi.org/10.18653/v1/2023.emnlp-main.1031)\.In*Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing*, pages 16571–16582, Singapore\. Association for Computational Linguistics\.
- Lev \(1949\)Joseph Lev\. 1949\.[The Point Biserial Coefficient of Correlation](https://doi.org/10.1214/aoms/1177730103)\.*The Annals of Mathematical Statistics*, 20\(1\):125 – 126\.
- Lyu et al\. \(2024\)Qing Lyu, Marianna Apidianaki, and Chris Callison\-Burch\. 2024\.[Towards faithful model explanation in NLP: A survey](https://doi.org/10.1162/coli_a_00511)\.*Computational Linguistics*, 50\(2\):657–723\.
- Löfström et al\. \(2022\)Helena Löfström, Karl Hammar, and Ulf Johansson\. 2022\.[*A Meta Survey of Quality Evaluation Criteria in Explanation Methods*](https://doi.org/10.1007/978-3-031-07481-3_7), page 55–63\.Springer International Publishing\.
- Meng et al\. \(2024\)Tao Meng, Ninareh Mehrabi, Palash Goyal, Anil Ramakrishna, Aram Galstyan, Richard Zemel, Kai\-Wei Chang, Rahul Gupta, and Charith Peris\. 2024\.[Attribute controlled fine\-tuning for large language models: A case study on detoxification](https://doi.org/10.18653/v1/2024.findings-emnlp.779)\.In*Findings of the Association for Computational Linguistics: EMNLP 2024*, pages 13329–13341, Miami, Florida, USA\. Association for Computational Linguistics\.
- Miller \(1956\)George A\. Miller\. 1956\.[The magical number seven, plus or minus two: Some limits on our capacity for processing information\.](https://doi.org/10.1037/h0043158)*Psychological Review*, 63\(2\):81–97\.
- Miller \(2019\)Tim Miller\. 2019\.[Explanation in artificial intelligence: Insights from the social sciences](https://doi.org/10.1016/j.artint.2018.07.007)\.*Artificial Intelligence*, 267:1–38\.
- Min et al\. \(2023\)Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen\-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi\. 2023\.[FActScore: Fine\-grained atomic evaluation of factual precision in long form text generation](https://doi.org/10.18653/v1/2023.emnlp-main.741)\.In*Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing*, pages 12076–12100, Singapore\. Association for Computational Linguistics\.
- Nguyen et al\. \(2025\)Duy Nguyen, Archiki Prasad, Elias Stengel\-Eskin, and Mohit Bansal\. 2025\.[Multi\-attribute steering of language models via targeted intervention](https://doi.org/10.18653/v1/2025.acl-long.1007)\.In*Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pages 20619–20634, Vienna, Austria\. Association for Computational Linguistics\.
- Oozeer et al\. \(2025\)Narmeen Fatimah Oozeer, Luke Marks, Fazl Barez, and Amir Abdullah\. 2025\.[Beyond linear steering: Unified multi\-attribute control for language models](https://doi.org/10.18653/v1/2025.findings-emnlp.1278)\.In*Findings of the Association for Computational Linguistics: EMNLP 2025*, pages 23513–23557, Suzhou, China\. Association for Computational Linguistics\.
- Patel et al\. \(2023\)Ajay Patel, Delip Rao, Ansh Kothary, Kathleen McKeown, and Chris Callison\-Burch\. 2023\.[Learning interpretable style embeddings via prompting LLMs](https://doi.org/10.18653/v1/2023.findings-emnlp.1020)\.In*Findings of the Association for Computational Linguistics: EMNLP 2023*, pages 15270–15290, Singapore\. Association for Computational Linguistics\.
- Pearson \(1896\)Karl Pearson\. 1896\.[Vii\. mathematical contributions to the theory of evolution\.—iii\. regression, heredity, and panmixia](https://doi.org/10.1098/rsta.1896.0007)\.*Philosophical Transactions of the Royal Society of London, Series A: Containing Papers of a Mathematical or Physical Character*, \(187\):253–318\.
- Petroni et al\. \(2021\)Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Rocktäschel, and Sebastian Riedel\. 2021\.[KILT: a benchmark for knowledge intensive language tasks](https://doi.org/10.18653/v1/2021.naacl-main.200)\.In*Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pages 2523–2544, Online\. Association for Computational Linguistics\.
- Prabhumoye et al\. \(2019\)Shrimai Prabhumoye, Chris Quirk, and Michel Galley\. 2019\.[Towards content transfer through grounded text generation](https://doi.org/10.18653/v1/N19-1269)\.In*Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long and Short Papers\)*, pages 2622–2632, Minneapolis, Minnesota\. Association for Computational Linguistics\.
- Rajani et al\. \(2019\)Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher\. 2019\.[Explain yourself\! Leveraging language models for commonsense reasoning](https://doi.org/10.18653/v1/P19-1487)\.In*Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics*, pages 4932–4942, Florence, Italy\. Association for Computational Linguistics\.
- Reif et al\. \(2022\)Emily Reif, Daphne Ippolito, Ann Yuan, Andy Coenen, Chris Callison\-Burch, and Jason Wei\. 2022\.[A recipe for arbitrary text style transfer with large language models](https://doi.org/10.18653/v1/2022.acl-short.94)\.In*Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\)*, pages 837–848, Dublin, Ireland\. Association for Computational Linguistics\.
- Rong et al\. \(2024\)Yao Rong, Tobias Leemann, Thai\-Trang Nguyen, Lisa Fiedler, Peizhu Qian, Vaibhav Unhelkar, Tina Seidel, Gjergji Kasneci, and Enkelejda Kasneci\. 2024\.[Towards human\-centered explainable AI: A survey of user studies for model explanations](https://doi.org/10.1109/TPAMI.2023.3331846)\.*IEEE Trans\. Pattern Anal\. Mach\. Intell\.*, 46\(4\):2104–2122\.
- Rooein et al\. \(2023\)Donya Rooein, Amanda Cercas Curry, and Dirk Hovy\. 2023\.[Know your audience: Do LLMs adapt to different age and education levels?](https://arxiv.org/abs/2312.02065)*Preprint*, arXiv:2312\.02065\.
- Schneider and Handali \(2019\)Johannes Schneider and Joshua Peter Handali\. 2019\.[Personalized explanation for machine learning: A conceptualization](https://aisel.aisnet.org/ecis2019_rp/171)\.In*Proceedings of the 27th European Conference on Information Systems \(ECIS\)*, Stockholm & Uppsala, Sweden\.
- Siskou and Hautli\-Janisz \(2025\)Wassiliki Siskou and Annette Hautli\-Janisz\. 2025\.[Exploring the limits of prompting llms with speaker\-specific rhetorical fingerprints](https://aclanthology.org/2025.lm4dh-1.14)\.In*Proceedings of the First on Natural Language Processing and Language Models for Digital Humanities*, pages 143–154, Varna, Bulgaria\. INCOMA Ltd\., Shoumen, Bulgaria\.
- Sokol and Flach \(2020\)Kacper Sokol and Peter Flach\. 2020\.[One explanation does not fit all](https://doi.org/10.1007/s13218-020-00637-y)\.*KI \- Künstliche Intelligenz*, 34\(2\):235–250\.
- Stahl et al\. \(2024\)Maja Stahl, Leon Biermann, Andreas Nehring, and Henning Wachsmuth\. 2024\.[Exploring LLM prompting strategies for joint essay scoring and feedback generation](https://aclanthology.org/2024.bea-1.23/)\.In*Proceedings of the 19th Workshop on Innovative Use of NLP for Building Educational Applications \(BEA 2024\)*, pages 283–298, Mexico City, Mexico\. Association for Computational Linguistics\.
- Subramani et al\. \(2022\)Nishant Subramani, Nivedita Suresh, and Matthew Peters\. 2022\.[Extracting latent steering vectors from pretrained language models](https://doi.org/10.18653/v1/2022.findings-acl.48)\.In*Findings of the Association for Computational Linguistics: ACL 2022*, pages 566–581, Dublin, Ireland\. Association for Computational Linguistics\.
- Tausczik and Pennebaker \(2010\)Yla R\. Tausczik and James W\. Pennebaker\. 2010\.[The psychological meaning of words: Liwc and computerized text analysis methods](https://doi.org/10.1177/0261927X09351676)\.*Journal of Language and Social Psychology*, 29\(1\):24–54\.
- Turner et al\. \(2023\)Alexander Matt Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid\. 2023\.[Activation addition: Steering language models without optimization](https://arxiv.org/abs/2308.10248)\.*Preprint*, arXiv:2308\.10248\.
- Wachsmuth and Alshomary \(2022\)Henning Wachsmuth and Milad Alshomary\. 2022\.[“Mama always had a way of explaining things so I could understand”: A dialogue corpus for learning to construct explanations](https://aclanthology.org/2022.coling-1.27)\.In*Proceedings of the 29th International Conference on Computational Linguistics*, pages 344–354, Gyeongju, Republic of Korea\. International Committee on Computational Linguistics\.
- Wang et al\. \(2020\)Cunxiang Wang, Shuailong Liang, Yili Jin, Yilong Wang, Xiaodan Zhu, and Yue Zhang\. 2020\.[SemEval\-2020 Task 4: Commonsense validation and explanation](https://doi.org/10.18653/v1/2020.semeval-1.39)\.In*Proceedings of the Fourteenth Workshop on Semantic Evaluation*, pages 307–321, Barcelona \(online\)\. International Committee for Computational Linguistics\.
- Wang and Shu \(2023\)Haoran Wang and Kai Shu\. 2023\.[Explainable claim verification via knowledge\-grounded reasoning with large language models](https://doi.org/10.18653/v1/2023.findings-emnlp.416)\.In*Findings of the Association for Computational Linguistics: EMNLP 2023*, pages 6288–6304, Singapore\. Association for Computational Linguistics\.
- Wang et al\. \(2025\)Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma\. 2025\.[Adaptive activation steering: A tuning\-free llm truthfulness improvement method for diverse hallucinations categories](https://doi.org/10.1145/3696410.3714640)\.In*Proceedings of the ACM on Web Conference 2025*, WWW ’25, page 2562–2578, New York, NY, USA\. Association for Computing Machinery\.
- Wiley et al\. \(2018\)Jennifer Wiley, Allison J\. Jaeger, Andrew R\. Taylor, and Thomas D\. Griffin\. 2018\.[When analogies harm: The effects of analogies on metacomprehension](https://doi.org/10.1016/j.learninstruc.2017.10.001)\.*Learning and Instruction*, 55:113–123\.
- Yang et al\. \(2025\)An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 23 others\. 2025\.[Qwen2\.5 technical report](https://arxiv.org/abs/2412.15115)\.*Preprint*, arXiv:2412\.15115\.
- Zhang et al\. \(2025a\)Haobo Zhang, Qiannan Zhu, and Zhicheng Dou\. 2025a\.[A unified prompt\-aware framework for personalized search and explanation generation](https://doi.org/10.1145/3716131)\.*ACM Trans\. Inf\. Syst\.*, 43\(3\)\.
- Zhang et al\. \(2025b\)Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou\. 2025b\.[Qwen3 embedding: Advancing text embedding and reranking through foundation models](https://arxiv.org/abs/2506.05176)\.*Preprint*, arXiv:2506\.05176\.
- Zytek et al\. \(2024\)Alexandra Zytek, Sara Pido, Sarah Alnegheimish, Laure Berti\-Équille, and Kalyan Veeramachaneni\. 2024\.[Explingo: Explaining ai predictions using large language models](https://doi.org/10.1109/BigData62323.2024.10825114)\.In*2024 IEEE International Conference on Big Data \(BigData\)*, pages 1197–1208\.

## Appendix AGroup\-Specific Attribute Extraction

### A\.1Stack Exchange Dataset

To extract group\-specific attributes, we rely on English Stack Exchange data\.333Stack Exchange,[https://stackexchange\.com](https://stackexchange.com/)Stack Exchange is a collection of Q&A forums, each covering a specific domain \(e\.g\., game development or philosophy\)\. We used the official Stack Exchange Data Dump444Stack Exchange Data Dump,[https://archive\.org/details/stackexchange\_20250630](https://archive.org/details/stackexchange_20250630)of all Stack Exchange domains \(as of June 30, 2024\) and filter for explanatory questions based on the questions wordswhat,how,why[Miller \(2019\)](https://arxiv.org/html/2608.29215#bib.bib32)\. For our purposes, we make the simplifying assumption that the answers in a forum of a specific domain are provided by members of the respective group \(e\.g\., game developers\)\. Note, our method does not depend on Stack Exchange data specifically\. The underlying data can easily be exchanged to enable steering for other groups\. We only need texts written by members of a specific group\.

To ensure high\-quality data, we filtered the answers based on the following criteria: \(1\) a minimum length of 50 words, to consider detailed explanations\. \(2\) a maximum tokenized length of 512 tokens since this is the max length that the modeldeberta\-v3\-largecan handle \(see Appendix[B\.3](https://arxiv.org/html/2608.29215#A2.SS3)\), and \(3\) at least one upvote, as a minimum guarantee that anyone finds the answer helpful\.

Afterwards, to ensure a balanced distribution, we consider only Stack Exchange forums that have at least 6000 answers and randomly sample 6000 answers each from the resulting eleven groups:

> biologists, chemists, computer scientists, electrical engineers, game developers, historians, philosophers, physicists, pilots, politicians, software engineers

### A\.2Style Attribute Prompts

[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)implemented an unsupervised method to extract style attributes in order to create style representations of author style in text\. For this, they introduce a two\-stage prompting\. For the first stage, they created six open\-ended prompts that generate descriptions of a text on a broad dimension of style\. For example, this is the open\-ended prompt to get a description of the unique grammar style of the author:

> Write a long paragraph describing theunique grammar styleof the following passage without referring to specifics about the topic\. Passage: \.\.\. Description:

In addition, they created 87 targeted prompts to generate descriptions about specific dimensions of style derived from linguistic and psychological categories[Tausczik and Pennebaker \(2010\)](https://arxiv.org/html/2608.29215#bib.bib49)\. For example, this is the targeted prompt to get a description if the author uses any figurative language\. We added “If not, keep the answer short\.” since in our experiments the LLM hallucinated otherwise:

> Write a description of whether the author of the following passage has anyfigurative language\. If not, keep the answer short\. Passage: \.\.\. Description:

This results in a total of 93 prompts\. We adopt 92 of these prompts \(6 open\-ended prompts, 86 targeted prompts\) since one targeted prompt regarding swear words was created twice\.

For the second stage,[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)created a prompt to rewrite the descriptions to natural language style attributes beginning with “The author …”\. We added “These sentences have to be short\. Avoid examples\. Avoid negations\. Use only the information that is present in the description\.” since in our experiments the LLM hallucinated otherwise:

> Here’s a description of an author’s writing style for a passage: \.\.\. Rewrite this description as a long list of short sentences describing the author’s writing style where each sentence is in the format of ’The author is X\.’ or ’The author uses X\.’\. These sentences have to be short\. Avoid examples\. Avoid negations\. Use only the information that is present in the description\. Output:

After running this two\-stage prompting on texts written by 1,000 different authors, they extracted nearly 1\.3M style attributes\. For the final style vector, they selected only 768 attributes\. The first 87 attributes correspond to the style attributes extracted by the 87 targeted prompts\. The remaining 681 are downselected from the remaining style attributes based on filtering heuristics \(frequency, similarity, interpretability\)\.

### A\.3Knowledge Attribute Prompts

We extend the two\-stage prompting process of[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)by creating 18 additional open\-ended prompts to extract knowledge attributes from a given explanation\. We create prompts to get descriptions about six common types of knowledge[Bloom et al\. \(1956\)](https://arxiv.org/html/2608.29215#bib.bib6);[de Jong and Ferguson\-Hessler \(1996\)](https://arxiv.org/html/2608.29215#bib.bib13);[Anderson et al\. \(2001\)](https://arxiv.org/html/2608.29215#bib.bib4)\.

For each type of knowledge, we create a general prompt\. For example, this is the general prompt to extract the factual knowledge of the author:

> Write a long paragraph describing thefactual knowledgepresented in the following passage\.

In addition, we create more specific prompts related to this type of knowledge:

> Write a long paragraph describing theterminologypresented in the following passage\.

> Write a long paragraph describing thespecific factspresented in the following passage\.

We use the following prompts to extract the conceptual knowledge of the author:

> Write a long paragraph describing theconceptual knowledge presentedin the following passage\.

> Write a long paragraph describing theclassifications and categoriespresented in the following passage\.

> Write a long paragraph describing theprinciples and generalizationspresented in the following passage\.

> Write a long paragraph describing thetheories, models, and structurespresented in the following passage\.

> Write a long paragraph describing thetrends and sequencespresented in the following passage\.

To extract the procedural knowledge, we utilize the following prompts:

> Write a long paragraph describing theprocedural knowledgepresented in the following passage\.

> Write a long paragraph describing thesubject\-specific skills and algorithmspresented in the following passage\.

> Write a long paragraph describing thesubject\-specific techniques and methodspresented in the following passage\.

> Write a long paragraph describing thecriteria for selecting and applying appropriate procedurespresented in the following passage\.

To extract the metacognitive knowledge, we created the following prompts:

> Write a long paragraph describing themetacognitive knowledgepresented in the following passage\.

> Write a long paragraph describing thestrategic knowledgepresented in the following passage\.

> Write a long paragraph describing thecognitive tasks, including relevant contextual and conditional knowledge, presented in the following passage\.

> Write a long paragraph describing theself\-knowledgepresented in the following passage\.

To get descriptions about the conventional and sitiational knowledge of the author, we created the following prompts:

> Write a long paragraph describing theconventionspresented in the following passage\.

> Write a long paragraph describing thesituational knowledgepresented in the following passage\.

Following[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36), to rewrite the knowledge descriptions into lists of knowledge attributes, we designed a new rewrite prompt\. Since for our approach it is important that the sentences are understandable on its own, we include an additional instruction:

> Here’s a description of an author’s knowledge: \.\.\. Rewrite this description as a long list of short sentences describing the author’s knowledge where each sentence is in the format of ’The author knows X\.’ or ’The author understands X\.’ These sentences have to be short\. Each sentence must be understandable on its own, without the need for context from the other sentences\. Avoid unnecessary detail\. Avoid examples\. Avoid negations\. Use only the information that is present in the description\. Output:

### A\.4Attribute Extraction

We run the two\-stage prompting process of our approach to extract style and knowledge attributes\. We used a fixed random seed to ensure reproducibility\. The prompts can be found in Appendix[A\.2](https://arxiv.org/html/2608.29215#A1.SS2)and[A\.3](https://arxiv.org/html/2608.29215#A1.SS3)\. For this, we randomly sampled 500 answers per group from the eleven target groups \(i\.e\., 5500 in total\) from the Stack Exchange data \(see Appendix[A\.1](https://arxiv.org/html/2608.29215#A1.SS1)\)\. This way, we keep the total number of inference steps comparable to[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)who used 10k posts \(10 per author\), since we increase the total number of prompts to 110\. In particular, for the two\-stage prompting, our setup results in 1\.21M inference steps, compared to 1\.86M for[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)\. We got 605k lists of attributes from the two\-stage prompting process\.

To ensure comparability across groups, we select a final shared set of style attributesF\(S\)F^\{\(S\)\}and knowledge attributesF\(K\)F^\{\(K\)\}\. For this, we apply the LLM\-based filtering to keep only attributes that either belong to style or knowledge \(see Appendix[A\.4\.1](https://arxiv.org/html/2608.29215#A1.SS4.SSS1)for details\)\. This resulted in about 190k unique candidate style attributes and 1\.4M unique candidate knowledge attributes\. Next, we clustered the attributes based on a cosine similarity threshold of 0\.85 \(see Appendix[A\.4\.2](https://arxiv.org/html/2608.29215#A1.SS4.SSS2)for details\), resulting in about 31k style attribute clusters and about 470k knowledge attribute clusters\. We filtered out clusters for which the representative attribute was predicted for more than seven groups\. Moreover, we filtered out clusters when less then six answers were annotated with the representative attribute \(see Appendix[A\.4\.3](https://arxiv.org/html/2608.29215#A1.SS4.SSS3)for details\)\. After that, for the final set of attributesFF, we select the 86 clusters related to the targeted style prompts\. From the remaining clusters, we select the representative attribute based on how often it was extracted for the Stack Exchange answers\. Once an attribute is selected to be part ofFF, we do not select another attribute with a cosine similarity bigger than 0\.7 to avoid too similar duplicates\. We then remain with\|F\|=1250\|F\|=1250attributes withF=F\(S\)∪F\(K\)F=F^\{\(S\)\}\\cup F^\{\(K\)\}including\|F\(S\)\|=340\|F^\{\(S\)\}\|=340style attributes and\|F\(K\)\|=910\|F^\{\(K\)\}\|=910knowledge attributes\.

#### A\.4\.1LLM\-Based Filtering

To ensure that each extracted attribute describes either the writing style or the knowledge of an author, we apply a LLM\-based filtering\. In particular, we run a two\-stage prompting\. We used a fixed random seed to ensure reproducibility\.

First, we prompt the LLM to classify each attribute into the classes*knowledge*,*style*,*both*or*none*:

> You are an expert linguist specializing in authorship analysis\. Your task: Given a sentence about an author, classify it into exactly one of the following four categories: \*\*knowledge\*\*: The sentence describes what the author knows, including their skills applied\. \*\*style\*\*: The sentence describes how the author writes, including aspects such as spelling, grammar, punctuation, the vocabulary used, or the sentence and paragraph structure\. \*\*both\*\*: The sentence could describe both the author’s knowledge or their writing style\. \*\*none\*\*: The sentence does NOT describe the author’s knowledge or writing style\.

To ensure that the LLM predicts only one of the four classes, we use guided decoding implemented in the outlines library\.555[https://github\.com/dottxt\-ai/outlines](https://github.com/dottxt-ai/outlines)

Second, we reclassify all attributes that were labeled with*both*and reassign each one to the label*knowledge*or*style*that best fits\.

> You are an expert linguist specializing in authorship analysis\. Your task: Given a sentence about an author that could refer to either the author’s knowledge or their writing style\. Assign it to the single category it most accurately belongs to: \*\*knowledge\*\*: The sentence describes what the author knows, including their skills applied\. \*\*style\*\*: The sentence describes how the author writes, including aspects such as spelling, grammar, punctuation, the vocabulary used, or the sentence and paragraph structure\.

After running the two\-stage prompting, we filter out attributes labeled with*none*\. We retain attributes labeled with*knowledge*only if they were extracted using a knowledge attribute prompt\. Similarly, we retain attributes labeled with*style*only if they were extracted using a style attribute prompt\.

#### A\.4\.2Clustering

To cluster semantically similar attributes, we perform a radius\-based neighbor clustering in the embedding space based on[Ester et al\. \(1996\)](https://arxiv.org/html/2608.29215#bib.bib15)\. Each embedded attribute is treated as a node in a graph and connected to all other attributes within a radius defined by a cosine similarity threshold\. Instead of forming clusters via full graph connectivity[Ester et al\. \(1996\)](https://arxiv.org/html/2608.29215#bib.bib15), cluster centers are identified through a greedy nearest\-neighbor descent: Starting from each node, we repeatedly move to the most similar neighboring node within the radius until convergence\. All attributes that converge to the same center are grouped into a cluster\. This approach produces variable\-sized clusters that adapt to local similarity without requiring a predefined number of clusters\.

#### A\.4\.3Frequency\-Based Filtering

To increase the representativeness of the clusters, we filter out clusters based on two frequency\-based criteria\. First, we filter the representative attributes of the clusters based on a maximum number of groups for which they were extracted to avoid too general attributes, e\.g\.,The author uses words\.We follow the filtering heuristics of[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)and set the maximum number to0\.6⋅11≈70\.6\\cdot 11\\approx 7\. In addition, to ensure that the attributes are not overly specialized, we also take into account how frequently each attribute was extracted\. To determine an appropriate value, we start with 1 and gradually increase it until the resulting number of clusters is at most 10,000\. This threshold is chosen to balance granularity and generalization: It is large enough to capture meaningful distinctions between attributes, but small enough to avoid overly fragmented clusters that would be sparse or noisy\. Following this procedure, we arrive at a final value of six\. Thus, an attribute is selected only if it was extracted for at least six answers\.

#### A\.4\.4Ablation Study

In the steering experiments, we use Qwen \(7B\) for both attribute extraction and explanation generation\. To analyze whether this choice affects steering effectiveness of this model, we conduct an ablation study in which Llama \(8B\) and Ministral \(8B\) are used for the attribute extraction\.

Executing the entire generation pipeline, including the hyperparameter tuning of the layersIIand the factorλ\\lambda, would require a significant amount of computational resources\. Therefore, we only compute the steering vectors derived from attributes extracted by Llama and Ministral \(we follow the same procedure as described in[A\.4](https://arxiv.org/html/2608.29215#A1.SS4)\)\. We then compare these vectors with the original steering vectors derived from Qwen\-based attributes\. This comparison already provides insight into the robustness of the steering, as highly similar steering vectors are expected to induce comparable hidden\-state modifications and, consequently, similar generations\.

Table[4](https://arxiv.org/html/2608.29215#A1.T4)reports the cosine similarity between steering vectors obtained from Llama\- or Ministral\-based attributes and the original Qwen\-based steering vectors\. Across all evaluated models, the steering directions remain highly aligned, with cosine similarities consistently above 0\.8 and reaching up to 0\.96\. Notably, for Qwen, the steering vectors derived from attributes extracted by Llama and Ministral remain highly similar to the original steering vectors based on the attributes extracted by Qwen \(0\.9 and 0\.85\)\. This suggests that the strong steering performance observed when using Qwen for generating the explanations \(see Table[1](https://arxiv.org/html/2608.29215#S6.T1)and[2](https://arxiv.org/html/2608.29215#S6.T2)\) is not a consequence of Qwen extracting the attributes itself\.

Overall, the results indicates that our method is robust to variations in attribute extraction and does not critically depend on using Qwen\. Instead, the steering directions appears to be highly similar across different LLMs\.

Table 4:Cosine similarity of the steering vectors based on the attributes extracted by Llama \(Llama\-based\) or Ministral \(Ministral\-based\) and the original steering vectors based on the attributes extracted by Qwen\. The values represent the cosine similarity averaged across all steered layersIIand all 11 target groups, as well as the standard deviation\.

### A\.5Most Important Attributes

To identify themmmost important style and themmmost important knowledge attributes per groupgg, we compute the point\-biserial correlation[Lev \(1949\)](https://arxiv.org/html/2608.29215#bib.bib27)for all attribute vectors𝐯g∈\{0,1\}1250\\mathbf\{v\}\_\{g\}\\in\\\{0,1\\\}^\{1250\}created for each of the 500 answers from Stack Exchange during the two\-stage prompting process\. We also considered to identify the most important attributes based on the attributes vectors predicted by the model described in Appendix[B\.3](https://arxiv.org/html/2608.29215#A2.SS3)for the 6000 answers per group \. However, the most important attributes identified based on these attribute vectors seem to be less representative than those based on the 500 answers\. One reason for this could be that predicting the attribute vectors for the 6000 answers using the model is impacted by the inherent generalization error\. While this does not have a big effect when considering the whole attribute vector, the attribute\-specific quality is lower\. This is why we identify the most important attributes based on the 500 answers per group\. Table[5](https://arxiv.org/html/2608.29215#A1.T5)shows the most important style attributes form=5m=5per group ordered by the correlation values\. Table[6](https://arxiv.org/html/2608.29215#A1.T6)lists the top five most important knowledge attributes for for each group, also ranked by their correlation values\.

Table 5:Top five most important style attributes \(by point\-biserial correlation\) for each group based on the attribute vectors𝐯g∈\{0,1\}1250\\mathbf\{v\}\_\{g\}\\in\\\{0,1\\\}^\{1250\}created for the 500 answers per group from Stack Exchange\.Table 6:Top five most important knowledge attributes \(by point\-biserial correlation\) for each group based on the attribute vectors𝐯g∈\{0,1\}1250\\mathbf\{v\}\_\{g\}\\in\\\{0,1\\\}^\{1250\}created for the 500 answers per group from Stack Exchange\.

## Appendix BGroup\-Specific Explanation Generation

### B\.1GPU usage

We conducted all experiments on a shared cluster equipped with 16 H200 and 8 A100, with a total compute cost of approximately 200 GPU hours\.

### B\.2Prompting Baseline

In the experiments, we compare our approach against a prompting baseline\. We investigate two different system prompt settings\. Since the Stack Exchange explanations from our dataset contain around 168 words on average, we instructed the LLM to limit its responses to 200 words\.

The first system prompt includes only the target group \(e\.g\., game developers\):

> You are a helpful assistant specialized in answering explanatory questions asked bygame developers\. Explain in a way thatgame developerscan effectively understand the explanations\. Your explanations have to be shorter than 200 words\.

The second prompt additionally includes themmmost important style attributes and themmmost important knowledge attributes for that group\. We rewrite the extracted attributes \(see Section[3\.1](https://arxiv.org/html/2608.29215#S3.SS1)\) so that it fits into the prompt \(e\.g\.,The author includes code snippets\.→\\rightarrowThey include code snippets\.\):

> Take into account thatgame developershave the following explanatory style: They include code snippets\. \.\.\. Take into account thatgame developersusually have the following knowledge: They understand game development\. \.\.\. You are a helpful assistant specialized in answering explanatory questions asked bygame developers\. Explain in a way thatgame developerscan effectively understand the explanations\. Your explanations have to be shorter than 200 words\.

### B\.3Specificity Metric

To measure the group specificity of a generated explanation, we train a classifierc:\[0,1\]1250→ℝ11c:\[0,1\]^\{1250\}\\rightarrow\\mathbb\{R\}^\{11\}that predicts a value for each target group based on the attribute vector created for that explanation\. For this, we follow[Patel et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib36)by first training a model \(SFAM\) based ondeberta\-v3\-largethat takes an input text and predicts a probability value for each attributefj∈Ff\_\{j\}\\in Fhow likely the attribute is present in the text\. To train and evaluate this model, we split the Stack Exchange dataset into 400 train samples, 50 validation samples and 50 test samples per group\. This model is then trained on the attribute vectors𝐯∈\{0,1\}1250\\mathbf\{v\}\\in\\\{0,1\\\}^\{1250\}created for the 400 answers per groupggduring the two\-stage prompting process\. The test accuracy of that model is0\.8760\.876\. With this model, we extend our Stack Exchange dataset by constructing the attribute vectors𝐯∈\[0,1\]1250\\mathbf\{v\}\\in\[0,1\]^\{1250\}for all 6000 answers per group\. We again split the data into 5500 train samples, 300 validation samples and 200 test samples per group\. Based on these data, we train another model \(LISA\) based ondeberta\-v3\-largethat directly maps an input text to an attribute vector𝐯∈\[0,1\]1250\\mathbf\{v\}\\in\[0,1\]^\{1250\}\. It enables the fast computation of an attribute vector for a generated explanation\. This model achieves a mean squared error of0\.04370\.0437on the test set\. Finally, we train the classifier on the 6000 attribute vectors per group \(5500/300/200\)\. The classifier achieves a test accuracy of0\.8520\.852on the test set\. Given a generated explanation tailored to a specific group and the corresponding attribute vector𝐯\\mathbf\{v\}, thes​p​e​c​i​f​i​c​i​t​y∈\[0,1\]specificity\\in\[0,1\]is obtained by applying the softmax to the classifier output and selecting the component corresponding to the group\. For the test set, the classifier achieves an averages​p​e​c​i​f​i​c​i​t​yspecificityscore of0\.8010\.801\.

The group specificity serves also as the validation metric to find the best values for the hyperparametersIIandλ\\lambdaon the validation set\. However, since activation\-based steering leads the LLM to generate endless repetitive texts whenλ\\lambdais chosen too high[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), we additionally include a binary repetition penaltyrrand length penaltyll\. In addition, we check that the generated explanations contain only Latin\-1 characters since we want to generate explanations in English\. Since we prioritize fluent explanations, we weighted the repetition penalty twice as much as the length penalty for obtainings​p​e​c​i​f​i​c​i​t​yv​a​l∈\[−3,1\]specificity\_\{val\}\\in\[\-3,1\]:

s​p​e​c​i​f​i​c​i​t​yv​a​l≔softmax​\(c⁡\(𝐯\)\)g−2⋅r−lspecificity\_\{val\}\\coloneqq\\textrm\{softmax\}\(c\(\\mathbf\{v\}\)\)\_\{g\}\-2\\cdot r\-l\(5\)
Since the human study revealed that the steering success of the explanations generated by our approach was nearly maximum\(4\.17/5\)\(4\.17/5\), we conclude that the quality of the classifier is sufficient for hyperparameter tuning\. Therefore, we emphasize its real\-world usefulness in general\.

### B\.4Factuality Metric

To measure the factuality of a generated explanations, we use theFActScoremetric[Min et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib33)which computes the probability of the set of atomic facts extracted from the explanation that are supported by a reliable knowledge source\.

To foster reproducibility, we utilize the open\-source implementation[Lage and Ostermann \(2025\)](https://arxiv.org/html/2608.29215#bib.bib24)\. As[Lage and Ostermann \(2025\)](https://arxiv.org/html/2608.29215#bib.bib24)suggest, we selectedOLMo\-2\-1124\-7B\-SFTfor the atomic fact generation, andgemma\-3\-4b\-itfor the atomic fact validation according to Wikipedia \(as of: April 1, 2023\)\. Since activation\-based steering leads the LLM to generate nonsense texts whenλ\\lambdais chosen too high[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23),OLMo\-2\-1124\-7B\-SFTfails to generate atomic facts for such explanations in our experiment\. We decided to set the factuality score to zero in such cases since the implementation of[Lage and Ostermann \(2025\)](https://arxiv.org/html/2608.29215#bib.bib24)does not handle this\.

### B\.5Harmonic Mean Metric

We additionally compute the harmonic mean to capture the balance of the evaluation criteria\. Since the harmonic mean is calculated based on the mean values, we cannot report significance results for the harmonic mean\.

For the automatic evaluation, we compute the harmonic mean based on the mean values for specificity \(s\) and factuality \(f\):

HMean≔2⋅s⋅fs\+f\\text\{HMean\}\\coloneqq\\frac\{2\\cdot s\\cdot f\}\{s\+f\}\(6\)
For the manual evaluation, we compute the harmonic mean based on the mean values for steering success \(s\), helpfulness \(h\) and plausibility \(p\):

HMean≔3⋅s⋅h⋅ps⋅h\+s⋅p\+h⋅p\\text\{HMean\}\\coloneqq\\frac\{3\\cdot s\\cdot h\\cdot p\}\{s\\cdot h\+s\\cdot p\+h\\cdot p\}\(7\)

### B\.6Results for ELI5 Questions

Table[7](https://arxiv.org/html/2608.29215#A2.T7)shows the mean results of our main experiments for the 600 ELI5 questions where we evaluate our approach based on them=5m=5most important attributes against two baselines using different model families\. We did not compute thef​a​c​t​u​a​l​i​t​yfactualitysince theFActScoremetric[Min et al\. \(2023\)](https://arxiv.org/html/2608.29215#bib.bib33)requires specifying a topic that maps to a Wikipedia title but not all ELI5 questions can be clearly mapped to a Wikipedia title\.

Table 7:Main automatic evaluation results:*Specificity*of all tested*LLMs*,*prompt components*, and*steering vectors*on the ELI5 test questions, along with steering*factor*λ\\lambdaand steered*layers*II\. Bold components/vectors are part of our approach\. Best values per LLM bold\.
### B\.7Ablation Results

Table 8:Automatic ablation results, comparing*Qwen 7B*and*32B*with 3 and 10 attributes respectively on the science questions\. Best values per LLM bold\. See caption of Table[1](https://arxiv.org/html/2608.29215#S6.T1)for further explanations\.Table 9:Automatic ablation results, comparing*Qwen 7B*and*32B*with 3 and 10 attributes respectively on the ELI5 test questions\. Best values per LLM bold\. See caption of Table[7](https://arxiv.org/html/2608.29215#A2.T7)for further explanations\.As described in Appendix[A\.5](https://arxiv.org/html/2608.29215#A1.SS5), we identify the topmmmost important style attributes and themmmost important knowledge attributes per group\. To assess the impact ofmmon the steering effectiveness, we considerm∈\{3,5,10\}m\\in\\\{3,5,10\\\}for the group\-specific explanation generation\. Table[8](https://arxiv.org/html/2608.29215#A2.T8)shows the results of our approach based on the Qwen model when using them=3m=3andm=10m=10most important attributes for the 97 science questions\. To further examine how model capacity influences steering effectiveness, we additionally evaluate a larger variant from the Qwen family \(Qwen2\.5\-32B\-Instruct\)\. Table[9](https://arxiv.org/html/2608.29215#A2.T9)shows the mean results for the ELI5 questions\.

Overall, the results denote that there are only marginal performance differences betweenm=3m=3andm=10m=10across all dimensions indicating we found a sweet spot withm=5m=5\. The results of the science question shown in Table[8](https://arxiv.org/html/2608.29215#A2.T8)indicate that increasing the model size increases the factuality but decreases the specificity of the explanations\. Based on the results shown in Table[1](https://arxiv.org/html/2608.29215#S6.T1)and[8](https://arxiv.org/html/2608.29215#A2.T8), using Qwen 32B with the second system prompt setting achieves the best balance between specificity and factuality \(HMean=0\.342\\textrm\{HMean\}=0\.342\)\. For the ELI5 question shown in Table[9](https://arxiv.org/html/2608.29215#A2.T9), increasing the model size resulted in increased specificity\. These findings suggest that using larger models can help achieving a better balance between specificity and factuality\.

### B\.8Results per Group

Figure[5](https://arxiv.org/html/2608.29215#A2.F5)and[6](https://arxiv.org/html/2608.29215#A2.F6)illustrate the steering effectiveness in terms of specificity per group of the prompting baseline, the standard activation\-based steering baseline[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)and our approach\. In Figure[5](https://arxiv.org/html/2608.29215#A2.F5), only the explanations for the 97 science questions are evaluated\. Figure[6](https://arxiv.org/html/2608.29215#A2.F6)shows the results for the 600 ELI5 test questions\. As a result, the approach of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)has the highest specificity for nearly all target groups across both question types\. The prompting baseline the lowest and our approach is in the middle\. The highest specificity was achieved by our approach for philosophers\. The approach of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)additionally achieves high specificity values for software engineers and game developers, for whom the performance of the prompting baseline collapses to nearly zero for the science questions\.

Figure 5:Per\-group steering effectiveness based on the means​p​e​c​i​f​i​c​i​t​yspecificityscores for the 97 science questions\. The blue area shows the prompting and the orange area shows the activation\-based steering baseline[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), each using only the group in the prompt\. The green area shows our approach, which additionally uses the five most important attributes in the prompt\.Figure 6:Per\-group steering effectiveness based on the means​p​e​c​i​f​i​c​i​t​yspecificityscores for the 600 ELI5 questions\. The blue area shows the prompting and the orange area shows the activation\-based steering baseline[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), each using only the group in the prompt\. The green area shows our approach, which additionally uses the five most important attributes in the prompt\.
### B\.9Human Study

To manually assess the steering effectiveness of the generated explanations of the Qwen \(7B\) model, we conducted a user study on the online freelancing platform Upwork666Upwork,[https://www\.upwork\.com](https://www.upwork.com/)\. We hire nine participants consisting of three experts representing the target groups: biologists, game developers, and philosophers\. We selected these target groups since they are the best performing groups across all approaches of the three clusters \(see Figure[3](https://arxiv.org/html/2608.29215#S6.F3)and[4](https://arxiv.org/html/2608.29215#S6.F4)\)\. While a study with nine experts*might*be limited, we would like to point out that having three experts for each of three groups is more than is common in many NLP generation papers, in our experience\. We define an expert as someone with formal education \(e\.g\., a degree in the respective field\) or professional experience in the field\. In addition, we filter for people with a job success rate of at least 90% to ensure high quality results\.

We sample 25 questions from the 97 science questions and 25 questions from the 600 ELI5 test questions\. Thus, the task of the study consists of 50 randomly sampled questions with three explanations for each question\. The first explanation was generated by the prompting baseline that only includes the target group in the prompt\. The second explanation was generated by the approach of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23), only based on the prompt including the target group\. The third explanation was generated by our approach based on the five most important style and knowledge attributes\. We randomly shuffle the order of the three explanations for each question to avoid a learning effect\. The participants were instructed to assume that a person of their target group \(e\.g\., a game developer\) asked the question and is honestly interested in the explanations\. This was important to avoid a misunderstanding when the question is in the domain of the target group, e\.g\., “How does DNA work?” for biologists, since normally an expert in biology would already know the answer\. The task was to rate the explanations based on the following three criteria:

1. 1\.How well is each explanation tailored to <target group\>?
2. 2\.How helpful is each explanation for <target group\> to understand the topic?
3. 3\.Does each explanation seem plausible to you?

We assess the three criteria on a 5\-point Likert scale from*not at all*\(1\) to*fully*\(5\) using Google forms\. We estimate that it takes about seven minutes to read the explanations and to answer our questions, which results in about six hours of work in total\. We paid a fixed\-price of$100\\$100which is a hourly wage of about$16\\$16which we assume to be an adequate pay anywhere on earth\.

In the following, the job description for biologists posted on Upwork is shown:

> In this study, our goal is to investigate how to best explain a topic to people with different backgrounds and abilities\. The task consists of 50 questions with three explanations for each question\. You will be asked to judge how well the explanations are tailored to biologists\. You will also be asked to rate whether the explanations are helpful for understanding the topic and whether the content is plausible\. We estimate that it will take about 7 minutes to read the three explanations and to answer our questions, which results in about 6 hours of work in total\. The deadline for submission is 31\.12\.2025, 11:59 PM \(CET\)\. Requirements: \- You must have a formal education \(e\.g\., a degree in biology\) or professional experience in the field of biology\. \- You must be fluent in English, with the ability to understand scientific explanations\.

In the following, the instruction for the participants \(in this case biologists\) are shown:

> Hi xxx, Thank you for your quick reply\! :\) As mentioned in the job description, the goal is to compare different explanations and evaluate how well they are tailored to biologists\. A fully tailored explanation considers the background knowledge and experience of biologists\. You will be given 50 questions \(e\.g\., Why is the sky blue?\) and for each of the questions three explanations \(A/B/C\)\. Assume that a biologist asked these questions and is honestly interested in the explanations\. Since you have experience in biology, your task is to rate the explanations based on the following three criteria: \(1\) How well is each explanation tailored to biologists? \(2\) How helpful is each explanation for biologists to understand the topic? \(3\) Does each explanation seem plausible to you? Rate the explanations on a 5\-point scale, from "not at all" to "fully"\. IMPORTANT: Evaluate each criterion independently\. For example, an explanation may be fully tailored by making use of good analogies, yet the topic may still be difficult to understand\. In that case, rate the first question as "fully" and the second question as "not at all," not low for both\. You will get a link to a test Google form, which includes only one question\. This is to ensure that you can open, navigate, and submit the Google form without any issues\. Once you have successfully submitted the test form, I will send you a link to the final Google form with all 50 questions\. Note: If you want to save your progress and reload it at a later point, you will need to log in with a Google account\. We would recommend conducting the study on a laptop or tablet, not on a smartphone\. If you have any questions, please don’t hesitate to reach out\! Otherwise, let me know if you are ready to start the study\. :\) Test Link: <Google forms link\> Participant ID: <id\> Best, xxx

Table[10](https://arxiv.org/html/2608.29215#A2.T10)shows the mean results in terms of steering success \(first question\), helpfulness \(second questions\) and plausibility \(third question\) assessed for the generated explanations of the ELI5 test questions\. Significance difference it determined with the Wilcoxon signed\-rank test\. In addition, the inter\-annotator agreement \(IAA\) based on the Krippendorff’s alpha is shown\. Table[11](https://arxiv.org/html/2608.29215#A2.T11)lists the mean results for the science questions\. Our approach achieves the highest steering success across both question types by a significant margin\. While the approach of[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)has the lowest values for helpfulness and plausibility, the prompting baseline performs best for these two dimensions\. However, there is no significant difference in the helpfulness values \(4\.014\.01vs\.3\.903\.90\) of the prompting baseline and our approach for the science questions as shown in Table[11](https://arxiv.org/html/2608.29215#A2.T11)in contrast to the ELI5 questions \(see Table[10](https://arxiv.org/html/2608.29215#A2.T10)\)\. We speculate that the complexity of the science questions enables more benefit of group\-specific explanations as generated by our approach, while the rather general topics of the ELI5 questions could already be well\-understood with generic explanations as generated by the prompting baseline\.

While the IAA is low to moderate for all evaluation criteria, this is expected for subjective tasks, since the goal is not to achieve a high IAA, as it is in annotation\. Rather, the IAA serves as a means to quantify the amount of subjectivity in the average of various opinions\. Our manual inspection of the evaluators’ results shows that they all worked reliably\.

Table 10:Manual evaluation results based on 25 randomly sampled ELI5 test questions: We evaluatesteering successin terms of how well the explanations are tailored to the target group, theplausibilityof the explanations and, howhelpfulthe explanations are to understand the topic generated by Qwen 7B for the two baselines and our approach\. Significant difference to the prompting baseline are marked with†\\daggerand significant difference to[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)are marked with‡\\ddagger\(Wilcoxon signed\-rank test,p<\.05p<\.05\)\. Inter\-annotator agreement \(IAA\) was calculated using Krippendorff’s alpha\.Table 11:Manual evaluation results based on 25 randomly sampled science questions: We evaluatesteering successin terms of how well the explanations are tailored to the target group, theplausibilityof the explanations and, howhelpfulthe explanations are to understand the topic generated by Qwen 7B for the two baselines and our approach\. Significant difference to the prompting baseline are marked with†\\daggerand significant difference to[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)are marked with‡\\ddagger\(Wilcoxon signed\-rank test,p<\.05p<\.05\)\. Inter\-annotator agreement \(IAA\) was calculated using Krippendorff’s alpha\.
### B\.10Explanation Samples

Table[12](https://arxiv.org/html/2608.29215#A2.T12)\-[17](https://arxiv.org/html/2608.29215#A2.T17)show the generated explanations for two questions generated by Qwen 7B for the two baselines and our approach\. Upon manual analysis, we find that the explanations of the baseline based on[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)are somewhat tailored to the target group but do not answer the question anymore \(e\.g\., code\-based explanation in Table[13](https://arxiv.org/html/2608.29215#A2.T13)\)\. This might explain the low helpfulness scores\. In contrast, the explanations of the prompting baseline perfectly answer the questions but are barley tailored the target group\. As shown in the human study, our approach demonstrates the best steering success by successfully tailoring the explanations to the target group \(Table[14](https://arxiv.org/html/2608.29215#A2.T14)and[17](https://arxiv.org/html/2608.29215#A2.T17)\)\. Given that higher steering success does not appear to translate into greater helpfulness, future work should focus on determining when successful tailoring is helpful and when it is not, in order to ensure that explanations are both tailored and helpful\. For example, rather simple topics could already be well\-understood with generic explanations as demonstrated by the prompting baseline\.

Table 12:Explanation samples of the*prompting baseline*generated by Qwen 7B for the question “How does DNA work?” tailored to three different target groupsg∈\{b​i​o​l​o​g​i​s​t​s,g​a​m​e​d​e​v​e​l​o​p​e​r​s,p​h​i​l​o​s​o​p​h​e​r​s\}g\\in\\\{biologists,gamedevelopers,philosophers\\\}\. The prompting baseline uses the target groupggin the prompt\.Table 13:Explanation samples of the*activation\-based steering baseline*[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)generated by Qwen 7B for the question “How does DNA work?” tailored to three different target groupsg∈\{b​i​o​l​o​g​i​s​t​s,g​a​m​e​d​e​v​e​l​o​p​e​r​s,p​h​i​l​o​s​o​p​h​e​r​s\}g\\in~\\\{biologists,gamedevelopers,philosophers\\\}\. The baseline uses the target groupggin the prompt and the activation\-based steering vectors based on the average of the activation vectors extracted for the 500 Stack Exchange answers ofgg\.Table 14:Explanation samples of*our approach*generated by Qwen 7B for the question “How does DNA work?” tailored to three different target groupsg∈\{b​i​o​l​o​g​i​s​t​s,g​a​m​e​d​e​v​e​l​o​p​e​r​s,p​h​i​l​o​s​o​p​h​e​r​s\}g\\in\\\{biologists,gamedevelopers,philosophers\\\}\. Our approach uses the five most important attributes identified forggin the prompt and the attribute\-based steering vector according to Equation[1](https://arxiv.org/html/2608.29215#S3.E1)\.Table 15:Explanation samples of the*prompting baseline*generated by Qwen 7B for the question “What is the role of enzymes in digestion?” tailored to three different target groupsg∈\{b​i​o​l​o​g​i​s​t​s,g​a​m​e​d​e​v​e​l​o​p​e​r​s,p​h​i​l​o​s​o​p​h​e​r​s\}g\\in\\\{biologists,gamedevelopers,philosophers\\\}\. The prompting baseline uses the target groupggin the prompt\.Table 16:Explanation samples of the*activation\-based steering baseline*[Konen et al\. \(2024\)](https://arxiv.org/html/2608.29215#bib.bib23)generated by Qwen 7B for the question “What is the role of enzymes in digestion?” tailored to three different target groupsg∈\{b​i​o​l​o​g​i​s​t​s,g​a​m​e​d​e​v​e​l​o​p​e​r​s,p​h​i​l​o​s​o​p​h​e​r​s\}g\\in~\\\{biologists,gamedevelopers,philosophers\\\}\. The baseline uses the target groupggin the prompt and the activation\-based steering vectors based on the average of the activation vectors extracted for the 500 Stack Exchange answers ofgg\.Table 17:Explanation samples of*our approach*generated by Qwen 7B for the question “What is the role of enzymes in digestion?” tailored to three different target groupsg∈\{b​i​o​l​o​g​i​s​t​s,g​a​m​e​d​e​v​e​l​o​p​e​r​s,p​h​i​l​o​s​o​p​h​e​r​s\}g\\in\\\{biologists,gamedevelopers,philosophers\\\}\. Our approach uses the five most important attributes identified forggin the prompt and the attribute\-based steering vector according to Equation[1](https://arxiv.org/html/2608.29215#S3.E1)\.