PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Leandra Fichtel, Janek Prange, Henning Wachsmuth

arXiv:2608.29215v1cs.CL

TL;DR

The paper asks whether LLMs can generate explanations tailored to specific groups when prompting alone is insufficient and group-specific supervised data is unavailable. It extracts group style and knowledge attributes, converts them into activation-based steering vectors, and applies them during inference. Across automatic and expert evaluations, the approach achieves better tailoring while largely maintaining factuality, though group profiles can oversimplify heterogeneous individuals.

  • Problem

    Prompting is insufficient for fine-grained group-specific explanations, while existing explanation data generally lacks explainee-group information needed for supervised training.

  • Method

    The method extracts target-group style and knowledge attributes from group-member explanations, then builds attribute-based steering vectors added to selected LLM activations during inference.

  • Results

    The approach generates significantly better-tailored explanations than prompting and state-of-the-art steering baselines while largely maintaining factuality.

  • Takeaways & Limitations

    The results provide evidence that explanations can be computationally tailored to groups by learning from explanations written by members of those groups.

  • Takeaways & Limitations

    Group profiles can oversimplify heterogeneous individuals and overlook variations in abilities or background within a target group.

Abstract

from arXiv · show

To effectively enable people to understand new topics, explanations should be tailored to their backgrounds and abilities. So far, prompting alone has been shown to be insufficient for creating such explanations and other computational methods are missing. 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 of our approach 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 largely maintaining factuality.

1 Introduction

The paper addresses the lack of computational methods for generating explanations tailored to target groups, beyond prompting alone. It proposes attribute-based activation steering and finds that the approach improves group tailoring while largely maintaining factuality.

  • Motivation: Group-specific explanation generation lacks computational methods that explicitly tailor explanations to target-group preferences without supervised explanation data.Prompting is sensitive to formulation and has low steering effectiveness, while existing explanation datasets generally lack explainee group information.
  • Contributions: The approach combines prompting and clustering to extract explanatory-style and knowledge attributes from texts written by target-group members.Examples include using code snippets as a style attribute and understanding game development as a knowledge attribute.
  • Contributions: Attribute-based steering vectors are constructed from style and knowledge attribute activations and added to selected LLM layers during inference.This design enables fine-grained steering toward a target group rather than relying only on broad prompting.
  • Contributions: The paper proposes a training-free method for attribute-based activation steering of LLMs to generate group-specific explanations.The method obtains style and knowledge attributes for a target group and uses them to steer internal activations during inference.
  • Results: Expert evaluation provides empirical evidence that the proposed steering is effective for group-specific explanation generation.The introduction reports evaluation against prompting and steering baselines using specificity and factuality, alongside a human study.

2 Related Work

Prior work studies personalized, group-specific, persona-, and activation-based steering, but existing approaches have limitations in fine-grained control, scalability, or attribute coverage. The paper positions its method as combining group modeling with attribute-based activation steering.

  • Personalization and group adaptation: Personalized explanations should reflect both explanatory style and the explainee’s background, but individual preferences are often unavailable.Group adaptation is presented as an alternative when information about the specific explainee cannot be accessed.
  • Personalization and group adaptation: Existing human-written explanation datasets generally lack personal information about explainees, limiting their direct use for group-specific generation.Related work therefore investigates identifying preferences of specific groups from other sources or methods.
  • Existing computational methods: Manual rhetorical features can limit scalability, while attribute-controlled fine-tuning and style-neuron methods are mainly evaluated on attributes with clear lexical indicators.These limitations motivate methods that can represent less overt group-specific style and knowledge attributes.
  • Prompting and personalization: Prompting does not reliably support fine-grained adaptation across groups such as different age or education levels.Prior findings report that LLMs do not adapt to different age levels when prompted accordingly, and prompting is unsuitable for fine-grained steering in this setting.
  • Activation steering: Recent activation-steering methods support inference-time output control, including multi-attribute steering, but this paper instead constructs vectors from natural-language style and knowledge attributes.The distinction is the paper’s route to fine-grained group-specific control.

3 Approach

The approach identifies target-group style and knowledge attributes, derives attribute-specific activation vectors, and combines them into weighted steering vectors applied during inference. This three-step pipeline provides fine-grained control over group-specific explanation generation.

  • Approach overview: The approach has three steps: extract target-group attributes, create attribute-specific activation vectors, and compute steering vectors for inference-time generation.The attributes model group abilities and background in terms of explanatory style and knowledge.
  • Attribute extraction: A two-stage prompting process extracts style descriptions and knowledge attributes from texts written by members of each target group.Knowledge extraction covers factual, conceptual, procedural, metacognitive, situational, and conventional knowledge.
  • Attribute selection: LLM-based filtering retains attributes classified as style or knowledge, after which semantic clustering creates shared attribute sets across target groups.Clustering reduces noise and supports comparison using the same style and knowledge attributes for all groups.
  • Attribute selection: The final attribute inventory retains all surviving clusters rather than enforcing a fixed number of attributes.Attributes are selected by frequency while rejecting those too similar to previously selected attributes.
  • Group-specific profiles: For each group, binary text-level attribute vectors are averaged into a group profile that identifies important style and knowledge attributes.An attribute value of 1 indicates that the attribute was predicted for the text during the two-stage prompting process.
  • Attribute-specific activations: Attribute-specific activation vectors are obtained by contrasting forward-pass activations for an attribute with activations for its negation.Style attributes are phrased as second-person instructions, whereas knowledge attributes use first-person statements.
  • Attribute-based steering: The group steering vector sums attribute-specific activation vectors and weights them using the softmax of the group’s attribute profile.The resulting vector is added to the original activation at layer i, with λ controlling steering strength.

4 Data

The experiments use Stack Exchange answers to extract group-specific attributes and general explanatory questions to evaluate steering. The data cover eleven target groups, with ELI5 and science-question sets serving as evaluation inputs.

  • Data sources: The data section uses Stack Exchange texts to extract attributes and separate explanatory questions to generate group-specific explanations.The source data and evaluation questions serve different roles in the experimental pipeline.
  • Group-specific attribute extraction: Stack Exchange is treated as a collection of domain-specific forums whose answers approximate texts written by members of the respective target groups.The authors acknowledge this group-membership assumption will not always hold and randomly select 6,000 answers from each of eleven groups.
  • Explanation-generation inputs: Evaluation questions are selected to be general enough to plausibly be asked by diverse target groups.The study employs ELI5 questions and scientific questions as two input types.
  • Explanation-generation inputs: The ELI5 evaluation uses 1,507 validation questions and 600 test questions from the KILT benchmark.Because ELI5 answers target a general audience, they are used as questions rather than ground-truth explanations.
  • Explanation-generation inputs: The science-question set contributes 97 additional questions after duplicate filtering.These questions originate from a prior study of adaptation to age and education levels.

5 Experiments

The experiments evaluate attribute extraction and group-specific explanation steering across multiple models, using automatic metrics and a human study. The setup compares prompting and activation-based baselines while measuring specificity, factuality, and their harmonic mean.

  • Two main experiments assess the adequacy of extracted attributes and the effectiveness of group-specific steering.
  • The pipeline identifies important attributes for 11 target groups from Stack Exchange answers using prompting, filtering, and clustering.The experimental setup retains 1,250 attributes: 340 style and 910 knowledge attributes.
  • The steering evaluation compares the proposed approach with prompting and standard activation-based steering across Qwen2.5-7B, Llama-3.1-8B, Ministral-8B, and Qwen2.5-32B models.The final evaluation uses 600 ELI5 test questions and 97 science questions after hyperparameter tuning on 100 validation questions.
  • The proposed attribute-based vectors are built from the most important group attributes, whereas the baseline averages target-group activations against those of other groups.Steering vectors are added to selected internal layers during inference, with layer windows and steering strength tuned on validation data.
  • Group specificity is estimated by a classifier that predicts scores for 11 groups from attribute vectors, with the target-group softmax score interpreted as tailoring probability.
  • Factuality is measured with FACTSCORE, while the harmonic mean captures the balance between specificity and factual accuracy.FACTSCORE estimates support for atomic facts using a knowledge source; factuality results are reported only where questions map to Wikipedia titles.

6 Results and Discussion

The extracted group profiles show broadly shared explanatory styles but more differentiated domain knowledge, and the proposed steering method generally offers the best specificity–factuality balance. Automatic and expert evaluations support improved group tailoring, while also revealing a tradeoff with factuality, plausibility, and helpfulness.

  • Attribute profiles: Style profiles are generally homogeneous across groups, whereas knowledge profiles are more diverse and keep non-technical groups distinct.These correlations suggest the profiles capture both shared explanatory structure and meaningful domain-specific differences.
  • Automatic evaluation: Including extracted attributes in prompts improves group specificity across approaches and models, but generally lowers factuality.The results frame attribute conditioning as a specificity–factuality tradeoff rather than a uniformly positive change.
  • Automatic evaluation: 0.313 and 0.299 are Qwen’s highest harmonic means, while 0.291 is its second-best value across the reported configurations.For Ministral, extracted attributes with Konen et al. (2024) achieve the best harmonic mean of 0.274; prompting has the best factuality for all models at 0.322–0.376.
  • Human evaluation: 4.17 steering success is significantly higher for the approach, while prompting achieves the best plausibility at 4.56.The approach also achieves the best overall harmonic mean of 3.88, whereas Konen et al. (2024) records helpfulness 2.70 and plausibility 2.75.
  • Human evaluation: Higher steering success does not translate into significantly greater helpfulness, partly alongside low inter-annotator agreement of Krippendorff’s α = 0.178–0.489.The authors also report no significant helpfulness difference for the science questions and identify question type as a factor steering methods should model explicitly.

7 Conclusion

The paper presents a three-step method for group-specific explanation generation and evaluates it across 11 target groups using automatic metrics and a human study. The approach achieves the best balance between tailoring explanations and maintaining factuality, although steering success does not necessarily predict helpfulness.

  • The approach identifies group-specific style and knowledge attributes, extracts attribute-specific activation vectors, and combines them into steering vectors for inference-time LLM steering.
  • Evaluation covered 11 target groups with automatic metrics and a human study involving nine experts from three domains.
  • The method achieves the best balance between generating well-tailored explanations and maintaining factuality.
  • Steering success does not necessarily correlate with explanation helpfulness, motivating further study of what tailoring is most helpful.

8 Limitations

The approach has scope, measurement, and modeling limitations. Its evaluation is constrained to English and fixed groups, factuality relies on relative LLM-based comparisons, and increased specificity can reduce factual precision.

  • The method only accounts in a limited way for individuals whose abilities or backgrounds notably deviate from the mean group profile.
  • Experiments are limited to English and a fixed set of target groups, so generalization to other languages and groups remains unstudied.
  • Group profiles assume Stack Exchange contributors represent target groups, but users may cross domains, differ in expertise, or write atypical answers.
  • LLM-based attribute extraction inherits biases from the underlying model, although the extracted profiles appeared adequate for the 11 evaluated groups.
  • FACTSCORE uses LLMs for fact validation, so the authors interpret factuality through relative differences rather than absolute values.
  • Greater specificity can introduce oversimplifications that reduce factual precision, creating a tradeoff between group specificity and factuality.

9 Ethical Considerations

The paper identifies ethical risks in group-specific explanation generation, especially stereotyping, source-data bias, and misuse. Its data and extraction process also shape the profiles used for steering.

  • Group profiles may oversimplify knowledge and abilities, mix heterogeneous subgroups, or overlook important within-group variation.
  • Explanations derived from Stack Exchange attributes may reflect biases present in the source data.
  • The approach could be misused for targeted persuasion or misinformation beyond harmless educational settings, despite those not being its intended use.
  • The human study has limited sample size and potentially subjective ratings of steering success, helpfulness, and plausibility.
  • Attributes are extracted from English Stack Exchange explanations using prompts that cover explanatory style and multiple knowledge types.

A.4.4 Ablation Study

The ablation study tests whether attribute extraction depends on the model used and examines implementation choices affecting steering. Steering directions remain highly aligned across extraction models, while larger models improve factuality but reduce specificity.

  • Attribute extraction uses Qwen 7B, with Llama 8B and Ministral 8B tested in an ablation study.
  • Steering-vector comparisons provide a robustness check without rerunning the full pipeline and its layer and scaling-factor tuning.
  • Cosine similarities between Llama- or Ministral-derived and Qwen-derived steering vectors exceed 0.8 and reach 0.96.
  • The results indicate that steering directions are highly similar across different attribute-extraction LLMs and that the method does not critically depend on Qwen.

B.3 Specificity Metric

The paper measures group specificity from attribute-based representations of generated explanations and uses this metric to tune steering. It also evaluates factuality and combines specificity with factuality through a harmonic mean.

  • Specificity measurement: A classifier predicts each target group from an explanation’s binary attribute vector to measure group specificity.The classifier is trained on Stack Exchange-derived attribute vectors and achieved 0.876 test accuracy.
  • Specificity measurement: The specificity classifier is used on generated explanations and also validates the steering hyperparameters on a held-out validation set.The tuning procedure additionally applies repetition, length, and character-validity checks.
  • Human-evaluation linkage: The human study reports near-maximum steering success for the approach, with a mean score of 4.17/5.The authors use this result to argue that the classifier is sufficient for hyperparameter tuning.
  • Factuality measurement: FACTSCORE measures factuality as the probability that atomic facts extracted from an explanation are supported by a reliable knowledge source.The implementation uses OLMo-2-1124-7B-SFT for fact generation and gemma-3-4b-it for Wikipedia-based validation.
  • Harmonic means: The automatic harmonic mean combines mean specificity and factuality, while the manual harmonic mean combines steering success, helpfulness, and plausibility.Because the automatic harmonic mean is computed from mean values, significance results are not reported for it.

B.7 Ablation Results

The ablations identify five attributes as a practical setting because using three or ten produces only marginal differences. Model size changes specificity and factuality differently across science and ELI5 questions, while manual evaluation favors the proposed approach for steering success.

  • Attribute-count ablation: Using m = 3 and m = 10 attributes produces only marginal performance differences, supporting m = 5 as a practical setting.The comparison covers all evaluated dimensions in the ablation results.
  • Model-size ablation: For science questions, larger models increase factuality but decrease specificity, whereas for ELI5 questions they increase specificity.The best reported science-question balance is Qwen 32B with the second system prompt setting, achieving HMean = 0.342.
  • Automatic steering comparison: Across both question types, the Konen et al. (2024) baseline has the highest specificity for nearly all groups, while prompting is lowest and the proposed approach is intermediate.The proposed approach achieves the highest specificity for philosophers.
  • Human evaluation: The study used nine experts from biology, philosophy, and game development to rate 50 sampled questions on tailoring, helpfulness, and plausibility.The sample included 25 science questions and 25 ELI5 questions, with three explanations per question.
  • Human evaluation: In the expert study, the proposed approach achieves the highest steering success across both question types by a significant margin.Prompting performs best on helpfulness and plausibility in the reported comparison, while its helpfulness advantage over the proposed approach is not significant for science questions.
  • Human evaluation: Low-to-moderate inter-annotator agreement reflects subjectivity in the evaluation criteria, although manual inspection indicated reliable evaluator behavior.The authors frame agreement as a measure of subjectivity rather than an annotation-quality target.

B.10 Explanation Samples

The explanation samples illustrate a trade-off among prompting, standard activation steering, and attribute-based steering. The proposed approach best balances answering the question with tailoring the explanation to the target group in the reported examples.

  • Baseline examples: Prompting answers the questions well but produces explanations that are barely tailored to the target group.The prompting baseline uses only the target group in the prompt.
  • Baseline examples: The Konen et al. (2024) activation-steering baseline can tailor explanations but sometimes no longer answers the question.The paper gives a code-based explanation as an example and links this behavior to low helpfulness.
  • Interpretation: The human-study results show that the proposed approach achieves the best steering success among the three explanation-generation methods.The paper notes that stronger tailoring does not necessarily translate into greater helpfulness, especially for simple topics.
  • Automatic evaluation layout: Table 7 compares specificity across tested language models, prompt components, steering vectors, steering factors, and steered layers on ELI5 questions.Bold entries identify components of the proposed approach and the best value for each language model.
  • Proposed approach: The proposed approach uses the five most important style and knowledge attributes with an attribute-based steering vector for each target group.The examples cover “How does DNA work?” and “What is the role of enzymes in digestion?” for biologists, game developers, and philosophers.
Loading 2608.29215v1…