PaperlayerView exact revision on arXiv

Source-linked AI summary

Can Activation Steering Capture Multidimensional Authorship Style?

Hieu Tran, Calvin Bao, Marine Carpuat

arXiv:2609.04792v1cs.CLcs.AI

TL;DR

Authorship style is multidimensional and difficult to represent without copying target content, motivating richer exemplar representations than textual descriptors or single steering directions. A3S extracts rhetorical-aspect contrastive directions in activation space, merges them to manage interference, and tunes steering per instance. It improves transfer most for genuinely multi-aspect styles and outperforms a trained baseline in preference evaluations, particularly out of domain.

  • Problem

    Authorship style combines multiple rhetorical dimensions, but existing descriptors and single directions inadequately represent arbitrary exemplar-based style without copying content.

  • Method

    A3S uses structured rhetorical contrastive prompting to construct per-aspect activation directions, merges them with interference-aware aggregation, and tunes steering strength per instance.

  • Results

    A3S captures fine-grained style more effectively than natural-language descriptors, generalizes better out of domain than a learned model, and shows strongest gains for genuinely multi-aspect authorship.

  • Takeaways & Limitations

    Aspect-specific residuals carry genuine stylistic signal, so naive direction aggregation can fail and requires interference-aware merging.

  • Takeaways & Limitations

    The evaluation is primarily English, and A3S incurs nontrivial inference-time overhead from direction construction and per-instance strength search.

Abstract

from arXiv · show

Activation steering has shown promise for controlling LLM generation along well-defined attributes, but it remains unclear whether it can handle the multidimensional and hard-to-define nature of authorship style. We ask whether structured contrastive prompting along rhetorically-motivated dimensions can construct rich style representations directly in activation space, bypassing the need for natural language style descriptors or dedicated training. We find that the resulting directions share a common authorship backbone while conflicting on aspect-specific residuals that carry genuine stylistic signal, explaining why naive aggregation fails. We operationalize this in Aspect-Aware Activation Steering (A3S), a training-free framework that merges per-aspect contrastive directions with interference-aware aggregation and tunes steering strength per instance. A3S improves authorship style transfer where it is genuinely multi-aspect, outperforms a trained baseline in preference evaluations on out-of-domain benchmarks, and keeps target-exemplar overlap consistently low.

1 Introduction

Authorship style is inherently multidimensional, making single-attribute or single-direction transfer inadequate. A3S addresses this by extracting rhetorical-aspect directions directly in activation space, merging them interference-awarely, and tuning steering per instance.

  • Authorship style combines lexical, rhythmic, organizational, and perspectival choices that must be transferred without changing the source message.
  • Single-direction activation steering can suffer interference because different stylistic aspects may push in opposing directions.
  • Structured contrastive prompting rewrites target exemplars along rhetorical dimensions to extract richer activation-space representations without natural-language descriptors or dedicated training.
  • A3S constructs per-aspect directions, merges them with activation-space Parameter Competition Balancing, and tunes steering strength per instance using Hybrid Adaptive Search.
  • Across MUD, LaMP, and LongLaMP, A3S gains most where authorship is genuinely multi-aspect and outperforms TinyStyler in preference evaluations, especially out of domain.
  • Per-aspect directions share an authorship backbone while retaining aspect-specific residuals, providing a geometric basis for interference-aware aggregation.
  • A3S captures fine-grained style more effectively than natural-language descriptors and generalizes better out of domain than a learned style-transfer model.

2 Related Work

Prior style-transfer methods rely on textual descriptors, trained embeddings, or single-attribute steering, which inadequately represent arbitrary exemplar-based authorship. A3S instead recomputes rhetorical-aspect activation directions per exemplar and aggregates them to address interference.

  • Text-style transfer has expanded from predefined attributes such as sentiment and formality toward arbitrary styles defined implicitly by reference text.
  • Prompt-based methods are training-free but can offer limited control, while direct exemplar prompting may inflate style match by copying target wording.
  • Activation steering avoids weight updates but prior methods generally treat style as one attribute with fixed strength, poorly matching user-specific authorship.
  • Authorship embeddings support attribution and generation conditioning, whereas A3S evaluates them but steers with aspect-decomposed contrastive activations.
  • Multi-attribute steering methods address interference for predefined labels or coarse categories, unlike A3S’s arbitrary-exemplar setting.
  • A3S decomposes target exemplars into rhetorical aspects, resolves interference through PCB-Merging, and dynamically tunes injection strength.

3 Methodology

A3S represents authorship style as multiple activation-space directions extracted from rhetorically motivated contrastive rewrites, then merges and tunes them at inference time. Its aggregation addresses conflicts between shared authorship structure and aspect-specific stylistic signals while constraining degeneration.

  • Problem formulation: A3S frames authorship-style transfer as representation learning from an exemplar, using activation-space directions instead of natural-language style descriptions.The target style is modeled as multiple aspect directions while preserving the source text’s meaning.
  • Aspect decomposition: The method decomposes authorship style into Tone, Diction, Perspective, Structure, Figurative Language, Syntax, and Surface.These dimensions are derived from literary criticism and composition theory.
  • Contrastive direction extraction: For each aspect, A3S generates meaning-preserving contrastive variants and computes activation directions from positive and negative representations.The variants isolate stylistic dimensions while keeping semantic content fixed; local aspects are more reliably isolated than discourse aspects.
  • Style aggregation: A3S merges aspect directions with interference-aware aggregation, comparing mean, median, TIES-Merging, and PCB-Merging strategies.PCB reweights components using relative magnitude and inter-aspect directional consistency, preserving shared and conflicting stylistic components.
  • Style aggregation: The aspect directions share an authorship backbone but contain conflicting residuals, so naive averaging can cancel genuine stylistic signal.This structure motivates PCB’s simultaneous use of magnitude and consistency weighting.
  • Adaptive steering: A3S searches for a sample-specific steering coefficient within a feasible region that maximizes steering intensity while penalizing distributional collapse.The search rejects suffix periodicity and truncation, and the objective penalizes unlikely continuations, distributional drift, and mode collapse.

4 Experiments

The experiments evaluate A3S across three authorship-style transfer benchmarks with prompting, activation-steering, and training-based baselines. They measure style matching, meaning preservation, target overlap, length shift, and pairwise preferences across in-domain and out-of-domain settings.

  • Datasets: The evaluation spans MUD, LaMP, and LongLaMP, covering short-form, Twitter, and long-form authorship transfer.MUD is Reddit authorship, LaMP is unsupervised Twitter-user transfer, and LongLaMP uses a topic-writing task.
  • Baselines: A3S is compared with prompting methods, activation-steering methods, and TinyStyler, a few-shot model conditioned on learned authorship embeddings.The prompting baselines include Simple Prompting, STYLL, RG, and Aspect-Aware Prompting; activation baselines include Konen et al. and Global Steering.
  • Baselines: Because TinyStyler is trained on part of MUD, MUD is treated as in-domain while LaMP and LongLaMP are treated as out-of-domain.This split frames the comparison between the training-based baseline and A3S across domain conditions.
  • Evaluation: Automatic evaluation measures style matching with LUAR, StyleCAV, and StyleDistance, and meaning preservation with MIS and SBERT.Style metrics report similarity toward the target and dissimilarity away from the source using cosine similarity.
  • Evaluation: Target overlap is measured with ROUGE-1, ROUGE-2, and ROUGE-L, while length shift tests whether gains reflect verbosity rather than stylistic transfer.The length analysis includes the output-to-source character-length ratio, the fraction of outputs over 10% longer, and mean absolute length shift.
  • Evaluation: Pairwise preferences are evaluated with GPT-4.1 as an LLM judge on all benchmarks and with a human study on MUD.The two best systems are compared using the preceding evaluation criteria.

5 Results and Analysis

Across three benchmarks, A3S improves style fidelity over prompting and activation-steering baselines while maintaining low target overlap, with preference gains strongest out of domain. Analyses attribute these gains to aspect-specific residuals and per-instance steering-strength adaptation, though automatic comparison with TinyStyler remains confounded.

  • 5.1 Main Results: A3S leads LUAR-based style fidelity across all three benchmarks for Qwen and keeps target-exemplar overlap lowest on MUD and LaMP.It also maintains near-source length ratios across both backbones.
  • 5.1 Main Results: A3S improves all three style-fidelity metrics over two non-decomposed activation-steering baselines on MUD and LaMP across both backbones.On LongLaMP, A3S (Qwen) also outperforms both baselines on all three style-fidelity metrics.
  • 5.1 Main Results: A3S is competitive with TinyStyler on LUAR and StyleDistance, but automatic comparisons are confounded by TinyStyler’s output-length differences, metric alignment, and mixed meaning preservation.TinyStyler expands short-form outputs 2.3x–4.4x and truncates long-form outputs to 0.3x.
  • 5.1 Main Results: In preference evaluations, A3S (Qwen) is preferred to TinyStyler on MUD by humans and GPT-4.1, with the strongest preference on out-of-domain LaMP and LongLaMP.Human annotators preferred A3S in 114 of 185 non-tie MUD judgments, or 61.6%.
  • 5.2 Are Rhetorically-Motivated Dimensions Distinct in Activation Space?: The results support interference-aware, multi-aspect steering: A3S consistently beats Global Steering, while residual information explains its advantage over single-direction aggregation.A3S’s strongest gains occur where target authorship is genuinely multidimensional.
  • 5.3 Impact of Per-Instance Steering Strength: Hybrid Adaptive Search reaches StyleCAV 0.48 while maintaining SBERT at 0.69, outperforming fixed α settings on the Pareto frontier.The search is roughly 25% faster than Binary Search, with construction costs amortized across downstream queries.

6 Conclusion

Structured contrastive prompting can construct rich authorship-style representations in activation space without dedicated training. A3S uses their shared backbone and aspect-specific residuals to motivate interference-aware merging, with the greatest gains expected for genuinely multidimensional styles.

  • Structured contrastive prompting along rhetorically motivated dimensions produces activation-space representations sufficient for authorship style transfer without dedicated training.
  • Aspect directions decompose into a shared authorship backbone and aspect-specific residuals carrying genuine stylistic signal.
  • This decomposition explains why naive aggregation fails and motivates interference-aware merging.
  • A3S is expected to yield its largest gains when authorship combines multiple rhetorical dimensions, narrowing to baseline parity for single-attribute styles.
  • The findings suggest activation space contains rich authorship-style structure accessible through structured prompting without parameter updates.

7 Limitations

A3S depends on the quality of generated contrastive variants and has untested crosslingual generalization, while its adaptive inference procedure adds substantial computational cost.

  • A3S aspect directions may reflect unintended stylistic changes when the generator cannot vary one aspect while holding others relatively fixed.The quality of the generated variants therefore constrains the quality of the resulting directions.
  • The shared PC1 backbone is an assumption whose robustness could be assessed with alternative generators.
  • Crosslingual transfer of the aspect geometry remains untested because the datasets are primarily English.
  • Direction construction requires K=16 generations per target profile, and per-instance α search adds several forward passes for each input.Using fewer aspects could reduce cost but would entail some quality loss.

AI Use Statement

The paper reports generative-AI assistance and describes implementation procedures for adaptive search, model backbones, direction extraction, and inference-time steering.

  • Generative AI tools assisted drafting and editing, while the authors reviewed the resulting text, equations, citations, and code-derived descriptions.
  • Adaptive search: Calibration uses binary search to find the largest feasible base steering strength α under degeneration and soft-penalty checks.
  • Adaptive search: Adaptation probes upward from a valid anchor or recovers downward from an invalid one using a fixed step.
  • Backbones and decoding: Experiments use frozen Llama-3.2-3B-Instruct and Qwen3-4B backbones with greedy decoding and steering during prefill and decoding.
  • Direction extraction: For each aspect, mean-pooled residual-stream differences between the exemplar and generated variants form a direction, which PCB then merges across layers.
  • PCB and Hybrid Search: PCB retains the top-r=0.5 coordinate importance scores, while calibration searches α in [0, 1] with tolerance ϵ=0.01.

A.3 Dataset Construction Details

The evaluation constructs authorship-transfer benchmarks from Reddit, Twitter, and long-form writing, then measures stylistic alignment, source divergence, meaning preservation, target overlap, and preference judgments. A3S shows its strongest advantage on multidimensional authorship and smaller gains for single-attribute formality transfer.

  • Reddit Authorship (MUD): MUD evaluates few-shot user imitation using 15 source authors, 15 target authors, and 3,600 source-target evaluation pairs.
  • Twitter Authorship (LaMP): LaMP uses distinct target-user profiles built from K=9 historical posts, yielding 1,496 Twitter evaluation samples.
  • Long-Form Authorship (LongLaMP): LongLaMP reformulates topic writing as transfer between distinct authors and evaluates a randomly sampled subset because long-form generation is computationally costly.
  • Evaluation metrics: Stylistic alignment is measured with embedding-based authorship or style models, while lower ROUGE scores indicate less direct reuse of target-exemplar content.
  • Evaluation metrics: Towards scores measure similarity to the target exemplar, Away scores measure divergence from the source, and MIS measures bidirectional entailment.
  • Formality control: In the formality control, A3S shows smaller improvements than on multi-aspect tasks; TinyStyler leads three style metrics, while A3S achieves the highest Qwen SBERT scores in both directions.A3S and TinyStyler are effectively tied in GPT-4.1 preferences, with 50% and 45%, respectively.
  • Representation analysis: Figure 5 measures how well six stylistic dimensions distinguish target-author text from foil text, with most dimensions peaking in middle network layers.This supports intervening at those layers during generation.
  • Additional analyses: The appendices provide complete ablations across stylistic dimensions, aggregation strategies, evaluation metrics, and adaptive-α settings.

A.10 Additional Length and Preference Results

Additional analyses show that A3S’s aspect directions contain substantial conflicting residual structure, while preference evaluations favor A3S over TinyStyler. These results support interference-aware aggregation and stronger out-of-domain performance.

  • Preference Results: 61.6% is A3S’s tie-excluded human win rate against TinyStyler on MUD, based on 210 judgments across 70 instances.A3S received 114 selections versus 71 for TinyStyler, with 25 ties; the 95% item-level CI was [53.3%, 69.7%].
  • Preference Results: GPT-4.1 prefers A3S over TinyStyler with both Llama and Qwen backbones, most strongly on the out-of-domain LaMP and LongLaMP benchmarks.A3S also exceeds TinyStyler on LUAR and SBERT across all three benchmarks.
  • Aspect Geometry: 60% of aspect pairs have negative residual cosine similarity after PC1 removal, with Perspective and Structure opposing local realization aspects.Raw aspect vectors remain positively aligned because of a shared authorship component.
  • Aspect Geometry: 0.627 is the mean residual-to-total norm ratio, indicating that roughly two-thirds of each aspect direction lies outside the shared PC1 backbone.Tone and Vocabulary have lower residual ratios of 0.36 and 0.37, respectively.
  • Aggregation: 0.378 at α=0.5 is achieved by raw PCB, exceeding the 0.339 from explicitly separating PC1 and residual signals.This supports jointly aggregating shared and residual signals rather than treating them as separate additive components.

A.12 Quality–Compute Comparison

The quality–compute analysis compares activation-based and prompting pipelines, emphasizing that A3S’s profile-construction cost can be amortized. Additional verbalized reasoning increases token use but does not improve style fidelity in prompting baselines.

  • Compute Cost: A3S’s main first-use cost is 16 contrastive generations for constructing a new target profile, while Hybrid Search adds relatively little token overhead.The resulting directions can be cached and reused across subsequent inputs.
  • Compute Cost: Profile construction is a one-time cost that can be amortized over later inputs because A3S directions are reusable.Table 13’s A3S costs are driven primarily by constructing the target profile.
  • Reasoning Ablation: StyleCAV decreases from 0.581 to 0.420 for STYLL, from 0.589 to 0.442 for RG, and from 0.516 to 0.389 for Aspect Prompt when explicit reasoning is added.SBERT changes only modestly by +0.025, +0.012, and +0.014, respectively.
  • Reasoning Ablation: Additional verbalized reasoning shifts prompting baselines toward semantic preservation while reducing target-style fidelity.The evaluation uses reasoning-enabled variants of STYLL, RG, and Aspect Prompt under the same backbone and dataset.

B Human Evaluation

The human evaluation directly measures stylistic alignment and meaning preservation through randomized pairwise judgments, complemented by model-based style comparison. Its materials operationalize authorship style across multiple rhetorical dimensions.

  • Evaluation Design: Human evaluation asks annotators to judge both style-transfer success against a target exemplar and meaning preservation against the original input.The two judgments use the same candidate pair and allow Candidate A, Candidate B, Tie, or Neither responses.
  • Evaluation Design: Annotators optionally explain style preferences using Structure, Tone, Vocabulary, Surface, Narrative perspective, and Figurative language.These dimensions provide the rationale categories after the primary style judgment.
  • Evaluation Design: System outputs are anonymized and randomly assigned to Candidate A or Candidate B, with an item-level mapping retained for analysis.Annotators are not informed which system produced either candidate.
  • Evaluation Design: 210 judgments cover 70 instances, with three independent judgments per item and shortest-instance sampling to reduce reading burden.The sampled items were drawn from instances already used for automatic-metric reporting.
  • Activation Construction: A3S constructs aspect-specific contrastive activation vectors from meaning-preserving paraphrases that isolate individual stylistic dimensions.All paraphrases use the same base model and decoding configuration, and prompts request only rewritten text.
  • Baseline Construction: The Aspect-Aware Prompting baseline uses a three-turn chain to analyze six rhetorical dimensions, distill descriptors, and rewrite the source input.The six dimensions are Tone, Figurative Language, Vocabulary, Structure, Syntax, and Perspective.
Loading 2609.04792v1…