PaperlayerView exact revision on arXiv

Source-linked AI summary

Predicting Steering Vectors and Adapter Weights for Few-Shot Author-Style Transfer

Leonard Popp, Danni Liu, Supriti Sinhamahapatra, Jan Niehues

arXiv:2610.03163v1cs.CLcs.AI

TL;DR

Few-shot author-style transfer is difficult, particularly in scientific writing where style and topic are entangled. The paper compares contrastive steering, predicted steering vectors, and predicted LoRA adapters; on seen authors, the hypernetwork offers a style–quality trade-off. The authors also caution that style scores are relative within their setup, not measures of imitation quality.

  • Problem

    Few-shot author-level style control is difficult, and scientific writing makes it harder because formal conventions limit surface variation while author and topic are entangled.

    Source: §1 · p. 1

  • Method

    The paper compares contrastive activation steering, predicted steering vectors, and hypernetwork-predicted LoRA adapters in a three-abstract cold-start setting without history or retrieval.

    Source: §4 · p. 4, §6.2 · p. 6, §8 · p. 9

  • Results

    36.5% style accuracy for Hypernet LoRA came with a 91.3% preference rate, while direct fine-tuning forfeited that preference rate.

    Source: §6.2 · p. 6

  • Takeaways & Limitations

    Manual and predicted steering vectors are near-orthogonal yet score comparably, so vector similarity alone cannot validate steering; downstream metrics are needed.

    Source: §6.2 · p. 6, §8 · p. 9

    Style accuracy is entangled with content and supports only relative rankings within this setup, not a measure of author imitation.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Adapting large language models to an individual author's style from a few examples is challenging, and scientific writing sharpens the difficulty: formal conventions leave little surface variation, and authors write about their own topics, so extracted ``style'' easily entangles with content. We study style-conditioned abstract generation from a few example abstracts per author and propose three methods: (1) contrastive activation steering, (2) a network that predicts steering vectors, and (3) a hypernetwork that predicts LoRA adapters. We find a consistent trade-off between style imitation and output quality: fine-tuning buys most of the available style signal but forfeits fluency, while the hypernetwork achieves the best trade-off on both seen and unseen authors. Our steering operates at author level, contrasting an author's abstracts against style-neutral generations for the same content. This holds topic fixed, removes the need for a predefined style inventory, and outperforms inventory-based steering. % [EDIT 1a] softened "no single optimal axis" claim Moreover, our analyses demonstrate that manually extracted and predicted steering vectors are near-orthogonal yet score comparably, indicating that style conditioning here can admit at least two unrelated directions rather than requiring one particular axis.

1 Introduction

Few-shot author-level style control remains difficult, especially in scientific writing where formal conventions limit surface variation and style is entangled with topic. The paper compares three proposed controls and finds a style–quality trade-off, with hypernetwork LoRA occupying the most efficient position.

1 Introduction

  • Scientific abstracts make author-style extraction difficult because formal conventions leave little surface variation and authors write about their own topics.The resulting style signal can be entangled with content.Source: §1 · p. 1
  • The study compares zero-shot, few-shot, and LoRA baselines with contrastive activation steering, predicted steering vectors, and predicted LoRA adapters using three abstracts per author.The setting is intended to represent author cold start, where neither a user history nor retrieval is available.Source: §1 · p. 1
  • The six methods exhibit a consistent trade-off between style imitation and output quality, with the hypernetwork appearing most efficient.The paper also reports near-orthogonal manual and predicted steering vectors with comparable style accuracy, and author-level contrastive steering matching or exceeding inventory-based steering.Source: §1 · pp. 1–2
  • Prior activation-steering work targets coarse stylistic distinctions or settings with substantial attribute supervision, rather than few-shot individual-author style.Existing hypernetwork adapters condition on task descriptions or profiles assembled from history and retrieval.Source: §1 · p. 1

2 Related Work

Prior work addresses style control and personalization, but the paper targets author-level scientific style in a strict few-shot, cold-start setting. It adapts contrastive steering to matched content and predicts steering vectors or adapters directly from a few examples.

2 Related Work

  • Per-user fine-tuning can achieve strong personalization but scales poorly across users and degrades when histories are limited.This work instead targets cold-start author-level style in scientific writing, where formal constraints leave comparatively little surface variation.Source: §2 · p. 2
  • Prior contrastive steering uses a fixed style inventory, whereas this work contrasts an author's abstracts with style-neutral generations for the same content.The matched-content approach holds topic constant, removes the need for a predefined inventory, and supports author-level steering.Source: §2 · p. 2
  • Unlike stored per-user contrastive vectors, the proposed network predicts steering vectors in one forward pass from three examples, with no per-author artifact to compute or store.The passage frames this as the primary contribution beyond adapting matched-content contrastive steering to author-level scientific writing.Source: §2 · p. 2
  • Prior hypernetwork methods generate adaptation parameters, while related approaches condition on task descriptions or user profiles built from history summaries and retrieval.This work predicts style-conditioned adapters for a fixed task from an aggregated embedding of examples, using three abstracts alone.Source: §2 · p. 2

3 Task and Evaluation

The task is to generate an abstract from a structured outline and three author abstracts, with a frozen base model and no per-author deployment training. Evaluation measures authorship attribution, topic preservation, and text quality, while treating style scores as comparative rather than absolute.

3 Task and Evaluation

  • Given three abstracts by a target author and an outline of an unseen paper, the model generates an abstract in that author's style; the original abstract is not supplied as input.The base model is frozen, deployment requires no per-author training, and evaluation includes authors seen in training and held-out authors.Source: §3 · p. 2
  • For unseen-author attribution, the evaluation uses k-NN in a learned embedding space so new authors can be introduced at inference without retraining.It reports accuracy and mean reciprocal rank using example abstracts as the reference database.Source: §3 · p. 2
  • Figure 1 shows that both control networks embed the author examples and condition an MLP on layer information, but their output heads differ.One emits a residual-stream vector; the other emits LoRA weights for attention projections.Source: Figure 1 · p. 3
  • Style and content are not fully separable in the attribution metric, so style accuracy is used to compare models rather than as an absolute measure.Abstracts generated without author information are attributed above chance.Source: §3 · pp. 2–3
  • A binary LLM judge checks topic preservation by deciding whether a generated abstract and the paper introduction originate from the same document.The judge is GPT-OSS-20B.Source: §3 · p. 3
  • Text quality is measured by pairwise preference against the human-written abstract, with each pair judged twice in swapped order to cancel position bias.The authors treat differences between preference rates, rather than their absolute level, as informative.Source: §3 · p. 3

4 Style Control Methods

The section develops manual contrastive steering, a network that predicts steering vectors, and a hypernetwork that predicts LoRA adapters. The methods differ in their conditioning and deployment requirements, and the adapter update varies with the model's layer input.

Manual Contrastive Steering

  • Manual contrastive steering subtracts the model's neutral-generation activations for a reference paper from the author's reference activations, then averages the three per-reference vectors for injection.The neutral generation uses the same paper content, and the steering intensity and target layers are selected on validation.Source: §4 · p. 3
  • Manual steering requires one neutral generation and the corresponding paper outline per reference, so it consumes more than the three abstracts available in the deployment setting.It also requires a validation search over intensity; the test configuration uses λ = 0.875.Source: §4 · p. 3

Learned Steering

  • The learned steering network predicts a residual-stream vector directly from the reference set, avoiding manual steering's neutral generations and intensity search.The references are embedded with a frozen STAR encoder, then projected and averaged into an author representation; the network learns per-layer magnitudes.Source: §4 · p. 4
  • A single predicted vector per layer shifts every token uniformly, regardless of what is being written.The method therefore motivates the subsequent input-dependent LoRA intervention.Source: §4 · p. 4

Hypernetwork for LoRA Weights

  • The hypernetwork predicts LoRA adapters whose updates depend on the layer input, allowing the intervention to vary with what the model is writing.It conditions on layer and module embeddings and predicts adapters for q_proj and v_proj.Source: §4 · p. 4

Training

  • Both learned methods train end-to-end with the base model and encoder frozen; the model receives the zero-shot prompt, so style enters through predicted parameters.The adapter heads use a no-op initialization, and restricting adapted layers to a middle-layer subset acts as a regularizer and improves downstream metrics.Source: §4 · p. 4

5 Experimental Setup

The setup describes how the paper builds its abstract-generation datasets, compares personalization baselines, and configures its models and evaluation checks.

5.1 Data Creation

  • The filtered arXiv corpus contains 1,232 papers from 203 authors, with 826 training papers and one validation and test paper per author.Papers are from permissively licensed cs.AI and cs.LG categories through 2022; each outline is paired with its abstract, introduction, and three same-author examples.Source: §5.1 · p. 4
  • The unseen-author test set contains 1,312 papers from 328 authors, with three examples and one test sample per author.Each author has exactly four papers, disjoint from the base dataset.Source: §5.1 · pp. 4–5

5.2 Baselines

  • The baselines range from outline-only zero-shot and three-example few-shot prompting to shared rank-8 LoRA adapters conditioned on author examples.The LoRA adapters are optimized across training authors rather than separately per author.Source: §5.2 · p. 5

5.3 Implementation Details

  • All methods use a frozen Qwen3-4B base; learned steering predicts shifts for all 36 layers, while the hypernetwork predicts rank-8 query and value adapters.Both learned methods encode the three example abstracts with STAR; manual steering uses layers 18–20.Source: §5.3 · p. 5
  • The topic-preservation judge achieves 99.3% accuracy, misclassifying 18 of 2,464 abstract–introduction pairs.Negative pairs are created by pairing abstracts with unrelated introductions.Source: Table 1 · p. 5
  • Removing style while holding content fixed costs 22.2 points in author-classifier accuracy, while 31.0% remains attainable from content alone.Styling a different author’s paper falls to chance, showing that content and style source are varied separately in the classifier evaluation.Source: Table 2 · p. 5

6 Quantitative Results

The hypernetwork offers the strongest style–quality balance across seen and unseen authors, while direct fine-tuning gains more style accuracy at a substantial quality cost. Metric checks also show that the style classifier responds to content as well as author style.

6.1 Metric Validation

  • The topic-preservation judge achieves 99.3% accuracy on 2,464 abstract–introduction pairs; the style classifier scores 53.2% on held-out abstracts versus 31.0% on style-neutral abstracts.The 31.0% residual is not a pure content component, and neither classifier score isolates style or content alone.Source: §6.1 · pp. 5–6

6.2 Results on Seen Authors

  • On seen authors, hypernetwork LoRA reaches 36.5% style accuracy with 91.3% preference rate; direct LoRA reaches 49.8% style accuracy but preference rate falls to 38.1% from zero-shot’s 97.9%.Manual steering reaches 34.5% style accuracy at a 94.3% preference rate, while predicted steering reaches 35.0% at 89.6%.Source: Table 3 · p. 6, §6.2 · p. 6

6.3 Generalization to Unseen Authors

  • On unseen authors, hypernetwork LoRA gains 8.5 style-accuracy points over zero-shot for 1.9 points less preference rate; LoRA again has highest style accuracy at a 39.0% preference rate.The trade-off structure reappears, with hypernetwork LoRA ranking second on all three metrics.Source: Table 3 · p. 6, §6.3 · p. 7

6.4 Style-Quality Trade-Off

  • The hypernetwork is the only proposed approach Pareto-optimal on both author sets at the reported configurations, between zero-shot quality and LoRA style accuracy.At the common layer scope 18–20 on seen authors, the hypernetwork has higher style accuracy (36.5% vs. 34.0%), while learned steering has higher preference rate (97.5% vs. 91.3%), so neither dominates there.Source: Figure 2 · p. 7, §6.4 · p. 7

7 Analyses

The analyses find that steering vectors with comparable style accuracy can point in unrelated directions, while layer-scope effects vary across interventions. Narrow scopes improve the hypernetwork’s results, but learned steering retains strong performance at full depth.

7.1 No Single Style Axis

  • Manual and predicted steering vectors have near-zero cosine similarity at layers 18–20 (−0.03, −0.05, −0.04) despite comparable style accuracy, so matching the manual direction is unnecessary.This does not rule out a privileged direction; the authors say that testing for one requires comparing vectors from disjoint same-author reference subsets against vectors from different authors.Source: §7.1 · p. 8, Table 4 · p. 8
  • Predicted vectors are dense while manual vectors are sparse, although both perturbations have L2 norms between 18% and 36% of the corresponding hidden activation norm.Predicted-vector L1/L2 values are 39.5–40.1, compared with 17.3–19.4 for manual vectors.Source: §7.1 · p. 8, Table 4 · p. 8

7.2 Layer Selection

  • Manual steering performs best in a narrow middle band: layers 18–20 reach 36.0% style accuracy versus 35.0% at 20–22, while all 36 layers yield 2.5%.The authors suggest that accumulating perturbations across the residual stream may push hidden states away from the model’s training distribution; they also note that mid-depth may balance later reshaping against having too few layers to realize a late intervention.Source: §7.2 · pp. 8–9, Table 5 · p. 8
  • Restricting the hypernetwork to layers 18–20 raises style accuracy to 36.5% vs. 34.5%, MRR to 0.46 vs. 0.43, and preference rate to 91.3% vs. 71.4% over full depth.Only these two scopes were evaluated, so layers 18–20 are not established as optimal.Source: Table 5 · p. 8, §7.2 · p. 9
  • At full depth, learned steering reaches 35.0% style accuracy and 89.6% preference; its 18–20 scope keeps accuracy near 34.0% while raising preference to 97.5%.The layer-scope pattern differs from the hypernetwork’s: narrowing learned steering improves preference rather than style accuracy.Source: Table 5 · p. 8, §7.2 · p. 9

8 Conclusion

The paper studies author-level style conditioning in a strict cold-start setting and recommends treating it as a trade-off across interventions. Its analyses caution against seeking a single steering direction and recommend using same-content model defaults as contrastive negatives.

8 Conclusion

  • The study uses three abstracts, no history or retrieval, and a frozen base model to examine author-level style conditioning.The conclusion frames style conditioning as navigating a trade-off paid by each intervention.Source: §8 · p. 9
  • Near-orthogonal manual and learned vectors support comparable style accuracy, so refining a contrast set may tune an arbitrary direction and vector similarity alone cannot validate a method.The authors recommend judging methods by downstream metrics instead.Source: §8 · p. 9
  • For contrastive steering, the negative should be the model’s own style-neutral generation for the same content the output must deviate from.This keeps the contrast tied to the desired deviation rather than a predefined style inventory.Source: §8 · p. 9

Limitations

The authors identify limitations in measuring style, attributing gains specifically to authorship, selecting intervention scope, and generalizing beyond their experimental configuration.

Limitations

  • Because style-neutral generations receive above-chance authorship attribution, style accuracy supports only ordinal comparisons within this setup, not absolute measurement of author imitation.The authors say disentangling style from content requires a content-invariant classifier and a probe that isolates the generator from the metric.Source: Limitations · p. 9
  • Without a control for generic shifts toward human-written abstracts, the results do not establish that learned interventions are author-specific.Within- versus between-author similarity of predicted vectors and effective updates remains unmeasured.Source: Limitations · p. 9
  • Only two intervention scopes were evaluated, establishing that a narrow band beats full depth but not that the selected band is optimal.A full sweep over bands and widths per method remains future work.Source: Limitations · p. 9
  • Evaluation uses one base model, one broad arXiv domain, and English only, so whether the observed trade-off generalizes remains open.The authors call for replication across model scales, families, domains, and languages.Source: Limitations · p. 9

System Prompt

The system prompt asks GPT-OSS-20B to judge whether an abstract and introduction describe the same topic. It defines binary response rules and specifies one-word output.

System Prompt

  • The prompt compares an abstract with an introduction to determine whether they describe the same specific subject matter or topic.The figure caption says the abstract placeholder is replaced with the paper summary and the introduction placeholder with the introduction used for comparison.Source: System Prompt · p. 12, Figure 3 · p. 12
  • The prompt requires “True” for matching core content and “False” for unrelated, different, or contradictory topics.Source: System Prompt · p. 12
  • The response must be exactly one word, either “True” or “False,” with no explanation.Source: System Prompt · p. 12

A Prompt Templates

The appendix describes prompts for creating paper outlines and for generating abstracts with or without author examples. Long papers are chunked and summarized in two passes, while shorter papers are summarized directly.

A.1 Topic Preservation Judge Prompt

  • The topic-preservation prompt asks GPT-OSS-20B whether a generated abstract and the paper introduction describe the same document, returning a binary label.It decodes greedily with the default medium reasoning effort.Source: §A.1 · p. 12

A.2 Preference Rate Judge Prompt

  • The quality prompt asks GPT-OSS-20B to compare two abstracts for the same paper on clarity, scientific rigor, and narrative flow.The judge decodes greedily with the default medium reasoning effort.Source: §A.2 · p. 12
  • For topic-preservation evaluation, prompts are capped at 3,500 tokens; only the introduction is truncated when needed, leaving abstracts of up to 242 words intact.Source: §A.2 · p. 12

A.3 Outline Creation Prompt

  • Long papers are split into chunks of up to 7,750 tokens with 250-token overlap; each chunk receives a compact bulleted outline before a second pass combines them.Source: §A.3 · p. 12
  • Papers that fit the context window are summarized directly, without chunking, using the plain-text paper input prompt.All outline creation uses 4-bit Qwen3-4B-Instruct-2507.Source: §A.3 · p. 12

A.4 Abstract Generation Prompt

  • The zero-shot abstract prompt requests a formal academic abstract from a structured outline without information about the target author.Source: §A.4 · p. 12
  • The few-shot prompt adds three target-author abstracts and asks the model to adopt that style while staying faithful to the outline.Source: §A.4 · p. 12

B Implementation Details

The implementation uses prompt templates for evaluation, outline creation, and abstract generation, alongside tuned classifier, LoRA, and style-control components.

B Implementation Details

  • GPT-OSS-20B compares generated and human-written abstracts for the same paper on clarity, scientific rigor, and flow, evaluating each pair in both orders.Source: Figure 4 · p. 13, B Implementation Details · p. 13
  • For chunked papers, one prompt summarizes each text chunk and another combines the chunk outlines using one numbered section block per chunk.The ellipsis in the combination prompt is schematic.Source: Figure 5 · p. 13, Figure 6 · p. 14
  • The direct outline prompt receives the paper’s plain text when it fits the model context window.Source: Figure 7 · p. 14
  • Zero-shot, manual steering, learned steering, and hypernetwork prompts provide an outline without author information; few-shot and LoRA prompts add three author abstracts as style examples.Source: Figure 8 · p. 15, Figure 9 · p. 15
  • The longest inputs are 1,851 tokens for outline-only and 2,740 with three examples, so no input truncation occurs.Qwen3-4B-Instruct-2507 is loaded in full precision for generation and the judge in 4-bit.Source: B Implementation Details · p. 16
  • The selected authorship classifier is fine-tuned distilbert-base-uncased k-NN with k = 3 and cosine distance.For unseen authors, its database contains three reference abstracts per author.Source: B Implementation Details · p. 16
  • The LoRA baseline uses q_proj and v_proj adapters with r = 8 and α = 16, trained with a 2×10−4 learning rate and early stopping.Cross-entropy is computed on the abstract only, with the prompt masked.Source: B Implementation Details · p. 16
  • Both style-control models use a four-block MLP; the steering network has a single output head, while the hypernetwork has four LoRA-matrix heads.Their inputs are 164-dimensional for steering and 196-dimensional for the hypernetwork, which also encodes the target module.Source: B Implementation Details · pp. 16–17
Loading 2610.03163v1…