AuthorMix: Modular Authorship Style Transfer via Layer-wise Adapter Mixing

1. Problem Statement

Authorship Style Transfer involves rewriting text to match the specific writing style of a target author while strictly preserving the original semantic meaning. Existing methods face three primary challenges:

• Meaning Preservation vs. Style Transfer: High-fidelity style transfer often leads to "semantic drift," where the rewritten text loses the original meaning.
• Low-Resource Adaptation: Most state-of-the-art (SoTA) methods require training a single, monolithic model on massive corpora containing thousands of author pairs. This makes adapting to a new, specific target author (especially one with few available texts) costly and inflexible.
• Interpretability: Existing approaches (e.g., embedding manipulation or policy optimization) act as "black boxes," offering little insight into how specific stylistic features are learned or combined.

2. Methodology: AuthorMix

The authors propose AuthorMix, a lightweight, modular, and interpretable framework that decomposes style transfer into three stages. Instead of training a new model for every target, it constructs a target-specific model by mixing pre-trained adapters.

Stage 1: Author Adapter Training

• Base: A large language model (LLM) is used as a frozen base (LLaMA-3.1-8B-Instruct).
• Adapters: Individual LoRA (Low-Rank Adaptation) adapters are trained for a small set of "high-resource" authors (e.g., 10 classic authors with abundant text).
• Training Data: Pseudo-parallel data is generated by paraphrasing the high-resource author's texts into a "neutral" style (removing stylistic markers but keeping meaning) using a large LLM. The LoRA adapter is then fine-tuned to reconstruct the original stylistic text from the neutral version.

Stage 2: Layer-wise Weight Mixing

For a new low-resource target author (with only a handful of example texts), AuthorMix does not train a new model from scratch. Instead, it creates a target adapter ($\hat{\theta}$) by linearly mixing the pre-trained high-resource adapters ($\theta^{(i)}$).

• Granularity: Unlike previous work that mixes adapters at a global level, AuthorMix performs layer-wise mixing. For a model with $L$ layers, the weights for the target adapter at layer $j$ are a linear combination of the weights of the $n$ source adapters at that same layer:
  $$\hat{\theta}_j = \sum_{i=1}^{n} W_{i,j} \cdot \theta^{(i)}_j$$
• Optimization: The mixing weights ($W$) are optimized to maximize a Joint Score, which is the geometric mean of:
  1. Toward Score: Stylistic similarity to the target author.
  2. Mutual Implication Score (MIS): Semantic similarity to the source text.
• Optimization Strategies:
  • LoRAHub: Uses gradient-free optimization (NGOpt) to find weights.
  • GRPO (Group Relative Policy Optimization): A novel application of Reinforcement Learning (RL) to learn the mixing weights. The Joint Score serves as the reward function.

Stage 3: Adapter Subset Selection

To improve efficiency and performance, the system does not mix all available adapters. It first calculates style similarity (via cosine distance of style embeddings) between the target and the high-resource authors, selecting only the top-$k$ most similar adapters for the mixing process.

3. Key Contributions

1. Modular Framework: AuthorMix enables the rapid creation of a style-transfer model for any target author by mixing a library of pre-trained adapters, eliminating the need for massive retraining.
2. Layer-wise Mixing: The paper introduces and validates layer-wise weight mixing over adapter-wise mixing. This allows the model to capture author-specific characteristics that reside at different depths of the neural network.
3. RL-based Optimization: The authors are the first to apply GRPO (a reinforcement learning algorithm) to learn adapter mixing weights, demonstrating it outperforms gradient-free methods (LoRAHub) in balancing style and meaning.
4. Interpretability: The learned layer-wise weights reveal where in the network specific stylistic traits are encoded (e.g., middle layers often contribute positively to style, while deep layers may be suppressed).

4. Experimental Results

Experiments were conducted on 100 source-target author pairs using Project Gutenberg literary texts.

• Performance: AuthorMix (specifically the GRPO layer-wise variant with $k=4$) achieved the highest Joint Score (0.34), outperforming SoTA baselines (ASTRAPOP, TinyStyler, STYLL) and even GPT-5.1 (few-shot).
• Meaning Preservation: AuthorMix achieved the highest MIS (0.83) with the lowest variance among all methods. This indicates it preserves semantic meaning significantly better than methods that rely on neutral pivots or aggressive policy optimization, which often degrade meaning.
• Style Transfer: It matched the style transfer capability (Toward Score) of the best baselines (0.16).
• Efficiency:
  • Training Cost: Adapting to a new target takes only ~40 minutes on a single GPU (learning ~128 scalar weights).
  • Comparison: This is drastically faster than ASTRAPOP-DPO (~48 hours) or TinyStyler (>48 hours).
• Ablation Findings:
  • Layer-wise > Adapter-wise: Layer-wise mixing consistently outperformed global adapter mixing.
  • Optimization: GRPO significantly outperformed LoRAHub in layer-wise scenarios, likely because gradient-free methods struggle with the high-dimensional search space of layer-wise weights.
  • Subset Size: Performance peaked at $k=4$ (mixing 4 similar authors) and degraded with too many adapters, confirming the importance of selecting a relevant subset.

5. Significance and Implications

• Scalability: AuthorMix solves the "low-resource" bottleneck in style transfer. It allows for the creation of specialized models for niche authors or new targets with minimal data and compute.
• Semantic Fidelity: By explicitly optimizing for meaning preservation alongside style, it addresses a critical failure mode of current style transfer models (semantic drift).
• Interpretability: The layer-wise mixing weights provide a window into the internal mechanics of LLMs, showing that different layers encode different aspects of an author's style.
• Ethical Considerations: The authors acknowledge the risk of impersonation but position the work as a research tool for personalized adaptation, emphasizing the need for safeguards in deployment (e.g., transparency, consent).

In conclusion, AuthorMix represents a shift from monolithic, expensive style transfer models to a modular, efficient, and interpretable approach that leverages the strengths of parameter-efficient fine-tuning (LoRA) and advanced optimization techniques (RL) to achieve superior results in low-resource settings.