New LLM Framework Uses Style Transfer to Identify Authorship Without Labeled Training Data

Researchers have developed a novel approach to authorship identification by tapping into the capabilities of large language models in ways that don't require labeled training examples. The framework measures how easily a text's writing style can be transferred between documents using an LLM's internal probability distributions.

The Core Approach#

Computational stylometry—the scientific study of writing style patterns—has long tackled problems like authorship attribution and plagiarism detection. The new method adds a machine learning twist by using log-probabilities from causal language models (a standard pre-training approach for LLMs) to assess whether two texts share the same authorial voice. Because the approach avoids explicit supervision, it can operate on diverse texts without requiring paired examples where authors' works are already known.

Key Results#

The framework demonstrates strong performance across several fronts:

• Substantially outperforms systems built around LLM prompting alone when applied to authorship verification
• Matches or surpasses contrastive learning baselines in most scenarios, particularly when using larger models
• Shows consistent gains as model size increases, suggesting the method becomes more reliable with stronger language models
• Generalizes well to non-English languages

Flexible Trade-offs#

For authorship verification specifically, the team introduces an optional mechanism that performs additional computation at test time to boost accuracy. This provides practitioners with a dial they can adjust based on whether they prioritize speed or correctness.

Why It Matters#

By avoiding the need for supervised training data, this approach opens authorship analysis to scenarios where labeled examples are scarce. The scaling benefits also hint that future, more capable language models may make these tools even more effective for security and forensic applications.