EDITORS Know Your Style! Editing LoRA Subspaces for Stylistic Attribution and Imitation
Organizations: School of Computing, National University of Singapore, Singapore · Agency for Science, Technology and Research (A*STAR), Singapore · Aix Marseille Univ, CNRS, I2M, Marseille, France
Abstract
Disentangling writing style from semantic content is a fundamental challenge in literary text modeling. Style-content entanglement causes models to rely on semantic content for authorship attribution (AA) tasks and memorize author-specific content for style imitation (SI) tasks. We propose EDITORS (EditingLoRA Subspaces), a diagnose-then-deflate framework that adapts pretrained large language models and prioritizes stylistic features, making it effective for both AA and SI tasks. EDITORS first trains a diagnostic LoRA adapter using style-neutral statements derived from the training corpus to identify a content-oriented subspace. It then trains the final style adapter under activation-space deflation that projects input activations away from the identified content-oriented directions. We empirically demonstrate that EDITORS is effective. On three AA benchmarks spanning literature, social media, and news, EDITORS achieves state-of-the-art accuracy, outperforming the strongest existing baseline by up to 29%. On SI, EDITORS generates stylistically faithful and diverse passages, outperforming baselines in achieving style similarity while adhering to content specifications.
Figures & tables
| Object | Trained on | Purpose | Used at inference? | Initialization |
|---|---|---|---|---|
| Pretraining | Frozen base model | Yes | – | |
| (optional) | Domain/task warm-up | No | ||
| Style-neutral | Diagnose content directions | No | ||
| Derived from | Deflate content directions | Yes | SVD of | |
| (under deflation) | Final style adapter | Yes | Task-dependent |
| Method | CCAT10 | CCAT50 | Blogs10 | Blogs50 | Gutenberg30 |
|---|---|---|---|---|---|
| BertAA | 84.00 | 65.72 | 65.40 | 59.69 | 42.76 |
| STAR | 91.75 | 67.33 | 73.40 | 65.43 | 53.30 |
| StyleEmbedding | 18.20 | 4.28 | 40.67 | 15.60 | 5.00 |
| StyleDistance | 20.60 | 8.00 | 52.33 | 15.40 | 11.67 |
| GPT-4 + LIP | 57.80 | – | 68.00 | – | – |
| GPT-5-nano + LIP | 57.00 | – | 35.18 | – | – |
| Backbone | Setting | CCAT10 | CCAT50 | Blogs10 | Blogs50 | Gutenberg30 | ||
|---|---|---|---|---|---|---|---|---|
| BERT | BertAA | 84.00 | 65.72 | 65.40 | 59.69 | 42.76 | – | – |
| EDITORS | 90.75 | 77.47 | 92.68 | 83.39 | 43.57 | +14.1 | 10.0 | |
| Phi-4-mini-3.8B | SFT | 93.70 | 69.07 | 91.25 | 71.60 | 48.03 | – | – |
| EDITORS | 95.70 | 70.40 | 92.08 | 77.74 | 53.37 | +3.1 | 2.2 | |
| Mistral-7B | SFT | 93.00 | 80.53 | 94.17 | 75.79 | 72.37 | – | – |
| EDITORS | 97.25 | 86.80 | 95.77 | 85.65 | 73.00 | +4.5 | 3.3 |
| Model | Style Similarity | BLEU | Self-BLEU |
|---|---|---|---|
| GPT-4 | 23.93 8.55 | 4.42 1.20 | 27.35 3.49 |
| StyleTunedLM | 53.39 9.43 | 40.38 5.42 | 37.48 5.60 |
| EDITORS | 56.50 8.74 | 27.09 4.76 | 24.35 4.90 |
| Model | Style Similarity | Content Preservation | BLEU | Self-BLEU |
|---|---|---|---|---|
| GPT-4 | 10.82 0.47 | 99.88 0.06 | 16.54 0.12 | 11.03 0.13 |
| StoryTrans | 13.44 0.52 | 0.68 1.12 | 4.96 0.95 | 14.36 0.38 |
| ASTRAPOP | 46.03 0.56 | 3.73 0.49 | 27.08 0.30 | 14.85 0.50 |
| StyleTunedLM | 44.34 0.97 | 59.63 1.34 | 47.18 0.16 | 20.54 0.39 |
| EDITORS | 55.08 0.56 | 68.16 1.27 | 27.42 0.64 | 13.14 0.55 |
| Model | Style Similarity | Content Preservation | BLEU |
|---|---|---|---|
| EDITORS w/ | 54.10 0.68 | 39.23 2.54 | 45.64 0.26 |
| EDITORS w/o | 55.08 0.56 | 68.16 1.27 | 27.42 0.64 |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Split Size (Passages) | Train Dist. (Per Author) | Length | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Dataset | Domain | Train | Val | Test | Avg | Min | Max | Avg (Words) | ||
| Gutenberg30 | Literature | 12,000 | 300 | 3,000 | 400 | 400 | 400 | 700 | ||
| Blogs50 | Social Media | 42,874 | 5,346 | 5,346 | 643 | 424 | 1,967 | 161 | ||
| CCAT50 | News | 4,000 | 250 | 750 | 80 | 80 | 80 | 504 | ||
| ID | Representative genre | Content specification | Authors |
|---|---|---|---|
| A | Children / youth moral tale | A child learns that telling the truth is better than hiding a mistake. | Charlotte Mary Yonge; Jacob Abbott. |
| B | Adventure / survival / exploration | A group works together to survive after getting stranded. | Charles Kingsley; Daniel Defoe; Edward Stratemeyer; George Alfred Henty; Harold Bindloss; Jack London; James Otis; R. M. Ballantyne; Robert Louis Stevenson; Rudyard Kipling. |
| C | Mystery / detective / suspense | A missing item is found by careful thinking, not force. | Edward Phillips Oppenheim; G. K. Chesterton; Wilkie Collins. |
| D | Social realism / local color | A small-town rule causes trouble, and someone speaks up. | Bret Harte; Hamlin Garland; John Galsworthy; John Morley; William Dean Howells. |
| E | Satire / wit / comic society | A silly rule is taken too seriously and leads to a mess. | Frank Richard Stockton; George Bernard Shaw; James Matthew Barrie; Jerome Klapka Jerome; Oscar Wilde; William Makepeace Thackeray; William Wymark Jacobs. |
| F | Speculative science fiction / social thought | A new technology spreads too fast and causes problems. | Herbert George Wells. |
| Dataset | EDITORS w/o | EDITORS w/ |
|---|---|---|
| CCAT50 | 66.80 | 80.40 |
| Blogs50 | 62.07 | 83.80 |
| Gutenberg30 | 73.30 | 82.33 |
| Model | Llama | Gemma | Phi4 | Mistral | STAR | BertAA | StyleSim Avg. | BLEU | Self-BLEU | |
|---|---|---|---|---|---|---|---|---|---|---|
| EDITORS w/ | 63.67 | 67.89 | 60.78 | 60.56 | 39.00 | 67.00 | 59.82 | 10.64 | 47.17 8.28 | 34.83 6.62 |
| EDITORS w/o | 59.33 | 60.22 | 55.22 | 56.11 | 40.89 | 67.22 | 56.50 | 8.74 | 27.09 4.76 | 24.35 4.90 |
| GPT-4 | 18.00 | 22.73 | 27.82 | 15.45 | 20.46 | 39.09 | 23.93 | 8.55 | 4.42 1.20 | 27.35 3.49 |
| StyleTunedLM | 56.56 | 58.11 | 52.11 | 55.11 | 35.56 | 62.89 | 53.39 | 9.43 | 40.38 5.42 | 37.48 5.60 |
| Model | Llama | Gemma | Phi4 | Mistral | STAR | BertAA | Avg. | |
| GPT-4 | 5.79 | |||||||
| ASTRAPOP | 11.34 | |||||||
| StoryTrans | 6.95 | |||||||
| StyleTunedLM | 11.96 | |||||||
| EDITORS w/o | 56.67 0.67 | 63.33 1.40 | 55.08 2.34 | 56.83 1.12 | 36.19 0.45 | 62.38 1.35 | 55.08 0.56 | 9.84 |
| EDITORS w/ | 56.19 2.43 | 60.16 2.65 | 56.67 1.03 | 56.83 1.12 | 33.81 0.53 | 60.95 0.98 | 54.10 0.68 | 10.14 |
| Model | StyleDistance | StyleEmbedding |
|---|---|---|
| GPT-4 | 4.92 1.10 | 6.03 0.55 |
| StoryTrans | 9.37 0.99 | 7.94 1.45 |
| ASTRAPOP | 17.30 1.45 | 14.60 2.44 |
| StyleTunedLM | 7.78 2.15 | 11.11 1.10 |
| EDITORS | 22.70 2.15 | 19.52 2.65 |
| Model | GPT-4o mini | Qwen3.5-27B | NLI-based entailment rate |
|---|---|---|---|
| GPT-4 | 100.00 0.00 | 99.76 0.12 | 16.18 3.23 |
| StoryTrans | 0.83 1.34 | 0.52 0.89 | 3.49 2.79 |
| ASTRAPOP | 4.40 0.48 | 3.06 0.50 | 3.65 1.10 |
| StyleTunedLM | 66.79 1.49 | 52.46 1.19 | 4.92 1.67 |
| EDITORS | 72.54 1.35 | 63.77 1.19 | 10.48 0.48 |
| Model | OR@25 (%) Target author | OR@25 (%) Global corpus | Maximum copied span (characters) |
|---|---|---|---|
| GPT-4 | 0.10 0.12 | 0.94 0.04 | 39 |
| StoryTrans | 0.01 0.01 | 0.09 0.04 | 25 |
| ASTRAPOP | 1.74 0.39 | 6.88 0.32 | 1934 |
| StyleTunedLM | 1.11 0.04 | 6.32 0.33 | 51 |
| EDITORS | 0.74 0.01 | 4.64 0.25 | 44 |
| Model | Style Sim. (%) | Content Pres. | BLEU | Self-BLEU | LCS max |
|---|---|---|---|---|---|
| GPT-4 | 10.82 0.47 | 99.88 0.06 | 16.54 0.12 | 11.03 0.13 | 9.67 1.53 |
| ASTRAPOP | 46.03 0.56 | 3.73 0.49 | 27.08 0.30 | 14.85 0.50 | 161.00 240.92 |
| StoryTrans | 13.44 0.52 | 0.68 1.12 | 4.96 0.95 | 14.36 0.38 | 9.00 0.58 |
| StyleTunedLM | 44.34 0.97 | 59.63 1.34 | 47.18 0.16 | 20.54 0.39 | 12.00 0.00 |
| EDITORS w/ | 54.10 0.68 | 39.23 2.54 | 45.64 0.26 | 21.07 0.43 | 14.33 1.53 |
| EDITORS w/o | 55.08 0.56 | 68.16 1.27 | 27.42 0.64 | 13.14 0.55 | 12.67 1.53 |
| Method | Style Sim. (%) | Content Pres. | BLEU-4 | Self-BLEU | OR@25 |
|---|---|---|---|---|---|
| StyleTunedLM | 22.14 1.00 | 86.43 1.30 | 36.62 1.60 | 15.50 0.52 | 0.92 0.14 |
| EDITORS | 37.62 1.33 | 94.29 0.36 | 35.02 1.52 | 11.74 1.38 | 0.75 0.20 |
| Dataset | AA accuracy on | AA accuracy on |
|---|---|---|
| CCAT50 | 52.80 | 80.40 |
| Blogs50 | 33.67 | 83.80 |
| Gutenberg30 | 27.17 | 82.33 |
Authorship Attribution
Authorship attribution identifies or verifies who wrote a text, code sample, or other creative work from patterns in its style and content.
See all 71 papers