Authorship Attribution (AA) requires capturing fine-grained stylistic characteristics, making it particularly challenging in zero-shot (ZS) settings where no task-specific supervision is available. In this work, we investigate the effect of author representations on ZS AA by evaluating a label-only prompting baseline together with three author representation strategies: representative writing samples, LLM-generated descriptions, and style embeddings (LISA). The first three approaches perform attribution using LLM prompting, while the embedding-based approach uses style embeddings with cosine similarity. We investigate the influence of prompt design and propose a two-stage embedding-based attribution framework that combines candidate space reduction with embedding-dimension selection. The results show that label-only ZS AA is ineffective, while incorporating author-specific representations consistently improves attribution performance. Among the evaluated approaches, the proposed two-stage LISA framework achieves the strongest overall performance, whereas LLM-generated style descriptions provide a substantially more compact representation of author style at the cost of some attribution performance. These findings demonstrate the importance of author representation in ZS AA, while indicating that current open-source LLMs remain insufficient for robust attribution without more effective representation learning.
Figures & tables
Figure 1: Confusion matrices for Label Only ZS AA showing selection bias in Mixtral: (a) 3-authors, (b) 15-authors.
Subset
nrows
nauthors
Mixtral
Gemma
Qwen
df3
30
3
30.0
43.3
30.0
df5
50
5
22.0
28.0
16.0
df10
100
10
12.0
12.0
6.0
df15
150
15
3.3
8.0
7.3
Table 1: Label-only ZS AA Accuracy across three open-source LLMs under different numbers of candidate authors.
Figure 2: Confusion matrices for label-only ZS AA on the 10-author subset using three open-source instruction-tuned LLMs . All three models exhibit prediction concentration, although the distribution of predictions across candidate authors differs substantially between models.
Subset
AA Prompt
Acc
W-F1
μ F1
M-F1
Unsure
Mixtral
Huang et al. (2024) subset*
no_guidance
16.6
11.9
16.6
11.0
8
little_guidance
6.6
4.1
6.6
3.7
11
grammar
10.0
6.5
10.0
5.8
10
LIP
13.3
9.7
13.3
8.8
6
Top 10 authors
no_guidance
10.0
7.5
10.0
7.0
1
Table 2: Definition-enhanced ZS AA using representative writing samples across three sample-selection settings and three open-source LLMs .
Description P
AA Prompt
Acc (%)
W-F1
μ F1
M-F1
Mixtral
P1
no_guidance
23.3
14.3
23.3
14.3
little_guidance
23.3
14.7
23.3
14.7
grammar
23.3
15.2
23.3
15.2
LIP
26.6
16.0
26.6
16.0
P2
no_guidance
16.6
6.7
16.6
6.7
Table 3: Definition-enhanced ZS AA across three open-source LLMs . Style descriptions are generated using 4 description prompts; P1(Baseline), P2(Content-agnostic), P3(Linguistic) and P4(Discriminative) explained in Table 6 and evaluated using the attribution prompts in Table 7 .
Setting
Method
Stage-1 Recall (%)
Acc (%)
W-F1
μ F1
M-F1
k =2, N =20
Single-stage
–
36.6
34.67
36.6
34.6
2-stage: P
56.67
43.3
44.5
43.3
44.5
2-stage: A
56.6
40.0
37.57
40.0
37.5
2-stage: P+A
56.6
36.6
35.1
36.6
35.1
k =3, N =20
Single-stage
–
36.6
34.6
36.6
34.6
2-stage: P
73.3
40.0
39.2
40.0
39.2
Table 4: Definition-Enhanced ZS AA: Style embeddings as definitions under different values of k and N . P denotes presence and A denotes absence-based selection.
Figure 3: Comparison of the highest attribution accuracy achieved across the evaluated AA approaches.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Prompt Type
Description
ZS (Label-only)
You are a helpful assistant for author attribution. Given the following text, identify who is the most likely author. Choose ONLY from the following IDs: <author IDs>. Text: <text>. Answer:
Appendix
Table 5: ZS -shot (ZS) label-only prompt without definitions used for AA .
Prompt
Description
P1
you are an expert in writing style extraction, Given the writing samples of the author, extract writing style from it and write it in one consise paragraph.
P2
P1 + Focus purely on how the author writes, not what they write.
P3
P1 + Pay special attention to stylistic signals such as sentence length, grammatical looseness or rigidity, punctuation patterns, slang usage, formality level, and lexical choices.
P4
P1 + Highlight the most distinctive stylistic features that uniquely identify the author and distinguish them from others.
Appendix
Table 6: Prompts used for generating author descriptions.
Prompt
Description
System Instruction
Respond with a JSON object including two elements: “analysis” (reasoning behind the answer) and “answer” (the predicted author ID of the query text).
No Guidance
Given a set of texts with known authors and a query text, determine the author of the query text. Input query text: <query text>; Texts from potential authors: <candidate texts>.
Style Guidance
Given a set of texts with known authors and a query text, determine the author of the query text. Do not consider topic differences. Input query text: <query text>; Texts from potential authors: <candidate texts>.
Grammar Guidance
Given a set of texts with known authors and a query text, determine the author of the query text. Focus on grammatical styles. Input query text: <query text>; Texts from potential authors: <candidate texts>.
LIP
Given a set of texts with known authors and a query text, determine the author of the query text. Analyze writing styles while disregarding topic and content. Focus on linguistic features such as phrasal verbs, modal verbs, punctuation, rare words, affixes, quantities, humor, sarcasm, typographical errors, and misspellings. Input query text: <query text>; Texts from potential authors: <candidate texts>.
Appendix
Table 7: AA prompts adopted from Huang et al. (2024) .