When a system tries to match an anonymous tweet to the journalist who wrote it, it compares writing style. Part of what it calls "style" is not personal at all. It is the shared register of the people who write that way: their age cohort, their outlet, their beat, their region. On a platform with a skewed user base, that shared register is strong. The resulting failure is concrete. A style embedding asked to find the author of a casual tweet among hundreds of journalists will tend to rank highest the journalists who belong to the same crowd as the author, and the true author only somewhere among them. This note proposes to measure how much of cross-genre attribution error comes from this "crowd" signal. It then proposes to remove that signal with a retrieval step. For each query, the step pulls in background authors who write like the query author on the query's platform, looks at how those same people write in the other genre, and subtracts that expectation before comparing candidates.
The idea borrows its diagnosis from retrieval-augmented generation, meaning systems that fetch documents from an external index at query time and condition a model on them. It borrows its mechanism from the same place. Retrieval Shift as a Source of Demographic Bias in Medical RAG showed that dense retrievers key on demographic surface cues and fetch the wrong evidence. Answer flip rates rose from below 3% for standalone models to 18.87–34.92% for medical retrieval systems under ethnicity perturbations, and 86–97% of errors traced to a different retrieved set. We hypothesize the stylistic analogue: style retrievers key on cohort register. We then turn retrieval into the remedy by retrieving cohort-matched impostors, meaning other authors who look like the query author but are known not to be the candidate.
Core Claim
On cross-genre attribution, the strongest current style embeddings make errors that fall disproportionately on candidates who share the true author's cohort. A cohort baseline estimated per query can remove a measurable share of those errors. The baseline comes from retrieval over a disjoint background datastore of authors who write in both genres. It needs no demographic labels at inference time, because cohort is defined implicitly by nearest neighbours in the query's genre. Demographic and outlet labels are used only to audit where the errors fall.
Concretely, let $e(\cdot)$ be a fixed style encoder. Let $q$ be a query tweet and $a$ a candidate author's article profile. Let $N_k(q)$ be the $k$ background authors whose tweets lie nearest to $q$. Let $m_{tw}(q)$ and $m_{art}(q)$ be the mean tweet and mean article embeddings of those $k$ authors. We score candidates by
$$ s(q,a) = \cos\big(e(q) - m_{tw}(q),\; e(a) - m_{art}(q)\big). $$
The term $m_{art}(q)$ is the key quantity. It answers "what do articles look like from people who tweet like this?", and it depends on the query, so it changes the ranking over candidates rather than shifting all scores equally. Replacing $N_k(q)$ with all background authors gives global mean-centring. Replacing it with random authors gives a classic impostor baseline. The difference between those variants and the matched version isolates the contribution of cohort matching.
Relation to Prior Work
Three lines of work motivate the gap. CROSSNEWS links journalists' articles to their tweets. It introduces SELMA, style embeddings from an instruction-tuned 7B encoder, and reports a drop from about 85.8% article–article to about 78.5% article–tweet verification accuracy. It treats that drop as a genre penalty and does not ask whose style is being confused with whose. Gender Attribution: Tracing Stylometric Evidence Beyond Topic and Genre shows that group-level signals survive topic control in character patterns (71.3% on topic-balanced blogs) and in syntax (76.0% on scientific papers). A group signal that strong can plausibly contaminate individual-level matching. The platform-skew literature supplies the reason such signals would be concentrated on social platforms: Reddit-derived corpora skew young, male and Western (OpinionGPT, MPCHAT), and Twitter discourse departs from population structure (Tweeting AI). OpinionGPT models such skews explicitly with per-group adapters. We do the reverse. We estimate the group component so that it can be removed from an individual-level decision.
Two lines constrain the design. Bias Analysis and Mitigation in the Evaluation of Authorship Verification shows that corpus-wide statistics computed over test data inflate results. Its naive baseline beat deep models on PAN until it was evaluated on the cleaned Webis-19 corpus, where it fell to about 70%. Our background datastore must therefore be author-disjoint from every test candidate, and centring statistics must never touch test documents. A Ship of Theseus shows that LLM rewriting erases personal style, with BERT macro-F1 falling from 0.77 to 0.33 after one paraphrase. We therefore do not use any generative rewriting to construct test queries.
The mechanism reuses the retrieval-conditioning idea behind Grounding Language Models with External Knowledge Retrieval. The difference is that what we retrieve is contrastive background rather than supporting evidence. Can Large Language Models Identify Authorship? is also relevant. Its forensic prompting (LIP) reaches 84.45% weighted F1 on blogs, but it needs every candidate in the prompt. We use it only as an optional reranker over a short candidate list.
Research Program
The work has three parts, and the first can end the project.
The first part is a diagnosis. On the CROSSNEWS Gold set, we run SELMA and a contrastive encoder in tweet-to-article attribution. We then ask whether wrong top-ranked candidates share the true author's cohort more often than the candidate pool's base rate predicts. Cohort attributes come from four sources: outlet, the benchmark's fine-grained topic labels, and, where Wikidata coverage allows, gender and birth decade taken from the Wikidata links used to build the Silver set. Coverage has to be checked; we do not assume it. We avoid name-based gender inference as the primary label, because Inferring Gender from Names on the Web documents its failures outside Western naming conventions. We also build candidate pools that are cohort-homogeneous and cohort-mixed but matched in size. If the crowd signal matters, accuracy should fall more in homogeneous pools. If neither the error excess nor the pool gap appears after controlling for topic, the hypothesis is false and we stop.
The second part is the correction. We build the background datastore from CROSSNEWS Silver journalists. This set contains roughly 2,260 authors with both tweets and articles, and we remove anyone overlapping the Gold test authors. We then implement the scoring rule above, sweeping $k$ and a symmetric variant that also centres each candidate against its own article-space neighbours. The encoder is frozen throughout, so any gain comes from the retrieved background and not from extra training.
The third part is an optional generative stage that tests whether contrastive retrieved evidence helps an LLM. For the top five candidates, an LLM receives the query, each candidate's articles, and articles from the query's retrieved cohort neighbours, labelled as not the author. It is asked, in the LIP style, to name which features separate the author from the crowd. This stage is kept only if it beats the same prompt without cohort exemplars.
Baselines
The naive reference is SELMA cosine ranking on the frozen encoder, with no normalization, run on exactly the same Gold tweet-to-article queries and candidate pools. It is the strongest existing method on this benchmark. CROSSNEWS reports a tweet-to-article attribution average rank of about 5.6 and article–tweet verification accuracy of about 78.5%. We first reproduce these numbers so that our pools and encoder match the reported setting. We count as beating it a lower average rank and higher top-1 accuracy, with confidence intervals from query bootstrap that exclude zero. The gain must also be larger on cohort-homogeneous pools than on mixed ones, which is the signature the hypothesis predicts.
The second baseline is global centring. It subtracts the mean tweet and mean article embedding of all background authors. It costs the same, uses the same datastore, and contains no cohort matching, so it is the ablation that isolates our key ingredient. The third is the random-impostor variant, with $k$ background authors drawn at random, following the impostor tradition in verification. The fourth is topic-matched impostors, drawn from the same CROSSNEWS topic label as the query. This tests whether "cohort" reduces to topic.
The fifth is a trained debiasing competitor. It is a contrastive encoder fine-tuned on Silver with an adversarial head that penalizes predictability of outlet and topic. It receives the same data and a tuning budget matched in GPU-hours to our $k$ sweep. BEDAA reports up to 19.69% F1 gains within its benchmarks. No reported result on CROSSNEWS was found on Lacuna for BEDAA, so we include a DeBERTa fine-tune as a representative supervised encoder rather than claim its number.
For the optional LLM stage, the baseline is LIP reranking of the same top five candidates without cohort exemplars. The only reported LIP figure available is 84.45% weighted F1 on blogs, which is a different task. No reported result for LIP on CROSSNEWS tweet-to-article attribution was found on Lacuna beyond the paper's statement that zero-shot prompting lagged SELMA.
Why not just…?
Why not just use random impostors or global score normalization, which already exist? This is the most serious threat, and it is why both variants sit in the baselines. If random or global centring recovers most of the gain, the contribution shrinks to "normalization helps SELMA", and we will report it that way. The proposal survives only if matched neighbours beat random and global centring. They must do so specifically on cohort-homogeneous pools, where a single global mean cannot tell apart candidates who already share the crowd.
Why not just control for topic, since CROSSNEWS has topic labels? Journalists' cohorts and beats overlap, so apparent cohort effects may really be topic. Two elements address this. The topic-matched impostor baseline tests it directly. The diagnosis stratifies the same-cohort error excess within topic. The gender-attribution evidence (Gender Attribution) suggests group signal persists after topic control, but only this within-topic stratification will show whether that holds for journalists.
Why not just fine-tune the encoder adversarially to forget cohort? It may work, and it is run at a matched budget. Two considerations favour the retrieval approach, and both are tested rather than assumed. First, adversarial removal needs cohort labels at training time, whereas the retrieval method defines cohort implicitly. Second, the retrieval method leaves the encoder untouched, so a new background datastore adapts it to a new platform without retraining. The comparison therefore also includes transfer to a held-out outlet group.
Why not just prompt a frontier LLM with all candidates? Can Large Language Models Identify Authorship? itself notes that the all-in-one prompt breaks beyond small pools. CROSSNEWS reports that prompting lagged SELMA across genres. We therefore use LLMs only over retrieved short lists.
Proposed Experiments and Decisive Checks
The main evaluation is tweet-to-article and article-to-tweet attribution on CROSSNEWS Gold, which has 500 authors. We report average rank, top-1 and top-10 accuracy, and verification accuracy. Every system runs on the same pools and uses the same Silver-derived, author-disjoint background datastore. Following the Webis-19 lesson, all statistics are computed without access to test documents.
The first decisive check is the diagnosis. It asks whether same-cohort confusions exceed base rate within topic strata, and whether homogeneous pools cost SELMA more than mixed pools. A null result here ends the project.
The second decisive check asks whether cohort-matched centring beats global, random and topic-matched centring, with a larger margin on homogeneous pools. We also measure whether it shrinks the error excess in the diagnosis, not just the aggregate rank. A method that improves rank while leaving same-cohort confusions unchanged would be helping for some other reason.
The third check is fairness. We measure per-cohort attribution accuracy, wherever Wikidata attributes exist, to test whether the correction narrows gaps between groups or merely helps the majority cohort that dominates the background datastore. Because the datastore inherits the platform's skew, it may represent minority cohorts too sparsely to yield good neighbours. We therefore report accuracy as a function of neighbour density.
As a within-platform generality check, we apply the same procedure to the blog authorship corpus used in the LLM-attribution work. There, cohort structure comes from the platform's own user base rather than from a genre boundary.
References
- Retrieval Shift as a Source of Demographic Bias in Medical RAG
- CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark
- Gender Attribution: Tracing Stylometric Evidence Beyond Topic and Genre
- OpinionGPT: Modelling Explicit Biases in Instruction-Tuned LLMs
- MPCHAT: Towards Multimodal Persona-Grounded Conversation
- Tweeting AI: Perceptions of Lay versus Expert Twitterati
- Bias Analysis and Mitigation in the Evaluation of Authorship Verification
- A Ship of Theseus: Curious Cases of Paraphrasing in LLM-Generated Texts
- Can Large Language Models Identify Authorship?
- Inferring Gender from Names on the Web: A Comparative Evaluation of Gender Detection Methods
- BEDAA: Bayesian Enhanced DeBERTa for Uncertainty-Aware Authorship Attribution