Guillaume Vimont

Francis Kulumba, Wissam Antoun, Guillaume Vimont et al.

Authorship attribution in natural language processing traditionally struggles to distinguish genuine stylistic signals from topical confounds. While contrastive learning approaches have addressed this by maximizing semantic overlap between positive pairs, creating large-scale datasets under strict topic constraints remains challenging. We introduce HALvest, a 17-billion-token multilingual corpus harvested from 778k open-access academic papers, and HALvest-Contrastive, a derived dataset designed to isolate stylometric signals through controlled topic variation. Unlike prior work that minimizes lexical overlap, we exploit natural topic drift between papers by the same author, treating residual lexical patterns as authorial fingerprints rather than noise. Comparing lexical baselines (BM25) against neural models trained on unrestricted (topic-rich) versus base (topic-decoupled) triplets, we demonstrate that models trained exclusively on topic-decoupled data achieve superior performance across all test conditions, outperforming both retrieval baselines and models exposed to topic-rich training data. Our analysis reveals that while lexical signals provide substantial performance gains for keyword-driven methods, neural architectures learn robust stylometric representations that plateau with moderate context length, suggesting they capture distributional style beyond surface-level tokens. Both datasets and code are publicly available.

Francis Kulumba, Guillaume Vimont, Laurent Romary et al.

Authorship attribution models fine-tuned with the same pretrained encoder, data, and loss can differ four-fold in performance depending only on their scoring mechanism. We use mechanistic interpretability tools to explain this gap. Stylistic features such as word length, punctuation density, and function-word frequency are equally available at every layer in every model, including in an off-the-shelf control encoder, hence the gap not coming from representation quality. Instead, causal intervention shows that the scorer determines where the encoder consolidates authorship signal. Mean pooling forces consolidation by early to mid layers, while late interaction defers it to later layers. We further derive this difference from the gradient structure of each scorer, and training dynamics reveal distinct learning trajectories that follow from that difference.