ATTRICITE: Training an Open 4B Model for Citation Recovery toward Faithful Attribution

Faithful citation attribution begins with identifying the intended source for a scientific claim. We study this source-identification capability through citation recovery: recovering the paper cited by the original author from a citation-bearing passage. Our evaluation adopts the published author's citation as an observable human attribution signal and uses target recovery as a proxy for progress toward faithful attribution. We introduce \modelname{}, an open 4B-parameter model trained for tool-using citation recovery within the CiteGuard retrieval environment, together with \datasetname{}, a 7,607-instance computer-science dataset drawn from recent scientific literature. For controlled evaluation, we construct a 709-instance benchmark subset of \datasetname{}, comprising 410 development instances from 2024 publications and 299 temporally held-out test instances from 2025 publications. Across three runs at an inference temperature of 0.7, GRPO fine-tuning improves Qwen3-4B from $49.4\%\mathbin{\pm}1.5\%$ to $59.8\%\mathbin{\pm}0.2\%$ target-match accuracy, a gain of 10.4 percentage points. Despite using only 4B parameters, \modelname{} outperforms gpt-oss-20b and comes within 3.9 points of GPT-5.4-mini, while Gemma 4 31B IT achieves the strongest overall performance at $72.0\%\mathbin{\pm}1.0\%$. We release the model and collection pipeline (\url{https://anonymous.4open.science/r/AttriCite-C6D5}) to support reproducible research on citation recovery toward faithful attribution in a continually evolving scientific literature.