Reproduce “ATTRICITE: Training an Open 4B Model for Citation Recovery toward Faithful Attribution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yee Man Choi, Xuehang Guo, Songcheng Cai, Yimu Wang, Yi R. Fung, Qingyun Wang
ATTRICITE trains an open 4B-parameter language model to recover the specific scholarly source masked from a citation-bearing passage through tool-based literature search. On a temporally separated computer-science benchmark, reinforcement learning raised target-citation accuracy from 49.39% to 59.75%, with most of the gain coming from improved retrieval of the correct paper; performance also improved on a biomedical transfer set.
Key insights
- Task-specific reinforcement learning primarily improves search behavior: retrieval recall increases from 58.64% for base Qwen3-4B to 68.12% for ATTRICITE, while selection accuracy among retrieved candidates rises more modestly from 84.22% to 87.73%.
- A compact 4B model can outperform the larger gpt-oss-20b on this task, although it remains below stronger systems and struggles with closely related papers or version ambiguity.
- Citation recovery should not be treated as citation verification: matching the published reference identifies intended attribution but does not show that the source supports the associated claim.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.