Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
AI Research
arXiv CS.AI·27d ago·5 min read
ATTRICITE: Training an Open 4B Model for Citation Recovery toward Faithful Attribution
TL;DR
ATTRICITE is an open 4 billion parameter model designed for citation recovery, achieving a target-match accuracy of 59.8%. It outperforms larger models in this task, supporting efforts for faithful attribution in scientific literature.
✦ Why It Matters
Explore ATTRICITE to enhance citation recovery in your research projects today.
Key Takeaways
- Utilize ATTRICITE for improved citation recovery in your work.
- Leverage the CITEALIGN dataset for training models on citation tasks.
- Consider fine-tuning techniques like GRPO for better model performance.
- Stay updated on citation attribution methods to enhance research credibility.
Full Summary
Faithful citation attribution is crucial for identifying the original sources of scientific claims. ATTRICITE, a 4 billion parameter model, was developed to enhance citation recovery, which involves retrieving the original paper from a citation-bearing passage.
The model was trained within the CiteGuard retrieval environment using the CITEALIGN dataset, which includes 7,607 instances from recent scientific literature. A benchmark subset of 709 instances was created for controlled evaluation, revealing that fine-tuning with GRPO improved the model's target-match accuracy from 49.4% to 59.8%.
Despite its smaller size, ATTRICITE outperformed the larger gpt-oss-20b model and approached the performance of GPT-5.4-mini. The best performance was achieved by Gemma 4 31B IT at 72.0%.
The model and its collection pipeline are released to facilitate reproducible research in citation recovery.