ReCite: Agentic Reasoning for Faithful Citation
Yuyang Huang, Bobo Li, Jiajia Song, Yuzhe Ding, Chong Teng, Fei Li, Donghong Ji
Read on arXiv →
Key claim
ReCite improves citation accuracy through claim-level reasoning.
In plain English
Imagine you're writing a research paper and need to cite relevant studies to support your arguments. With the sheer volume of literature available, it can be overwhelming to find the right papers that not only exist but also logically back up your claims. Currently, many systems rely on finding papers that are semantically similar to your text, but this often leads to misattributions — citing real papers that don’t actually support your argument. This is what's called misattribution, and it can undermine the credibility of your work.
To tackle this issue, a new framework called ReCite has been proposed. Instead of just looking for similar papers, ReCite focuses on understanding the specific claims being made and finding evidence that directly supports them. It does this by employing a method that involves reasoning about the claims and verifying whether the evidence aligns logically. This approach not only improves the accuracy of citations but also introduces a self-correction mechanism that helps refine the results when the initial candidates don’t fit well. Compared to existing generative models, ReCite shows a significant improvement in citation accuracy, making it a more reliable tool for anyone looking to automate academic writing.
The approach shifts from similarity-based citation to claim-level reasoning.
Experiments show improved citation accuracy compared to existing models, though details on baselines could be clearer.
Deep reliability assessment
The methodology supports the claim that ReCite improves citation accuracy by focusing on claim-level reasoning rather than semantic similarity, but it may overclaim its generalizability across different disciplines as it is trained only on computer science papers.
Reproducibility
No open source code or dataset is mentioned in the paper.
Key figure
Figure 1 compares traditional semantic retrieval with the ReCite framework, highlighting the shift from similarity-based search to claim-level reasoning.