ScholarCopilot – An AI Academic Writing Assistant Jointly Launched by the University of Waterloo and Carnegie Mellon University

What is ScholarCopilot?

ScholarCopilot is an AI tool specifically designed for academic writing, developed by a research team from the University of Waterloo in Canada and Carnegie Mellon University. Based on the Qwen-2.5-7B model, it employs dynamic retrieval of citations and joint optimization for generation and citation integration, enabling the precise creation of academic texts with accurate references. During the text generation process, ScholarCopilot inserts special retrieval markers, which query the citation database. The retrieved citation content is then integrated into the subsequent generation, improving the accuracy of citations and the coherence of the text.

The main functions of ScholarCopilot

• Context-aware Continuation: Predict the next three sentences based on the existing content to ensure logical coherence, such as automatically expanding the literature review section.
• Chapter Auto-generation: Input keywords, and AI generates a complete chapter framework, supporting adjustments to academic styles, such as empirical analysis or theoretical derivation.
• Multi-language Support: Enables Chinese-English hybrid writing, suitable for submissions to international journals.
• Dynamic Retrieval Enhancement: Insert markers during writing, and AI retrieves relevant literature from a database of 500,000 arXiv papers in real time, with an accuracy rate exceeding 40%.
• One-click Citation Insertion: Supports multiple formats such as APA/MLA, automatically generates BibTeX entries, and saves time on organizing references.
• Source Verification Feature: Click on a citation to directly jump to the original article, ensuring every reference is authentic and verifiable.
• Doctoral Team Training Data: Fine-tuned on professional academic corpora based on the Qwen-2.5-7B model, with an academic rigor score of 2.87/5 for generated text, far surpassing similar tools.
• Error Self-check System: Automatically flags suspected “hallucinated content” and prompts users for manual review, such as contradictory data or unverified conclusions.

The Technical Principles of ScholarCopilot

• Dynamic Retrieval Marker: During the text generation process, ScholarCopilot dynamically determines when to cite a reference and generates a special retrieval marker. This marker triggers the model to pause text generation and retrieve relevant literature from academic databases in real time.
• Joint Optimization of Generation and Retrieval: The retrieved literature content (such as abstracts or key paragraphs) is directly integrated into the subsequent text generation steps. In this way, the model can generate high-quality academic texts, ensuring the accuracy and relevance of citations.
• Contrastive Learning Optimization: The representations of retrieval tokens are optimized through contrastive learning, enabling the model to efficiently perform similarity searches and further improve retrieval accuracy.
• Citation Accuracy Improvement: ScholarCopilot achieves a top-1 retrieval accuracy of 40.1%, significantly outperforming traditional methods such as E5-Mistral-7B-Instruct (15.0%) and BM25 (9.8%).
• Generation Quality Enhancement: On a dataset of 1,000 academic writing samples, ScholarCopilot achieves a composite score of 16.2/25 across five dimensions—relevance, coherence, academic rigor, completeness, and creativity—surpassing models with larger parameters.
• Training and Data: ScholarCopilot is based on the Qwen-2.5-7B model and trained on a dataset comprising 500K papers from arXiv. By jointly optimizing text generation and citation retrieval tasks, the model achieves significant improvements in both efficiency and accuracy.

The project address of ScholarCopilot

• Project Website: https://tiger-ai-lab.github.io/ScholarCopilot/
• GitHub Repository: https://github.com/TIGER-AI-Lab/ScholarCopilot
• Hugging Face Model Hub: https://huggingface.co/TIGER-Lab/ScholarCopilot-v1
• arXiv Technical Paper: https://arxiv.org/pdf/2504.00824

Application scenarios of ScholarCopilot

• Academic Paper Writing: ScholarCopilot is designed specifically for academic writing, significantly enhancing the efficiency and quality of paper drafting. Through its dynamic “generate-and-retrieve” mechanism, it can determine in real-time when citations are needed during text generation and automatically retrieve relevant literature.
• Introduction and Related Work Sections: ScholarCopilot performs exceptionally well in drafting the introduction and related work sections of a paper. It can automatically predict the next few sentences and provide precise citation suggestions based on the context.
• Academic Writing Teaching and Training: ScholarCopilot can be used for academic writing teaching and training. It helps students and novice researchers master the skills and norms of academic writing and quickly get started on writing high-quality academic papers.
• Scientific Research Team Collaboration: For scientific research teams, ScholarCopilot can share the disciplinary knowledge base, helping team members quickly build paper frameworks. Especially for newly joined members, it enables them to quickly get started with writing literature reviews, thereby enhancing the overall writing efficiency of the team.
• Journal Review: The traceability verification function provided by ScholarCopilot enables journal reviewers to verify the authenticity of references with just one click.