• Dataset was taken from: arXiv summarization dataset https://huggingface.co/datasets/ccdv/arxiv-summarization (abstracts + full papers)
• Pre-trained LLM model i.e. Mistral 7B with LoRA fine-tuning
• Training Goal was to summarize research papers accurately using minimal trainable parameters
• Evaluation was done using ROUGE, BLEU, BERTScore, and LLM-as-a-Judge for qualitative evaluation

The system consists of five specialized agents orchestrated using LangGraph:

Enhances the user’s input by generating expanded and related keywords using an LLM, improving the accuracy and coverage of search queries.

Interfaces with academic search APIs (e.g., arXiv, Semantic Scholar, PubMed) to retrieve relevant papers based on the expanded keywords.

Scores and ranks papers using a multi-criteria strategy including:

• Citation count
• Publication date
• Relevance to keywords (inferred through an external LLM API such as Together.ai)

Processes selected top-ranked papers and generates structured summaries using a LoRA fine-tuned language model which you already trained for academic summarization.