<- back

AI Research Paper Processor

A team project for Apart Research's AI Capabilities and Risks Demo-Jam in August 2024. Enter an arXiv ID or let the app select a random paper. The app makes a shorter summary of the abstract. A different model writes an abstract from that summary. The app then compares the new abstract with the original.

We wanted to see what information remains after compression and what the second model invents.

Live demo / Source / Video showcase

I used Python and Streamlit with the Hugging Face API. BART writes the summary, and Mistral 7B writes the new abstract. A small sentence-embedding model compares the two abstracts. You need your own Hugging Face API key to run the app.

I wrote most of the code, working with Severian0 and Smartiespawn, two other members of my LEAF cohort.

Outcome

The result was a hackathon prototype for exploring compression and reconstruction. The score measures text similarity. It does not show whether the new abstract is factually correct.