Work in Progress: An Automated Workflow to Summarize and Simplify AI Research with Free Tools

AI
Automation
NLP
LLMs
Open Science

My attempt to build a fully automated pipeline to fetch, simplify, and publish weekly summaries of open-access AI research using OpenAlex, Hugging Face, and GitHub Actions — entirely with free tools.

Published

June 28, 2025

🧠 Why I Built This

We may be curious about the latest discoveries in AI research, but reading papers may be time consuming, and also a little boring — even for someone having an extensive research background. I wanted to:

  • ✅ Make new research easier to understand
  • 🧪 Test the capabilities of free LLMs for science communication
  • 🤖 Automate end-to-end publishing using GitHub Actions

So I built a workflow that fetches one open-access AI paper per week, among the most recent publications, and rewrites its abstract in plain English for curious general readers.

🔧 Tools I Used

This project only uses free resources:

  • OpenAlex – to fetch recent open-access AI papers
  • Hugging Face Inference API – to generate human-friendly explanations (using models like zephyr-7b-beta and mistral-7b-instruct)
  • GitHub Actions – to automate everything weekly
  • Quarto Blog – to publish the summaries as .qmd posts here
  • ChatGPT - to ease the coding and consult about the project

⚙️ The Pipeline

1. Fetch a Paper

Using the OpenAlex API, I query for recent (< 6 months old) papers in artificial intelligence, limited to those with open-access full text and indexed abstracts. I:

  • Decode the abstract_inverted_index
  • Filter out papers already posted (based on DOI or title)
  • Randomly select one paper with a valid abstract

2. Simplify the Abstract

The chosen abstract is either:

  • Used as-is (if short)
  • Or summarized first (if >1500 characters) to prevent exceeding token limits and also to provide a shorter input to the LLM which simplifies it, as longer abstracts tended to confuse the model.

Summarization prompt:

“Summarize the following academic abstract in a concise way, capturing the main points clearly”

Then, a second prompt was used to rewrite it entirely in simple, more engaging engaging language. I tried several prompts, including some suggested by ChatGPT. For instance:

“You are an expert science communicator. Your task is to rewrite the following academic abstract. Rephrase the text in a clear, engaging, and simple way, while staying fully faithful to the original content. Use your own words, avoid jargon, and do not add information that is not explicitly present in the original. Write one paragraph of 5 to 8 sentences, in third person, without using the words ‘abstract’, ‘original’, or meta-commentary. Use a neutral, friendly tone and explain the ideas as if to a curious teenager. Conclude with a sentence explaining why this research matters to society.”

And:

“You are an expert AI communicator. Your goal is to rewrite a complex academic abstract for a tech-savvy but non-expert audience, like a curious student or a journalist. Follow these instructions precisely: (1) Rewrite the abstract into a single, cohesive paragraph between 5 and 8 sentences. (2) Start by explaining the core problem the researchers are trying to solve. (3) Clearly describe their unique method or solution. (4) Mention the key result or finding. (5) The final sentence MUST explain the real-world importance or potential impact of this research. (6) Use simple, engaging language. Avoid jargon or explain it in the simplest terms. (7) Do not add any information not present in the original text. Do not use phrases like ‘This paper…’ or ‘The abstract describes…’.”

3. Generate and Save the Blog Post

The output includes:

  • Paper title and authors
  • Original abstract (cleaned)
  • Rewritten summary
  • A link to the full paper
  • Notes on the model used

The post is saved as a .qmd file in my Quarto blog.

4. Publish Automatically

GitHub Actions runs this pipeline every Friday. It:

  • Fetches and simplifies the paper
  • Commits the new post to the blog repo
  • It auto-deploys the blog

Models & Parameters

I tried the following models:

  • t5-small - mostly stays to close to the original text
  • HuggingFaceH4/zephyr-7b-beta - results vary a lot! 🤔

The current parameters I am using:

Parameter What It Does Value Why We Chose It
temperature Controls randomness. Lower = more focused/deterministic. Higher = more diverse. 0.35 Helps keep output controlled and factual.
top_p Nucleus sampling: only sample from top tokens whose cumulative prob ≤ p. 0.8 Allows slight diversity but avoids extreme randomness.
repetition_penalty Penalizes repetition of phrases or words. 1.2 Encourages varied wording and avoids repeated structures.
do_sample Whether to sample or pick the most likely next token. False Disables random sampling for more consistent results.
use_cache Reuse past attention (for faster decoding). False Disabled in inference API for fresh generation each time.
max_new_tokens Maximum number of tokens the model can generate in the response. 300 Keeps summaries concise (~150–200 words max).
return_full_text Whether to return both prompt + response (True) or just the response. False We only want the model’s generated output, not the prompt.

⚠️ Challenges So Far

I hit over and over these issues:

  • 🥱 Summaries are often too close to the original (not real rewrites)
  • 🤐 Occasionally, I get empty output or skipped generations
  • 👽 I got a case of hallucination

I’m still experimenting with prompt wording, sampling parameters, and maybe switching to more consistent models.

💬 What’s Next — and How You Can Help

I’d love to:

  • Try better models (open to suggestions!)
  • Avoid hallucinations more reliably
  • Add topics beyond AI (NLP, ML, etc.)
  • Get community feedback to improve the summaries

👉 If you’ve tried summarizing scientific research with LLMs — or have tips for better prompting — please reach out!

You can find the pipeline code and one of my (a bit disappointing) attempts here:

Thanks for reading! 🤖✍️