For people building with AI
Make better product bets with new AI research
Your engineers shouldn’t spend their week reading papers. We read the new ones with your product in mind and, when one matters, explain it in plain words: what it is, whether it fits, and whether it’s worth your team’s time. Like having a research department, without the headcount.
We read the new AI papers with your product in mind and tell you, in plain words, which ones are worth your team’s time.
No card needed.
Your day on one screen
Example
Corrective Retrieval Augmented Generation
Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, Zhen-Hua Ling · 2024 · arXiv 2401.15884
Problem
When retrieval returns the wrong documents, the model still answers from them.
Main idea
A small evaluator grades what was retrieved, then keeps it, replaces it with a web search, or combines both.
Key result
Improved both standard RAG and Self-RAG on all four test datasets.
Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection
Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, Hannaneh Hajishirzi · 2023 · arXiv 2310.11511
Problem
Always retrieving a fixed number of passages adds noise, and answers aren’t always backed by their sources.
Main idea
Train the model to decide when to retrieve, and to critique its own answers with special reflection tokens.
Key result
Beat ChatGPT and retrieval-augmented Llama 2 on several tasks, with better citation accuracy.
Lost in the Middle: How Language Models Use Long Contexts
Nelson F. Liu, Kevin Lin, John Hewitt, et al. · 2023 · arXiv 2307.03172
Problem
Models accept long inputs, but it’s unclear how well they use information in the middle.
Main idea
Move the document that holds the answer to different positions in the input, and measure accuracy.
Key result
Accuracy is highest when the answer is at the start or the end, and drops sharply in the middle.
RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval
Parth Sarthi, Salman Abdullah, Aditi Tuli, et al. · 2024 · arXiv 2401.18059
Problem
Retrieving short chunks misses the big picture of a long document.
Main idea
Cluster and summarise the chunks again and again into a tree, then retrieve from any level of it.
Key result
With GPT-4, it raised the best reported accuracy on the QuALITY benchmark by 20 points.
Seven Failure Points When Engineering a Retrieval Augmented Generation System
Scott Barnett, Stefanus Kurniawan, Srikanth Thudumu, et al. · 2024 · arXiv 2401.05856
Problem
RAG systems are hard to get right, and many failures only show up once they’re in use.
Main idea
An experience report from three real projects that names seven places where RAG systems fail.
Key result
Seven failure points to check for, and the lesson that a RAG system is only fully tested in use.
RAGAS: Automated Evaluation of Retrieval Augmented Generation
Shahul Es, Jithin James, Luis Espinosa-Anke, Steven Schockaert · 2023 · arXiv 2309.15217
Problem
Checking a RAG system’s answers usually needs hand-written reference answers.
Main idea
Score faithfulness, answer relevance and context relevance with an LLM, with no reference answers needed.
Key result
Its faithfulness and answer-relevance scores agreed closely with human judgements.
The app’s card adds the data, the limits and the source passage for each fact.
Example project: a support assistant that answers customer questions from product docs. Hover or tap a paper to see its card.
Real, well-known papers, shown as an example. Your map is built every morning from that day’s new papers.
“iReading Room keeps Mitlesen on the edge of language-learning research: in our first week, three of its recommended papers became real improvements to our German stories.”
Someone may have solved your problem already. You just haven’t seen the paper.
Hundreds of new AI papers come out every day. Nobody can read them all, so most people skim social media and hope they didn’t miss the important one.
Usually the paper that would have saved you weeks turns up months later. Sometimes it never does.
How it goes today
- Check arXiv
- Open 30 tabs
- Skim the abstracts
- Save a few, forget most
- Do it again tomorrow
With iReading Room
- Open one map
- Read the cards that matter
- Save them with your notes
- Search your notes later
From a goal to a merged pull request, iReading Room keeps your product improving.
Tell it what you want to improve
Describe your product and the problem you want solved, in your own words, the way you’d brief a new colleague. You only do this once.
It reads the new research for you
Every new paper in your area is checked against your goal, not against keywords. Most stop below your line.
See a ranked dayDecide in a minute, or let your agent decide
Each match comes with a breakdown in plain words: the problem, the idea, the result. Save the good ones to your Library, or let your coding agent pull them itself over MCP.
Read an example breakdownYour agent opens a pull request
Claude Code or Cursor builds the change. Your team reviews it like any other, and iReading Room goes back to reading.
Connect your agent
02 – 04 repeat
Integrations. Bring the papers that matter into the tools your team already works in.
iReading Room · Integrations
Claude Code
Coding agent · MCP
Pulls the day's matches and checks them against your code.
Cursor
Coding agent · MCP
Asks which of today's papers fit the code you have open.
OpenAI Codex
Coding agent · MCP
Reads the day's matches from your terminal, with your own token.
GitHub Copilot
Agent mode in VS Code · MCP
Brings the day's matches into Copilot chat in VS Code, if your organisation allows MCP.
Gemini CLI
Coding agent · MCP
Pulls the day's matches into Google's command-line agent.
Cline
Coding agent in VS Code · MCP
Checks the day's matches against your project inside VS Code.
OpenCode
Coding agent · MCP
Pulls the day's matches into the open-source terminal agent.
Gemini Notebook
Notebook · Chrome extension
Papers you move to your Library land in a notebook in your own Google account. Formerly NotebookLM.
Other MCP clients
MCP over HTTP · token header
One read-only tool, get_matches, behind a personal token you can revoke. Needs a client that can add a token header.
Your tool next
More on the way
Not on the wall yet? Write to us below and tell us where your team works.
Using something else? Tell us where your team works: support@ireadingroom.com
Who it’s for
For people who build with AI, and the people who lead them
Engineers building with LLMs
You ship RAG, agents or fine-tuned models. Hear about the paper that fixes your problem the week it comes out, not three months later.
Team leads and CTOs
Know what changed in your area this week without asking anyone to read arXiv. Your team can get a weekly briefing.
Researchers
Follow your exact question, not your whole field. Read, highlight and take notes on PDFs in one place.
For teams
A weekly briefing on what your team is building
Tell us what you’re building. Every week your team gets a map like the one above, with a short note on why each paper matters.
The first one is free. If it isn’t useful, you’ve lost five minutes.
Get a free sample briefing
Email us two or three sentences about what your team is building.
Email us for a free sampleor write to support@ireadingroom.com
We reply within 2 working days. We use your email and description only to prepare your sample and reply. Privacy
Private by default
Your project descriptions and context are private.
Your notes are yours
Export your saved papers, highlights and notes whenever you want.
Leave any time
Delete your account from Settings, and your data goes with it.
Stop scanning arXiv.
Read the few papers that matter for what you’re building.
No card needed.