CS 224V

Agentic AI

Fall 2026

Project Info

Each project will be in a group of 2 students.

A core part of CS 224V is a project you work on throughout the quarter. We have mentors and advisors across many disciplines who have signed on to help you with these projects, from Stanford and from external partners. All projects are mentored and supervised on a weekly basis.

Projects from CS 224V have produced key research papers that inspired commercial deep research products (OpenAI Deep Research, Google Gemini Deep Research, Databricks Genie Deep Research Mode). The open-source software has been downloaded and used in industry, and pilots based on the technology, hosted at WWKnowledge.org, have been used by over 800K consumers, journalists, and historians. Many projects build on previous years' technology to advance the state of the art for the year after. Will your project show up on that list next year?

The project proposals document is the starting point for choosing a topic. It collects state-of-the-art LLM research projects led by NLP researchers at Stanford, together with applications of LLMs developed with experts in biomedicine, history, journalism, medicine, and sustainability. You are most welcome to propose your own project — please post it in that document so it can attract partners.

Infrastructure for Your Project

This year we are making available AOS (Agentic OS), a new architecture on which you can build your own agents. You are welcome to use any tools you like. However, your project must use state-of-the-art tools and concepts, and cannot be built with vanilla LLMs by simply prompting them with the problem statement.

AOS is an open, accountable platform for collaborative agent systems, with a layered architecture: a semantic file system for persistent knowledge and artifacts, application platforms for long-horizon interaction and accountable decision making, and team workflow support for shared workspaces and collaboration. As an open-source, model-agnostic platform, it makes smaller and open-weight models more effective, and supports data sovereignty by letting organizations deploy, audit, extend, and govern it in their own environments.

Semantic file system

Ingestion

Semantic retrieval

Knowledge curation

Application frameworks

Project Axes

There are two major approaches to defining a project: you can start with an application area, or with a technique. For the latter, you will still need a domain to test your technique on.

A. Domain driven

  1. What is your domain of interest?
  2. What is the challenge in your project? For example: human-computer interaction; precision, recall, accountability; long context; long horizon; or unknown — you may need to build the first prototype to discover the challenge.
  3. Which techniques are you planning to use or to improve?
  4. How are you planning to evaluate your solution?

Take medicine as an example. OpenEvidence's accuracy on diagnosis and treatment in the MedXpertQA benchmark is limited to 44%. How can we improve it? Simply applying RAG to the PubMed literature does not work, because each paper reports the findings of one experiment; doctors rely on guidelines, which experts create by curating many papers on a topic. That opens up a range of problems:

B. Technique driven

You are welcome to improve the components of AOS or add new functionality. Some problems we have identified:

  1. CHURRO: no VLM today can handle Chinese newspapers typeset in the 1800s and early 1900s. Automatically synthesizing training data that teaches VLMs to handle such papers is a particularly attractive approach.
  2. SLIDERS: how do we support sophisticated queries over patient records?
  3. GRILL: how do we improve management of hypotheses and report generation? How do we perform deep analysis across domains, e.g. compiling information on rare diseases or compiling medical guidelines?
  4. DataSTORM and GRILL: how do we integrate the two so that literature search can be combined with data analysis?
  5. Genie Worksheets: can we derive worksheets from English descriptions, or from the accessibility information on websites?
  6. VERDICT: explore accountability in other domains — rare disease diagnosis and guideline adherence in medicine, governance compliance in finance.

Proposals and Deliverables

Mentor-written proposals

If you take a mentor-written proposal, you will submit a full project proposal on Gradescope. This is an extended version of the mentor-written proposal. It can be largely the same as what your mentor provided, but you should edit it if you are narrowing or expanding the scope and customizing it to your interests. It should have more detail on when each phase of the project will be completed, the datasets you will use, a proposed weekly schedule, and what each partner will work on.

In addition to the fields already in the mentor-written proposal, the full project proposal asks you to fill out:

Prior work:
Expected demo at the end of the quarter:
Weekly schedule:

Custom proposals

If you are doing a custom project, you will have to sign up to present it in class. Signing up is mandatory. If you sign up early, you will get more feedback, which you can use to update your final proposal.

Please still review the mentor-written proposals for examples of the level of detail we are looking for as you propose your custom project.

The full project proposal asks you to fill out:

Title:
Team Member(s):
Key Question:
Motivation:
Project description:
Mentor (if known):
Prior work:
Expected demo at the end of the quarter:
Weekly schedule:

Please use the following format: Project Template

Final Project Presentation, Poster, and Paper

At the end of the quarter we host a presentation and poster session for all final projects. Each group makes a 60-second presentation at the beginning, in our usual lecture location, and we move on to the poster session afterward. We adopt the same poster session guidelines as CS 224N. Dates are on the Schedule.

A final paper about your project is due in finals week. You should also submit your code, along with a README explaining how to run your program for a demo, at the same time as the paper. It is recommended to include a short video demo of your project along with your code.

Here is a suggested outline of the final paper:

(1) Abstract: 2-3 paragraphs summarizing the paper, including the results
(2) Introduction, which includes the motivation, main idea, and overall contribution
(3) Related work
(4) The core ideas
(5) Experimental results
(7) What you learned and future work
(8) Conclusion
(9) Appendix: Examples, Prompts that you engineered ...
We recommend using ACL style for your paper: ACL style files

Publications from Past CS 224V Projects

2023

2024

2025

2026