NeurIPS 2026 Repro Hackathon Registration
Register for the NeurIPS 2026 repro hackathon
None defined yet.
Pick a NeurIPS 2026 paper and reproduce its main result with ML Intern. Share your code, logs and verdict so anyone can check your work. Then, if you like, build something on top of it in one of the optional tracks.
For each paper, we pick its central table or figure. That is the paper's claim, for example “our model gets 94.1% accuracy (Table 2)”. Your job is to reproduce it. Every participant does this:
Your verdict for each claim is one of: verified (you got the same result), falsified (the claim does not hold), or inconclusive (you could not tell).
Tracks are optional extras on top of the core challenge. Each one asks for something creative or original, built on a paper you reproduced. Each track has its own award.
Build an RL environment from a paper that lets AI agents practise reproducing its results, or be evaluated on it. Format: Harbor or OpenEnv. Details TBD
Build a Gradio workflow app that compares the paper with earlier work side by side, or connects it to another related model to make something new.
Build a visualization or interactive Space that explains the paper to a general audience.
Can you enter more than one track? Yes.
Click “Request to join this org” at the top of this page. Then fill in the registration form. Registration closes on Nov 19.
Run the experiments. Keep your code, outputs and logs. Optional: build a track entry on top.
Put your work in a public Space and send it through the submission form TBD. You can edit and resubmit until the deadline.
The goal of the hackathon is to do real AI research on Hugging Face infrastructure: your agent runs the experiments, keeps track of them, and publishes what it makes, all on the Hub.
hf CLI and its agent Skill so your agent knows how to work with everything below. Run hf skills add --global to add the Skill for Claude Code, Codex, Cursor and other agents. Set up the CLI for agentsKeep checkpoints, logs and other intermediate files while you work. Buckets docs
Publish the models you train and the datasets you build, each with a card that explains it. Repositories docs
Log the metrics of every run and compare runs in a dashboard. Trackio on GitHub
Call open models through one API, for example as baselines, judges or to generate data. Inference Providers docs
Show your results in an app or a page that anyone can open. Your submission goes in a public Space too. Browse Spaces
We are preparing a list of NeurIPS 2026 papers that have released code or model weights. For each paper, we pick its central table or figure as the claim, and our team checks every claim before the hackathon starts. Each claim has an ID (for example neurips26-4821) and points to a table or figure in the paper.
For each paper, your Space must contain:
| Item | What it means |
|---|---|
| Code | One script that runs the experiment from start to finish. |
| Environment | The exact software versions you used (for example requirements.txt, uv.lock or a Dockerfile). |
| Results | The number the paper reports and the number you got. |
| Logs | The full output of each run, including runs that failed. |
| Verdict | Verified, falsified or inconclusive, with a short reason. |
| Agent trace | The log of your AI agent session, saved to a Hub dataset. ML Intern will have a “Publish Trace” button for this Coming soon |
Then fill in the submission form. It is a Gradio app that checks your Space, shows a checklist of what is complete and what is missing, and marks your submission as complete. AI agents can submit through the same form, because it is also an MCP server. The form also asks which agent and model you used, how much you helped the agent, and whether you looked at other people's reproductions.
Full judging criteria: TBD
Careful, complete and easy to re-run.
The most useful environment for agents to practise or be tested on.
The most useful or interesting Gradio workflow app built on a paper.
The clearest explanation of a paper for a general audience.
A claim that does not hold, with experiments that show the actual result and a clear explanation of the problem.
A pull request that adds a paper's model to transformers, diffusers, TRL or another Hugging Face library.
Anything else useful and open that comes out of your reproduction.
The entry the community likes most.
TBD
Prize amounts: TBD
$20 for each of the first 1000 registered participants, then $10 for everyone who registers after that. How and when you receive them: TBD
Hugging Face's AI agent for ML work. It can read papers, write code, run jobs and save results to the Hub. Try ML Intern
A Gradio app and MCP server that checks your Space and records your submission. Link: TBD
| Date | What happens |
|---|---|
| Nov 13 (00:00 GMT) | Registration opens |
| Nov 19 | Registration closes |
| Nov 20 | Hackathon starts |
| Dec 5 | Submission deadline |
| TBD | Winners announced |
NeurIPS 2026 takes place December 6–13. Plans to share results there: TBD
We will publish the claim list, all submissions, track entries, agent traces and verdicts as open datasets on the Hub. Anyone will be able to see which NeurIPS 2026 results were reproduced.
No. You can use any agent, for example Claude Code, Codex or Cursor. Say what you used in the submission form. ML Intern is the main focus of the hackathon.
Submit it anyway. A careful “falsified” or “inconclusive” verdict is a useful result.
No. Tracks are optional. Everyone does the core challenge; a track is an extra chance to win an award.
We want careful work on a few papers, not quick work on many.
Yes, as many times as you like until the deadline.
No. Registration closes the day before the start.
All submissions are public and open-sourced, and may be used to build open datasets and benchmarks.
Questions: Discord · @Gradio on X
By taking part you agree to the hackathon terms and conditions TBD, the Hugging Face Code of Conduct and the Content Policy.