SimulacraBench 2026

SimulacraBench is an online competition in the NeurIPS 2026 track that tests how well AI can fill in missing answers in development surveys. It is organised by a team of sixteen people from Stanford, UNHCR, UNICEF, the World Bank and the UN Innovation Network, with prizes sponsored by Schmidt Sciences. The total prize pool is US$16,500 (about ₹15.7 lakh), with a US$10,000 grand prize (about ₹9.5 lakh). The final phase ends on 19 November 2026 (Anywhere on Earth), which is 20 November in India.

It is open to teams from academia, industry and independent research anywhere, except where sanctions or law prevent it, so teams from India can enter. It is a technical machine learning contest rather than a startup grant, and the source does not mention an entry fee.

What the program offers

The grand prize is US$10,000 (about ₹9.5 lakh) for the top team on the overall leaderboard, which is the plain average of the scores on the three surveys. There are also three track prizes of US$1,500 each, one for the best score on each survey: the UNHCR survey of Syrian refugees, the UNICEF Faith and Immunisation survey, and the World Bank Skills Assessment. A Roots of the Data Prize of US$2,000 goes to the best team whose members are all affiliated with one of the 14 countries where the survey data was collected. Those countries are Egypt, Iraq, Jordan, Lebanon, Brazil, Guatemala, Kenya, Mali, Nepal, Nigeria, Pakistan, Zambia, South Africa and Lesotho, so India is not one of them and Indian teams cannot win this particular prize.

Winning teams present their work at NeurIPS and get to discuss their approach with the teams behind the datasets.

Who can apply

Academic, industry and independent research teams can enter, as long as sanctions or applicable law do not stop them. A person can be on only one team. For an applicant in India this means you can compete for the grand prize and the three track prizes. It suits machine learning engineers and data scientists, including startup teams working on survey modelling, synthetic respondents or social-impact AI. It does not suit anyone who is looking for funding for a business without doing the technical task.

How to apply

Create an account on Codabench and register on the SimulacraBench competition page. Download the starter kit, which is released under an MIT licence and contains three survey schemas, a generator that makes invented practice data, the grader, two baselines and a tutorial notebook. The real survey answers are never released to you.

Your entry is code, not a file of predictions. You submit a zip containing a main file that defines a predict function, plus a requirements file. The function returns, for every blank survey answer, a probability for each possible option. The organisers run it in a sandbox with no network on a single H100 GPU. During the development phase, which runs until 13 November 2026, you may submit three times per team per day. In the final phase you get one successful submission. Every team also has to send a 4-page method write-up through OpenReview, and source code has to be provided under a licence that allows non-commercial research reproduction. A practical tip: test with the practice-data generator long before the final phase, because the final run has a fixed time budget.

How applications are judged and what happens next

Winners are decided by the exact, unnoised score from the final phase, measured on respondents the development phase never used. Scores reward honest probabilities rather than just picking the most common answer, and are compared with a baseline that scores zero and a perfect prediction that scores one. Trying to recover or re-identify the underlying records, or helping someone else try, leads to disqualification.

About the organisers

The lead organisers sit at Stanford HAI, with partners from UNHCR, UNICEF and the World Bank who supplied the data, and the UN Behavioural Science Group involved. Schmidt Sciences sponsors the prizes and Modal provides computing.

Things to check before you apply

• The source does not say whether there is an entry fee, so confirm on the Codabench page before you start.
• The source does not mention travel support for presenting at NeurIPS.
• The final time budget is limited, and your code may not read files, use the network or import anything outside your requirements file.
• Development scores are noised and shown on only part of the data, so do not over-tune to the leaderboard.