AI4Science Hackathon @ Merantix AI Berlin

Build an autonomous AI agent that can do science — without knowing in advance what science it will be doing.

This is an autoresearch hackathon. Each team designs a system — agent scaffold, prompting strategy, tool-use pipeline, or any combination — capable of autonomously tackling data-driven scientific problems. You won't be solving the problems yourself. Your agent will.

How it works

We provide a set of development problems: real scientific tasks with known solutions, drawn from experimental and observational datasets. Use them however you like: study the task format, run your agent against them, evaluate its outputs, iterate on your design.

Final submissions are scored on a set of hidden test problems no team has seen. Same format, different tasks. Your agent runs fully autonomously during evaluation, no prompts, no nudges, no human in the loop.

What we're testing

Can your agent generalize across scientific domains? Can it load unfamiliar data, form an approach, execute the analysis, and produce meaningful results entirely on its own? That's autoresearch and that's what we want to see.

Problem categories

Tasks are drawn from three domains:

Materials
Science of AI / ML
Bio

Teams can specialize their agent for a single domain or enter the General Science super-category, where the same agent is run against test problems from all three domains at once.