The AI for Meta‑Science workshop has a companion hackathon, organized by Stockholm AI and Sweden AI Factory. After an in-person kick-off in Stockholm, teams spend one week online building working tools for problems the workshop studies: checking research questions, recording how AI-assisted research was done, and linking a paper’s claims to its evidence. The winning team presents its tool at the workshop at NeurIPS 2026 in Paris.
The winning team gets an all-expenses-paid trip to NeurIPS 2026 in Paris, a poster slot, a 5-minute lightning oral presentation at the AI for Meta‑Science workshop on December 13, and a winner’s certificate.
If the kick-off is full, join the waitlist on Luma and you will be notified when a spot opens.
Key Dates (Stockholm time, CET)
| Kick-off (in person) | Tuesday, November 10, 17:30–20:00, at RISE KTH, Drottning Kristinas väg 61, Stockholm. Food and drinks provided. |
| Build week (online) | November 10–17, teams work remotely |
| Submission deadline | Tuesday, November 17, 18:00 |
| Winner announced | Tuesday, November 24 |
| Winning team presents (in person) | Sunday, December 13, at the AI for Meta‑Science workshop, NeurIPS 2026, Paris |
Three Tasks
Each task asks you to build a working tool, and each one addresses a theme from our call for papers. The task brief gives full requirements.
1. Ideation stress-test tool
Build a tool where a researcher pastes in one research question and gets an early check on whether it is worth pursuing.
It must
- Extract and restate the key terms, flagging any that are ambiguous or used inconsistently in the literature
- Surface the implicit assumptions baked into the question
- Identify the evidence or data needed to answer it, and flag if that evidence is unlikely to be obtainable
- Recommend proceed, narrow, revise, or stop, with reasoning
Success criterion
Given a deliberately vague or overly broad question, it gives specific, actionable critique (“your term ‘engagement’ is used in three different senses in the literature”), not generic advice (“consider clarifying your terms”).
Related workshop theme: evaluating research ideas and AI scientists.
2. Provenance-preserving lab notebook
Build a lightweight logging wrapper around one AI-assisted research task, for example an agent cleaning data, running experiments, or iterating on a model.
It must automatically capture
- Each plan or hypothesis proposed, and the reasoning behind it
- What was actually run: code, parameters, data version
- What failed or contradicted expectations, and how the plan changed in response
- A timeline a human can walk back through to see the decision path, not just the final output
Success criterion
A reviewer who wasn’t present can read the log and answer “why did they change approach at step 4?” without asking the researcher.
Related workshop theme: reproducibility and accountability.
3. Claim-to-evidence linker
Build a tool that takes one paper or preprint and, for each major claim, locates and links its exact evidence: a specific figure, table, results paragraph, or code path.
It must flag
- Claims with no traceable supporting evidence in the document
- Claims whose evidence is weaker than the stated confidence (“we show X” backed only by a single small-n experiment)
- What was checked and what remains unverified
Success criterion
The output is a claim map (claim → evidence → confidence/caveat), not a single score, so the reader can inspect each link individually.
Related workshop theme: AI-assisted peer review and verification.
Organizers
The hackathon is run by Stockholm AI and Sweden AI Factory, in partnership with the AI for Meta‑Science workshop. Sweden AI Factory fully sponsors the hackathon and the prize.
Questions about the hackathon: info@stockholm.ai. Questions about the workshop: ai4metascience@googlegroups.com.