What's this about?
Scientific agents increasingly analyse data, run code and produce research artifacts, but most benchmarks score the final answer, an isolated program, or a single domain. FrontierChallenge is 300 end-to-end cross-domain scientific workflows, with 97 released and evaluated in this paper. 133 upvotes.
Why score the workflow instead of the answer?
Because in research the answer is often not the deliverable. What matters is whether the analysis was done correctly, whether the artifacts are reproducible, and whether someone else could build on them. An agent that reaches the right number through an invalid route has produced nothing usable. Scoring the process is much harder to automate, which is why almost nobody does it.
Why does this matter to you?
- It is the second scientific-workflow benchmark in today's digest, alongside Terminal-Bench-Science — a clear convergence.
- Cross-domain and end-to-end are both deliberate choices against the single-task benchmarks that currently drive model development.
- Only 97 of 300 tasks are released, which is a sensible hedge against training-set contamination.