What's this about?
The authors report a strong science/engineering assistant across 16 benchmarks and 769 participant task records. About one-third of assisted tasks were rated infeasible without AI; participant ratings are not an objective counterfactual.
What exactly is a research-assistant benchmark?
A set of tasks meant to probe hypothesis work, analysis or engineering, not only trivia. Results depend heavily on task selection and whether human help was allowed.
What's the community saying?
389 Hugging Face daily-paper upvotes.
- Research assistance may extend beyond drafting papers into useful technical work.
- Independent replication and blinded human comparisons are needed.
- Human final decisions still matter even when AI speeds subtasks.