Observed arrival · 2026-10-14
Rewind Gym turns research forks into AI test episodes
It describes RL environments that test whether models can make research decisions by replaying real projects from before a choice.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Frontier labs testing AI research judgment
- Worth noticing
- The page reports a 59.7% success rate and cites a 2026 arXiv paper, but provides no benchmark protocol in the extracted text.
Field notes
The homepage says a model sees the full record and next goal while post-decision events are hidden; each researcher choice becomes an episode, with the real outcome serving as grader. Its example is labeled “rare-celltype-fm / trajectory” and includes a “Show what happened” control; access is described as available to frontier labs by arrangement. The stated 59.7% figure is tied to a cited 2026 arXiv paper, but no benchmark protocol appears in the extracted text.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue