Cuebench

Frontier AI models still fail at specialized tasks like scientific reasoning, experiment design, and inference engineering, reportedly failing around 80% of the time on the hardest environments. Teams working on advancing these models have relied heavily on pre-training, but the page states this approach is now delivering diminishing returns. This leaves a gap in how models learn to handle highly specialized, domain-specific problems where current capability is weak.

Cuebench builds RL environments designed specifically around tasks where frontier models struggle, aiming to train models to improve in those areas. One example shown is a scientific reasoning task involving a hidden double pendulum system, where an agent must choose when to measure the pendulum head's position under a limited measurement budget and then reconstruct the full motion path from those measurements. The company also maintains leaderboards related to these environments and offers a bookable demo, suggesting a sales motion aimed at teams or labs training models.

The site argues that pre-training now delivers diminishing returns at the current stage of AI advancement, and that further progress depends on training models in specialized, hard environments instead.