Preference Model raises $16M and opens its RL environment framework

A weakness in a benchmark can distort a score; the same weakness in training can repeatedly reward the wrong behaviour. That distinction is central to Preference Model’s research into more robust learning environments.

The company says Karotte has been exercised through more than a million evaluation runs and controlled red-teaming. Releasing the framework gives other teams access to those environment-building tools while the company continues its research.