Researchers have long focused on the executive capabilities of AI, such as writing code, but a new study from the OpenMOSS team shifts focus toward "scientific taste." This term refers to the ability to distinguish which research ideas possess the highest potential for long-term impact. By introducing a framework called Reinforcement Learning from Community Feedback (RLCF), the team has demonstrated that AI can learn to evaluate and generate high-quality scientific hypotheses.

The training process involved a two-stage approach centered on preference modeling—the practice of teaching machines to understand human priorities. First, the team developed "Scientific Judge" by training it on 700,000 matched pairs of papers where one had significantly higher citations than the other. This allowed the model to act as a reward mechanism, identifying the traits that characterize groundbreaking science. Next, they used this judge to train "Scientific Thinker," a model designed to propose novel research directions that align with peer-review standards.

The results are striking, with specialized 30B-parameter models outperforming commercial systems like GPT-5.2 and Gemini 3 Pro on research benchmarks. This development marks a milestone toward creating "AI Scientists" capable of not just processing data, but actively steering the direction of discovery. By bridging the gap between raw power and qualitative judgment, this research suggests that the human trait of intuition is becoming a programmable feature.

Researchers have long focused on the executive capabilities of AI, such as writing code, but a new study from the OpenMOSS team shifts focus toward "scientific taste." This term refers to the ability to distinguish which research ideas possess the highest potential for long-term impact. By introducing a framework called Reinforcement Learning from Community Feedback (RLCF), the team has demonstrated that AI can learn to evaluate and generate high-quality scientific hypotheses.

The training process involved a two-stage approach centered on preference modeling—the practice of teaching machines to understand human priorities. First, the team developed "Scientific Judge" by training it on 700,000 matched pairs of papers where one had significantly higher citations than the other. This allowed the model to act as a reward mechanism, identifying the traits that characterize groundbreaking science. Next, they used this judge to train "Scientific Thinker," a model designed to propose novel research directions that align with peer-review standards.

The results are striking, with specialized 30B-parameter models outperforming commercial systems like GPT-5.2 and Gemini 3 Pro on research benchmarks. This development marks a milestone toward creating "AI Scientists" capable of not just processing data, but actively steering the direction of discovery. By bridging the gap between raw power and qualitative judgment, this research suggests that the human trait of intuition is becoming a programmable feature.