I have a benchmark that tests models for "here's a set of statements, express the common principle behind them" and the model expresses a principle.

I have a benchmark that tests models for "here's a set of statements, express the common principle behind them" and the model expresses a principle. Then in fresh context windows, the model is shown a series of statements and asks "which of these statements are examples of the principle?" jev skipped the first step (since it doesn't express) but came out on top finding which statements exhibit the principle Demo: https://yuwakisa.com/benchmarks/principle/ Results: https://yuwakisa.com/benchmarks/principle/report-s5xu-20260919/report