Whenever you're ready, walk me through how you'd approach this.
How do you design evaluation benchmarks to ensure a model’s reasoning capabilities are actually improving?
