What's this about?
Evolution strategies give broader reasoning coverage than GRPO — 16 upvotes. GRPO is the reinforcement-learning method behind much of the recent progress in model reasoning.
What is an evolution strategy?
Optimisation without gradients. Instead of computing which direction to nudge the weights, you perturb them randomly in many directions, keep whatever scored better, and repeat. It is inefficient per step and embarrassingly parallel, which used to make it uncompetitive. The claim here is about *coverage*: gradient methods follow one path and can converge on a narrow band of reasoning behaviours, while random perturbation explores more broadly and finds strategies gradient descent walks past.
Why does this matter to you?
- "Broader coverage" rather than "higher score" is the interesting claim — mode collapse in reasoning is a real and under-measured failure.
- Evolution strategies parallelise almost perfectly, which matters as inference compute becomes cheaper relative to coordination.
- Another entry in this week's pattern of re-examining whether the fashionable method is actually the best one.