Search papers, labs, and topics across Lattice.
Affiliation:
2
0
4
0
Evaluations across seven tasks and three Qwen scales show higher mean accuracy than RandOpt on Countdown, GSM8K, and MATH-500 at every scale, and perturbation geometry is highlighted as a key design choice for population-efficient, gradient-free search around pretrained models.
Optimizing transformer modules with tailored manifold geometries can significantly enhance performance, revealing that one-size-fits-all approaches may lead to instability.
