MEGA Hub

Can Bayesian Optimization Efficiently Find a Strong Single Expert in Neural Thickets?

Authors

Do you know Nigel Bastian Cendra?You can claim authorship or link another user.Do you know Abdelhamid Ezzerg?You can claim authorship or link another user.Do you know Fernando Julio Cendra?You can claim authorship or link another user.Do you know Jeremias Knoblauch?You can claim authorship or link another user.Do you know Jakob Zeitler?You can claim authorship or link another user.

Abstract

Gradient-free post-training has emerged as a compelling alternative to gradient-based optimization for large language models (LLMs), but existing approaches remain costly. We ask whether structured search can identify a strong single expert under a modest evaluation budget. Motivated by evidence that useful weight updates lie in low-dimensional subspaces, we apply Bayesian optimization within a random linear embedding of weight space. Our method requires no backpropagation and uses a Gaussian process surrogate to guide candidate evaluations efficiently. Across several reasoning benchmarks with Qwen2.5-Instruct models from 0.5B to 3B parameters, Bayesian optimization using five times less candidate evaluations matches or exceeds RandOpt. These results show that surrogate-guided search can substantially reduce the evaluation cost of gradient-free post-training while producing stronger deployable single experts.

Community

00