Bayesian optimization efficiently finds strong experts in LLMs

Researchers have developed a new method using Bayesian optimization to efficiently identify strong single experts within large language models, a process known as gradient-free post-training. This approach, which applies Bayesian optimization within a random linear embedding of weight space and uses a Gaussian process surrogate, requires no backpropagation. Experiments on reasoning benchmarks with Qwen2.5-Instruct models demonstrated that this method achieves comparable or superior results to RandOpt with significantly fewer candidate evaluations, reducing the cost of post-training while yielding stronger deployable models. AI

IMPACT Reduces the computational cost of fine-tuning LLMs, potentially accelerating the deployment of specialized models.

RANK_REASON The cluster contains an academic paper detailing a new method for optimizing large language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Bayesian optimization
• Gaussian process
• Hugging Face
• Neural Thickets
• Nigel Bastian Cendra
• Qwen2.5-Instruct
• RANDOPT: Optimisation par algorithmes stochastiques

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →