LlMBO
A method that applies Bayesian optimization to LLM fine-tuning, suited to extremely low-dimensional parameter spaces (up to a few dozen parameters). It requires no gradient computation, is sample-efficient, memory-light and naturally parallel, and can directly optimize discrete objectives such as greedy-decoding accuracy.
Timeline (last 90 days)
LlMBO (Large language Model Bayesian Optimization) introduces Bayesian optimization into LLM fine-tuning.
LlMBO has core advantages such as sample efficiency, being gradient-free, lightweight memory, natural parallelism, and directly optimizing the target metric (such as greedy decoding accuracy).
In low-dimensional fine-tuning experiments on a math benchmark suite, LlMBO's fine-tuning gains matched or even exceeded those of GRPO, while using only a small fraction of the latter's compute and memory.
LlMBO is currently still limited to low-dimensional fine-tuning scenarios, and for full-parameter training or large-scale adaptation, gradient methods remain irreplaceable.
All Facts (4)
LlMBO has core advantages such as sample efficiency, being gradient-free, lightweight memory, natural parallelism, and directly optimizing the target metric (such as greedy decoding accuracy).
50%UnverifiedIn low-dimensional fine-tuning experiments on a math benchmark suite, LlMBO's fine-tuning gains matched or even exceeded those of GRPO, while using only a small fraction of the latter's compute and memory.
50%UnverifiedLlMBO is currently still limited to low-dimensional fine-tuning scenarios, and for full-parameter training or large-scale adaptation, gradient methods remain irreplaceable.
50%UnverifiedLlMBO (Large language Model Bayesian Optimization) introduces Bayesian optimization into LLM fine-tuning.
50%