MOBO-Merge: Multi-Objective Bayesian Optimization for Model Merging
The MOBO-Merge framework employs multi-objective Bayesian optimization to determine merge parameters for the direct merging of trained AI models within weight space. This method presents a computationally efficient substitute for further fine-tuning. It conceptualizes the selection of merge parameters as a black-box multi-objective optimization challenge, striving to approximate the Pareto front while adhering to a restricted evaluation budget. The framework is agnostic to merge operators and was tested on the Qwen3-4B and Llama-3.1-8B models across two-model instruction-math and three-model instruction-math-code scenarios, utilizing Linear, SLERP, TIES, and block-wise merge operators. MOBO-Merge outperformed random search in 11 out of 12 benchmark comparisons, particularly excelling in multi-dimensional cases. The research is accessible on arXiv under identifier 2608.14264.
Key facts
- MOBO-Merge is a framework for model merging using multi-objective Bayesian optimization.
- It formulates merge-parameter selection as a black-box multi-objective optimization problem.
- The framework is merge-operator agnostic.
- Evaluated on Qwen3-4B and Llama-3.1-8B models.
- Tested in two-model instruction-math and three-model instruction-math-code settings.
- Used Linear, SLERP, TIES, and block-wise merge operators.
- Achieved higher mean hypervolume than random search in 11 of 12 comparisons.
- Gains are small for one-dimensional Linear interpolation but substantially large in other cases.
- Paper available on arXiv with identifier 2608.14264.
Entities
Institutions
- arXiv