SIP-BMM: Constructing the Capability--Efficiency Pareto Set for LLMs via Structural Importance Prior Bayesian Model Merging

Constructing a Pareto set is pivotal for navigating the capability--efficiency trade-offs in Large Language Models (LLMs). However, existing merging techniques remain inadequate for this task. Coarse-grained, model-level methods yield only a sparse set of suboptimal solutions, while fine-grained, layer-wise approaches suffer from the curse of dimensionality, rendering the search space computationally intractable. To resolve this dichotomy, we propose Structural Importance Prior Bayesian Model Merging (SIP-BMM), a framework that automatically constructs the LLM Pareto set. SIP-BMM renders high-dimensional layer-wise search tractable by introducing an importance-aware Sparse Axis-Aligned Subspace Bayesian Optimization (SAASBO) strategy. By leveraging a structural importance prior derived from task-vector differences, our method guides SAASBO to automatically identify critical layers, thereby dramatically reducing the effective dimensionality without sacrificing the granularity of full-model control. The entire process is automated within an evolutionary loop driven by the Log-Noisy Expected Hypervolume Improvement ($q$NEHVI) acquisition function. Experiments demonstrate that SIP-BMM discovers a stronger and denser Pareto front than competitive baselines, enabling agile model selection tailored to diverse operational constraints. Code is available at: https://github.com/MiLab-HITSZ/2026-SIPBMM.