SCF-RKL introduces sparse, distribution-aware model merging using reverse KL divergence to minimize interference. It selectively fuses complementary parameters, preserving stable representations and integrating new capabilities. Evaluations on 24 benchmarks show superior performance in reasoning, instruction following, and safety.
Key Points
- 1.Controls functional interference via sparse updates
- 2.Outperforms parameter arithmetic methods
- 3.Strong results across model scales and tasks
Impact Analysis
Reduces retraining costs for combining specialized LLMs. Enhances generalization and generation stability. Broad applicability to reasoning and instruction-tuned models.
Technical Details
Employs reverse Kullback-Leibler for divergence measurement. Mode-seeking sparsity preserves core functions. Tested on diverse architectures and benchmarks.