Leveraging Parameter Space Symmetries for Reasoning Skill Transfer in LLMs
This work addresses the issue of skill transfer across evolving LLM families, reducing redundant fine-tuning and enhancing adaptability, though it is incremental as it builds on existing task arithmetic methods.
The paper tackles the problem of negative interference in task arithmetic for transferring skills between Large Language Models by aligning parameter spaces using symmetries in Transformer architectures, resulting in consistent outperformance over standard task arithmetic on challenging reasoning benchmarks.
Task arithmetic is a powerful technique for transferring skills between Large Language Models (LLMs), but it often suffers from negative interference when models have diverged during training. We address this limitation by first aligning the models' parameter spaces, leveraging the inherent permutation, rotation, and scaling symmetries of Transformer architectures. We adapt parameter space alignment for modern Grouped-Query Attention (GQA) and SwiGLU layers, exploring both weight-based and activation-based approaches. Using this alignment-first strategy, we successfully transfer advanced reasoning skills to a non-reasoning model. Experiments on challenging reasoning benchmarks show that our method consistently outperforms standard task arithmetic. This work provides an effective approach for merging and transferring specialized skills across evolving LLM families, reducing redundant fine-tuning and enhancing model adaptability.