Leveraging Parameter Space Symmetries for Reasoning Skill Transfer in LLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Horoi, Stefan, Cho, Sangwoo, Chakraborty, Supriyo, Zhang, Shi-Xiong, Sahu, Sambit, Wolf, Guy, Winata, Genta Indra
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909901065814016
author Horoi, Stefan
Cho, Sangwoo
Chakraborty, Supriyo
Zhang, Shi-Xiong
Sahu, Sambit
Wolf, Guy
Winata, Genta Indra
author_facet Horoi, Stefan
Cho, Sangwoo
Chakraborty, Supriyo
Zhang, Shi-Xiong
Sahu, Sambit
Wolf, Guy
Winata, Genta Indra
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2511_10850
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Leveraging Parameter Space Symmetries for Reasoning Skill Transfer in LLMs
Horoi, Stefan
Cho, Sangwoo
Chakraborty, Supriyo
Zhang, Shi-Xiong
Sahu, Sambit
Wolf, Guy
Winata, Genta Indra
Computation and Language
Artificial Intelligence
Machine Learning
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.
title Leveraging Parameter Space Symmetries for Reasoning Skill Transfer in LLMs
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2511.10850