Speculate, then Collaborate: Fusing Knowledge of Language Models during Decoding

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Ziyao, Azmat, Muneeza, Li, Ang, Horesh, Raya, Yurochkin, Mikhail
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909542984450048
author Wang, Ziyao
Azmat, Muneeza
Li, Ang
Horesh, Raya
Yurochkin, Mikhail
author_facet Wang, Ziyao
Azmat, Muneeza
Li, Ang
Horesh, Raya
Yurochkin, Mikhail
contents Large Language Models (LLMs) often excel in specific domains but fall short in others due to the limitations of their training. Thus, enabling LLMs to solve problems collaboratively by integrating their complementary knowledge promises to improve their performance across domains. To realize this potential, we introduce a novel Collaborative Speculative Decoding (CoSD) algorithm that enables efficient LLM knowledge fusion at test time without requiring additional model training. CoSD employs a draft model to generate initial sequences and an easy-to-learn rule or decision tree to decide when to invoke an assistant model to improve these drafts. CoSD not only enhances knowledge fusion but also improves inference efficiency, is transferable across domains and models, and offers greater explainability. Experimental results demonstrate that CoSD improves accuracy by up to 10\% across benchmarks compared to existing methods, providing a scalable and effective solution for LLM-based applications
format Preprint
id arxiv_https___arxiv_org_abs_2502_08020
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Speculate, then Collaborate: Fusing Knowledge of Language Models during Decoding
Wang, Ziyao
Azmat, Muneeza
Li, Ang
Horesh, Raya
Yurochkin, Mikhail
Computation and Language
Artificial Intelligence
Large Language Models (LLMs) often excel in specific domains but fall short in others due to the limitations of their training. Thus, enabling LLMs to solve problems collaboratively by integrating their complementary knowledge promises to improve their performance across domains. To realize this potential, we introduce a novel Collaborative Speculative Decoding (CoSD) algorithm that enables efficient LLM knowledge fusion at test time without requiring additional model training. CoSD employs a draft model to generate initial sequences and an easy-to-learn rule or decision tree to decide when to invoke an assistant model to improve these drafts. CoSD not only enhances knowledge fusion but also improves inference efficiency, is transferable across domains and models, and offers greater explainability. Experimental results demonstrate that CoSD improves accuracy by up to 10\% across benchmarks compared to existing methods, providing a scalable and effective solution for LLM-based applications
title Speculate, then Collaborate: Fusing Knowledge of Language Models during Decoding
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2502.08020