Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Feng, Shangbin, Wang, Zifeng, Goyal, Palash, Wang, Yike, Shi, Weijia, Xia, Huang, Palangi, Hamid, Zettlemoyer, Luke, Tsvetkov, Yulia, Lee, Chen-Yu, Pfister, Tomas
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866918165888368640
author Feng, Shangbin
Wang, Zifeng
Goyal, Palash
Wang, Yike
Shi, Weijia
Xia, Huang
Palangi, Hamid
Zettlemoyer, Luke
Tsvetkov, Yulia
Lee, Chen-Yu
Pfister, Tomas
author_facet Feng, Shangbin
Wang, Zifeng
Goyal, Palash
Wang, Yike
Shi, Weijia
Xia, Huang
Palangi, Hamid
Zettlemoyer, Luke
Tsvetkov, Yulia
Lee, Chen-Yu
Pfister, Tomas
contents We propose Heterogeneous Swarms, an algorithm to design multi-LLM systems by jointly optimizing model roles and weights. We represent multi-LLM systems as directed acyclic graphs (DAGs) of LLMs with topological message passing for collaborative generation. Given a pool of LLM experts and a utility function, Heterogeneous Swarms employs two iterative steps: role-step and weight-step. For role-step, we interpret model roles as learning a DAG that specifies the flow of inputs and outputs between LLMs. Starting from a swarm of random continuous adjacency matrices, we decode them into discrete DAGs, call the LLMs in topological order, evaluate on the utility function (e.g. accuracy on a task), and optimize the adjacency matrices with particle swarm optimization based on the utility score. For weight-step, we assess the contribution of individual LLMs in the multi-LLM systems and optimize model weights with swarm intelligence. We propose JFK-score to quantify the individual contribution of each LLM in the best-found DAG of the role-step, then optimize model weights with particle swarm optimization based on the JFK-score. Experiments demonstrate that Heterogeneous Swarms outperforms 15 role- and/or weight-based baselines by 18.5% on average across 12 tasks. Further analysis reveals that Heterogeneous Swarms discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of language models.
format Preprint
id arxiv_https___arxiv_org_abs_2502_04510
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems
Feng, Shangbin
Wang, Zifeng
Goyal, Palash
Wang, Yike
Shi, Weijia
Xia, Huang
Palangi, Hamid
Zettlemoyer, Luke
Tsvetkov, Yulia
Lee, Chen-Yu
Pfister, Tomas
Computation and Language
We propose Heterogeneous Swarms, an algorithm to design multi-LLM systems by jointly optimizing model roles and weights. We represent multi-LLM systems as directed acyclic graphs (DAGs) of LLMs with topological message passing for collaborative generation. Given a pool of LLM experts and a utility function, Heterogeneous Swarms employs two iterative steps: role-step and weight-step. For role-step, we interpret model roles as learning a DAG that specifies the flow of inputs and outputs between LLMs. Starting from a swarm of random continuous adjacency matrices, we decode them into discrete DAGs, call the LLMs in topological order, evaluate on the utility function (e.g. accuracy on a task), and optimize the adjacency matrices with particle swarm optimization based on the utility score. For weight-step, we assess the contribution of individual LLMs in the multi-LLM systems and optimize model weights with swarm intelligence. We propose JFK-score to quantify the individual contribution of each LLM in the best-found DAG of the role-step, then optimize model weights with particle swarm optimization based on the JFK-score. Experiments demonstrate that Heterogeneous Swarms outperforms 15 role- and/or weight-based baselines by 18.5% on average across 12 tasks. Further analysis reveals that Heterogeneous Swarms discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of language models.
title Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems
topic Computation and Language
url https://arxiv.org/abs/2502.04510