Deletion Learning for Multi-Agent Systems (DL-MAS): A General Framework for Enhancing Collective Intelligence Through Pre-Training, In-Training, and Post-Training Deletion
Abstract
Multi-agent systems are commonly improved by adding agents, expanding communication networks, introducing new capabilities, increasing memory, incorporating additional knowledge, and strengthening coordination mechanisms. Although these approaches can improve collective performance, additional components may also introduce redundancy, communication overhead, conflicting decisions, unnecessary computation, coordination complexity, and interference between learned behaviors. Consequently, improving a multi-agent system may require not only adding or strengthening components but also selectively deleting components that are unnecessary, harmful, inefficient, or incompatible with the system's objectives.
This paper proposes Deletion Learning for Multi-Agent Systems (DL-MAS), a general conceptual framework that treats deletion as an explicit operation for designing, training, optimizing, and maintaining multi-agent systems. Unlike approaches that focus exclusively on removing training data, pruning neural networks, or eliminating individual agents, DL-MAS considers a broad range of potential deletion targets, including agents, capabilities, parameters, representations, memories, communication links, interaction rules, coordination mechanisms, task assignments, behavioral policies, computational operations, objectives, and external dependencies. The framework organizes deletion into three principal lifecycle stages: pre-training deletion, in-training deletion, and post-training deletion. It further considers operational deletion during system deployment and continued adaptation.
DL-MAS conceptualizes deletion as a selective, objective-dependent intervention whose value must be evaluated at the collective system level rather than inferred from the isolated properties of the removed component. A formal model is introduced to represent deletion actions, their costs, their effects on collective performance, and their interactions with retained components. The framework also develops a taxonomy of deletion targets, a general deletion procedure, evaluation criteria, research propositions, and an experimental methodology. The central proposition is that, under appropriate conditions, systematically identifying and removing unnecessary or counterproductive components can improve the efficiency, reliability, adaptability, and coordination of multi-agent systems while preserving required capabilities. DL-MAS provides a foundation for investigating deletion as a general design and optimization dimension of collective intelligence.
Related
- Deletion Learning (DL): A General Framework for Enhancing Learning Systems Through Pre-Learning, In-Learning, and Post-Learning Deletion
- Enhancing Intelligence Through the Suppression of Other Intelligence: A General Framework
- Enhancing Intelligence Through the Suppression of Other Intelligence: A General Framework for Multi-Agent Systems
- VORTEX DYNAMISM AND THE LEMNISCATIC MODEL OF SYNTHECRACY: An operational supplement to the methodology of collective intelligence and adaptive societal learning
- Enhancing Large Language Models Through the Suppression of Other Intelligence: A General Framework
BibTeX
@misc{zhang2026deletion,
title = {{Deletion Learning for Multi-Agent Systems (DL-MAS): A General Framework for Enhancing Collective Intelligence Through Pre-Training, In-Training, and Post-Training Deletion}},
author = {Jincheng Zhang},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23464366},
url = {https://zenodo.org/records/23464366},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Deletion Learning for Multi-Agent Systems (DL-MAS): A General Framework for Enhancing Collective Intelligence Through Pre-Training, In-Training, and Post-Training Deletion
AU - Jincheng Zhang
PY - 2026
DA - 2026/10/19
PB - Zenodo
DO - 10.5281/zenodo.23464366
UR - https://zenodo.org/records/23464366
AB - Multi-agent systems are commonly improved by adding agents, expanding communication networks, introducing new capabilities, increasing memory, incorporating additional knowledge, and strengthening coordination mechanisms. Although these approaches can improve collective performance, additional components may also introduce redundancy, communication overhead, conflicting decisions, unnecessary computation, coordination complexity, and interference between learned behaviors. Consequently, improving a multi-agent system may require not only adding or strengthening components but also selectively deleting components that are unnecessary, harmful, inefficient, or incompatible with the system's objectives. This paper proposes Deletion Learning for Multi-Agent Systems (DL-MAS), a general conceptual framework that treats deletion as an explicit operation for designing, training, optimizing, and maintaining multi-agent systems. Unlike approaches that focus exclusively on removing training data, pruning neural networks, or eliminating individual agents, DL-MAS considers a broad range of potential deletion targets, including agents, capabilities, parameters, representations, memories, communication links, interaction rules, coordination mechanisms, task assignments, behavioral policies, computational operations, objectives, and external dependencies. The framework organizes deletion into three principal lifecycle stages: pre-training deletion, in-training deletion, and post-training deletion. It further considers operational deletion during system deployment and continued adaptation. DL-MAS conceptualizes deletion as a selective, objective-dependent intervention whose value must be evaluated at the collective system level rather than inferred from the isolated properties of the removed component. A formal model is introduced to represent deletion actions, their costs, their effects on collective performance, and their interactions with retained components. The framework also develops a taxonomy of deletion targets, a general deletion procedure, evaluation criteria, research propositions, and an experimental methodology. The central proposition is that, under appropriate conditions, systematically identifying and removing unnecessary or counterproductive components can improve the efficiency, reliability, adaptability, and coordination of multi-agent systems while preserving required capabilities. DL-MAS provides a foundation for investigating deletion as a general design and optimization dimension of collective intelligence.
ER -