LDT-Coord: Coordinating Heterogeneous LLM Agents via a Digital Twin

Sh0ny

13 июля 2026

1 min read

In short

Researchers have proposed the LDT-Coord framework for effectively coordinating teams of heterogeneous LLM agents in physical environments. The approach uses a lightweight digital twin to resolve conflicts and reduces communication overhead by more than 70 times.

Teams of robotic agents based on diverse large language models (LLM) are increasingly being deployed in physical systems: smart factories, warehouses, and service robotics. However, their collaborative operation requires coordination mechanisms capable of functioning under limited network resources. Existing approaches, which rely on multi-turn natural-language dialogues, give rise to three problems: increasing communication costs as team size grows, quality dependence on LLM capabilities, and action delays due to iterative coordination.

Solution: LDT-Coord

To address these issues, we propose LDT-Coord—a network coordination framework built on a lightweight digital twin (DT). Here’s how it works:

Each agent independently selects an action and sends its decision to the DT server along with a structured time constraint on shared resources. This decouples the quality of coordination from the LLM’s ability to reason in natural language.
On the DT side, a training-free orchestrator (rule-based orchestrator) resolves conflicts between agents and returns instructions to prevent them.
To further reduce the communication load, agent reporting is formulated as a constrained partially observable Markov decision process (C-POMDP). The PPO-Lagrangian algorithm is used to solve the problem.

Results

Simulations showed that LDT-Coord achieves a task success rate comparable to that of traditional coordination methods. At the same time, the framework reduces communication costs by more than 70 times and maintains stability when using heterogeneous models.

Source: cs.AI updates on arXiv.org

новости ai llm агенты

Liked this write-up? Get one like it in your inbox every week

More AI-tool write-ups on the Telegram channel — short and to the point

Subscribe on Telegram

Comments

(0)

Solution: LDT-Coord

To address these issues, we propose LDT-Coord—a network coordination framework built on a lightweight digital twin (DT). Here’s how it works:

Each agent independently selects an action and sends its decision to the DT server along with a structured time constraint on shared resources. This decouples the quality of coordination from the LLM’s ability to reason in natural language.

On the DT side, a training-free orchestrator (rule-based orchestrator) resolves conflicts between agents and returns instructions to prevent them.

To further reduce the communication load, agent reporting is formulated as a constrained partially observable Markov decision process (C-POMDP). The PPO-Lagrangian algorithm is used to solve the problem.