LbMAS Implementation: Multi-Agent LLM System

• LbMAS is a multi-agent system architecture that uses a shared blackboard to coordinate diverse LLM agents for dynamic problem solving.
• The framework leverages a control unit for real-time agent selection and iterative consensus mechanisms, greatly reducing token usage.
• Experimental results indicate that LbMAS enhances accuracy and cost efficiency across complex reasoning benchmarks compared to static MAS designs.

LbMAS (Lattice Boltzmann Multi-Agent System) Implementation refers to the design and deployment of multi-agent architectures for LLMs where agent coordination, communication, and decision-making are managed via a blackboard architecture. In this paradigm, agents with diverse, potentially dynamic roles interact through a shared memory medium (the blackboard), enabling iterative, information-rich problem-solving cycles. This approach aims to facilitate complex, open-ended AI reasoning tasks by supporting dynamic agent selection, efficient information propagation, robust consensus mechanisms, and significant reductions in inference token cost (Han et al., 2 Jul 2025).

1. System Architecture and Core Principles

The LbMAS framework is anchored in three primary components: the blackboard, an LLM agent group, and a control unit.

• Blackboard: The central shared memory, composed of public (shared) and private (if needed) spaces, stores all agent-generated messages, intermediate inferences, and interaction histories. This centralized public memory replaces individual agent memory, reducing token usage and ensuring all agents operate over a synchronized conversational context.
• LLM Agent Group: Agents are instantiated with roles such as planner, decider, critic, conflict-resolver, and cleaner. In addition, query-specific experts are dynamically generated per query through an agent generator. Each agent reads the current blackboard and posts its output back to it, contributing knowledge, plans, or assessments.
• Control Unit: Implemented by an LLM, this module dynamically selects which agents will act in each round, based on the current task and the content of the blackboard. This enables on-the-fly adaptation of the agent workflow, matching system behavior to the evolving problem state.

This architectural separation permits a high degree of configurability and supports multi-round interactive reasoning, essential for tasks requiring iterative refinement or consensus (Han et al., 2 Jul 2025).

2. Blackboard Cycle: Information Flow and Agent Selection

The iterative solution process—the blackboard cycle—plays a central operational role:

• At each round, the control unit examines both the user query $q$ and the current blackboard state $B$ to select a subset of agents to activate for that round.
• Each agent $A_i$ receives the full blackboard as its context and produces a message $m_i$, which is appended to $B$.
• This cycle repeats until a consensus or stopping criterion is met (e.g., a decider agent produces an answer, or a round cap is reached).

This process is formalized as:

$\begin{algorithm} \KwIn{Query %%%%5%%%%, Agent group %%%%6%%%%, Control Unit %%%%7%%%%, Maximum rounds %%%%8%%%%, Blackboard contents %%%%9%%%%, Solution Extraction module %%%%10%%%%} \KwOut{Solution} %%%%11%%%% \While{%%%%12%%%%}{ %%%%13%%%% \For{%%%%14%%%% \KwTo %%%%15%%%%}{ %%%%16%%%% %%%%17%%%% } %%%%18%%%% } %%%%19%%%% \end{algorithm}$

The Agent Generation module for dynamic agent creation is denoted:

$\{(E_1, D_1), ..., (E_n, D_n)\} = AG(q, I)$

with $E_i$ representing identity and $D_i$ its domain, ensuring query-specific expertise and model diversity.

Agent selection at round $t$ employs:

$$\{E_{i_1}, ..., E_{i_j}\} = ConU(q, B, \{D_1, ..., D_n\})$$

This architecture supports both fixed and fully dynamic agent workflows, significantly generalizing prior static or purely sequential MAS designs.

3. Consensus and Decision Mechanisms

LbMAS terminates with consensus achieved by one of two methods:

• Decider Agent: A specially designated agent reads the blackboard and determines if a converged, final solution has been reached.
• Majority Voting (Consensus): Each active agent proposes an answer. Cumulative similarity scores are computed,

$$V(a_i) = \sum_{j=1, j\neq i}^N \text{sim}(a_i, a_j)$$

$\operatorname{argmax}_{a_i} V(a_i)$

selecting the solution most similar to the ensemble’s collective output.

This flexible consensus ensures the robustness of the final output against agent disagreement and noise, improving reliability relative to single-agent or rigidly-structured systems.

4. Experimental Findings and Benchmarks

In empirical evaluation across knowledge/reasoning (MMLU, ARC-Challenge, GPQA-Diamond, BBH) and mathematics (MATH, GSM8K) datasets, LbMAS achieved superior or competitive performance compared to both static multi-agent and dynamic autonomous MAS methods.

• On MMLU and GPQA, average accuracy improved by around 4.33–5.02% over Chain-of-Thought and static baselines.
• On MATH, LbMAS yielded 72.60% accuracy with only 4.72M tokens, reflecting substantial cost savings relative to methods such as GPTSwarm and AFlow, which require an explicit (and costly) search for the best workflow.

Dynamically generated query-specific agent ensembles (with randomized base LLM selection among models like Llama-3.1-70b-Instruct and Qwen-2.5-72b-Instruct) further improved performance and robustness. More complex queries required more blackboard rounds and benefited from extended inter-agent dialogue (e.g., for conflict resolution and message cleaning), while simple queries often converged in a single step (Han et al., 2 Jul 2025).

5. Unique Implementation Features and Advantages

Several implementation aspects distinguish LbMAS:

• Dynamic Agent Selection: The control unit adaptively chooses agents based on real-time blackboard content, avoiding rigid workflow templates and allowing complex, non-predefined reasoning paths.
• Decentralized, Public Memory: By using a single public blackboard, per-agent prompt lengths are reduced, alleviating the memory bottleneck and reducing token consumption.
• Specialized Agent Roles: Roles such as critic, conflict-resolver, and cleaner promote answer quality by filtering errors and redundant information.
• No Pretraining Overhead: Unlike autonomous MAS approaches requiring offline workflow search, LbMAS directly adjusts its strategy online based on current problem context, minimizing training and inference costs.

This architecture is particularly adept for domains where problem structure is unanticipated or evolves dynamically during inference.

6. Implications, Limitations, and Prospects

The integration of blackboard architecture within LLM-based MASs demonstrates that iterative, agent-mediated solution refinement can yield both strong accuracy and significant token efficiency. LbMAS supports complex, open-ended reasoning without statically engineered agent workflows—addressing a pronounced limitation in prior MAS frameworks that are poorly suited for dynamic or ill-structured problems.

A plausible implication is that LbMAS-style architectures may become a dominant pattern for integrating LLMs in complex, distributed, or collaborative AI workflows, particularly where problem structure emerges during computation (Han et al., 2 Jul 2025). However, full system performance depends on the sophistication of the control unit and the diversity of base LLMs; scaling to even larger agent pools or more diverse agent templates remains an open research target.

7. Summary Table: Key Elements of LbMAS

Component	Function	Implementation Highlights
Blackboard	Shared public memory for all agent messages	Centralized dialogue history, no per-agent memory required
Agent Group	Diverse specialized (and query-generated) agents	Dynamic instantiation by Agent Generation module
Control Unit	Chooses agents per round	Context-dependent selection, LLM-driven decision logic
Consensus Mechanism	Aggregates agent proposals to final answer	Decider agent or similarity-based majority voting
Efficiency	Reduces token usage, obviates workflow search	Multiround, context-adaptive cycles; no offline search

These design elements underpin the improved effectiveness and efficiency of LbMAS, as demonstrated by experimental results across multiple challenging benchmarks (Han et al., 2 Jul 2025). The approach shows a path toward scalable, adaptive, and collaborative multi-agent AI systems leveraging LLMs.