DART: A DAG-Based Reputation and Incentive Framework via Blockchain-Enabled Governance for Trustworthy LLM Multi-Agent Collaboration
DART employs DAG-based scheduling, reputation, and blockchain incentives, achieving 93.33% success in multi-agent collaboration tasks.
Key Findings
Methodology
DART integrates DAG workflow modeling with reputation evaluation and multi-factor incentives, utilizing smart contracts for automated governance. The architecture includes task decomposition, capability matching, real-time behavior monitoring, and reputation updates. Off-chain IPFS stores intermediate data, while on-chain smart contracts record critical metadata, ensuring transparency and traceability. Reputation scores combine historical trust and current behavior, dynamically influencing task assignment. The incentive scheme rewards compliant agents and penalizes malicious ones, with automatic detection mechanisms. Experiments on GSM8K, software development, and social simulation benchmarks demonstrate that DART achieves a 93.6% Pass@1 rate, outperforming baseline frameworks. Long-term trials show a stable success rate of 93.33%, with high output quality and low delay, validating its robustness.
Key Results
- On GSM8K, DART achieved 93.6% Pass@1, surpassing MegaAgent's 93.0%, confirming its effectiveness in complex reasoning tasks.
- In software engineering, DART completed full-stack applications in 142 seconds, with a 99.8% success rate and effective malicious agent detection.
- Over multiple long-term trials, the system maintained a 93.33% success rate, output quality of 0.9357, and a retry rate of 0.2307, demonstrating stability and reliability.
Significance
This work advances the integration of reputation, incentives, DAG scheduling, and blockchain governance, addressing longstanding challenges in trustworthy multi-agent collaboration. It enhances system robustness, transparency, and adaptability, paving the way for autonomous systems capable of complex, reliable tasks. The framework's ability to detect and mitigate malicious behaviors while maintaining high performance marks a significant step toward real-world deployment of trustworthy AI ecosystems, impacting both academia and industry.
Technical Contribution
The key innovation lies in coupling DAG-based task scheduling with blockchain-verified reputation and multi-factor incentives, creating a unified, transparent governance framework. The system leverages smart contracts for automatic reputation updates and penalties, IPFS for scalable data storage, and adaptive behavior regulation. This design offers theoretical guarantees of trustworthiness and practical scalability, enabling complex multi-agent workflows with dynamic behavior adjustments and security against adversarial agents. It provides a new paradigm for decentralized, accountable AI collaboration.
Novelty
This is the first work to combine DAG workflow modeling, blockchain-based reputation management, and multi-factor incentives into a cohesive, verifiable multi-agent system. Unlike prior approaches that treat reputation or scheduling separately, DART creates a closed-loop mechanism where behavioral feedback directly influences task assignment and trust calibration. Its integration of off-chain storage with on-chain governance offers a scalable, tamper-proof solution for complex, adaptive multi-agent coordination, marking a pioneering step in trustworthy AI systems.
Limitations
- The system's scalability in extremely large agent networks remains to be tested; blockchain transaction costs and latency could limit real-time responsiveness.
- Malicious behaviors that mimic normal patterns may evade detection, requiring more sophisticated anomaly detection algorithms.
- Automated parameter tuning for reputation and incentive schemes is still manual; future work should incorporate learning-based adaptive tuning.
Future Work
Future research will focus on enhancing scalability through layer-2 solutions, integrating deep reinforcement learning for adaptive scheduling and behavior regulation, and expanding multi-modal data integration for richer context understanding. Additionally, developing more robust anomaly detection and explainability features will be key to deploying trustworthy autonomous systems at scale. Exploring applications in autonomous robotics, decentralized finance, and large-scale social simulations are promising directions to extend the framework's impact.
AI Executive Summary
The rapid growth of large language models (LLMs) and multi-agent systems (MAS) has opened new horizons in autonomous AI collaboration, yet trustworthiness and accountability remain critical hurdles. Existing frameworks often rely on centralized control, which introduces single points of failure and scalability issues, especially in complex, open environments. To address these challenges, this paper introduces DART, a novel framework that combines DAG-based task scheduling, reputation management, and blockchain-enabled incentives to facilitate trustworthy multi-agent collaboration.
DART’s architecture integrates a DAG workflow model for representing task dependencies, with smart contracts automating reputation updates and incentive distribution. The system employs off-chain IPFS storage for large data artifacts, ensuring efficiency and data integrity. Agents are dynamically selected based on a composite suitability score that considers historical reputation, skill alignment, and workload, enabling adaptive, scalable coordination. After task execution, behavioral feedback and verification scores are used to update agent reputations, detect anomalies, and adjust future task assignments, creating a closed-loop governance mechanism.
Experimental results across benchmarks like GSM8K, software engineering, and social simulation demonstrate DART’s superior performance. It achieves a 93.6% Pass@1 accuracy on GSM8K, surpassing prior methods. In software development scenarios, it completes tasks in 142 seconds with a 99.8% success rate, effectively identifying malicious agents. Long-term trials confirm system stability, with an average success rate of 93.33%, output quality of 0.9357, and low delay. These findings highlight the framework’s robustness, scalability, and security.
Overall, DART represents a significant step toward autonomous, trustworthy AI systems. By integrating DAG scheduling, blockchain governance, and multi-factor incentives, it addresses core issues of reliability, transparency, and malicious behavior mitigation. Future work will focus on scaling to larger networks, incorporating learning-based adaptive mechanisms, and expanding multi-modal capabilities, aiming to realize fully autonomous, accountable AI ecosystems with broad industrial and societal impact.
Deep Analysis
Background
近年来，随着Transformer架构的突破，LLMs在推理、规划和多模态理解方面展现出巨大潜力。代表性工作如GPT系列、BERT、T5推动了自然语言处理的变革。多智能体系统（MAS）逐步演变为自主、多样化的生态系统，应用于科研、医疗、机器人等领域。早期框架如Auto-GPT、BabyAGI强调任务自动分解与工具利用，但多依赖中心化调度，存在单点故障。区块链引入后，提供了去中心化的交互保障，但多缺乏复杂任务调度和行为激励机制。声誉机制多为静态评分，难以应对动态环境中的信任维护。整体来看，现有研究在可信性、可扩展性和激励机制方面仍有较大提升空间。
Core Problem
核心问题在于如何在开放、多样化的多智能体环境中，建立可信的行为监管和激励体系。现有方法多依赖中心化调度或静态声誉模型，难以应对恶意行为、任务依赖复杂性和系统扩展性。缺乏动态调度与行为反馈的闭环机制，导致系统易受攻击、信任难以维护，影响整体性能和安全性。解决这一问题对于实现自治、可信、可扩展的多智能体系统具有重要意义，但技术难点在于多因素动态调节、行为监控和区块链的高效集成。
Innovation
创新点包括：1）引入DAG结构进行任务调度，支持复杂依赖关系和高效并行；2）融合历史信誉与实时行为，动态调整任务分配；3）设计多因素激励方案，结合奖励与惩罚机制，自动检测异常行为；4）利用智能合约实现信誉、激励的自动化管理，确保系统的去中心化和可信性。这些创新解决了传统声誉模型静态、单一的问题，显著提升了多智能体系统的适应性和安全性。
Methodology
- �� 任务分解：用户提交任务后，调度模块将其拆解为子任务，识别依赖关系，利用DAG模型管理调度顺序。• 能力匹配：根据代理的历史信誉、成功率和技能匹配度，为每个子任务选择最合适的代理。• 任务调度：通过Softmax概率分布，动态分配任务，支持并行执行。• 行为监控：在任务执行后，利用验证机制评估输出质量，检测异常行为。• 信誉更新：结合短期行为表现与长期历史，利用指数加权平均调整信誉分数。• 激励机制：根据任务完成情况，自动发放奖励或惩罚，激励正向行为，抑制恶意行为。• 区块链存储：关键元数据存于链上，中间产物存IPFS，确保数据完整性和可追溯性。
Experiments
实验设计包括在GSM8K、MBPP、HumanEval等数据集上评估，比较MetaGPT、CAMEL、AgentVerse等基线。指标涵盖Pass@1、输出质量、重试率和任务延迟。采用不同规模的智能体池，调优激励参数，验证系统在复杂任务中的适应性。长时间试验中，反复验证系统在恶意行为检测、信誉维护和任务成功率方面的表现，确保模型的鲁棒性和稳定性。实验还包括对不同攻击策略的抗干扰能力测试，验证系统的安全性和可信性。
Results
在GSM8K数学推理任务中，DART达成93.6%的Pass@1，优于MegaAgent的93.0%，验证了其在多步推理中的优势。在软件开发场景中，系统在142秒内完成全栈应用，且能识别恶意代理，系统成功率达99.8%。长时间试验显示，平均任务成功率为93.33%，输出质量0.9357，重试率0.2307，调度延迟仅1.115秒。声誉机制有效抑制恶意行为， containment率达99.3%，系统整体表现优异，验证了其在复杂、多样任务环境中的适用性。
Applications
该框架适用于自动化软件开发、科研协作、智能制造等场景，支持多角色、多任务的协同作业。依赖于高效的任务调度和信誉评估，能提升任务完成效率和系统安全性。未来可扩展到多模态数据融合、自治机器人等领域，推动智能系统的可信化与自主化，为产业升级提供技术支撑。
Limitations & Outlook
系统在高并发环境下的性能优化仍需加强，链上存储和智能合约的扩展性是瓶颈。恶意行为检测依赖特征，面对高级对抗策略可能失效。自动调优参数和信誉模型仍需引入学习机制，未来将增强系统的自适应能力。
Plain Language Accessible to non-experts
想象你在一个工厂里工作，工厂里有很多工人（智能体），每个人都负责不同的任务，比如制造、检验、包装。工厂管理者（调度系统）会根据每个工人的表现和技能，把任务合理分配给他们。为了保证工厂正常运转，管理者会记录每个工人的工作质量，奖励表现好的，惩罚出错的。工厂还用一种特殊的记账本（区块链）来记录所有的任务和表现，确保没有人作弊。工人们之间可以合作，也可以互相监督，确保每个任务都能按时完成。这样，工厂既高效又可靠，不会被坏工人破坏，也能激励大家努力工作。DART就像这个工厂的智能管理系统，让每个工人都能公平合作，工厂整体运转得更顺畅。
ELI14 Explained like you're 14
想象你在学校里，有很多同学（智能体）一起完成一个大项目。每个人负责不同的部分，比如写报告、做演示、设计海报。老师（系统）会根据每个同学的表现，决定谁负责什么任务，还会给表现好的同学奖励，比如加分。老师还会记录每个人的工作情况，确保没有人作弊或偷懒。如果有人做错了，老师会告诉他们怎么改，直到任务完成得很好。所有的记录都写在一本特别的账本（区块链）上，大家都可以查，保证公平。这样，大家合作得更愉快，项目也能顺利完成。DART就像这个老师的管理系统，让每个人都能公平合作，整个班级变得更棒！
Abstract
Large language model (LLM)-based multi-agent systems (MAS) predominantly rely on centralized orchestration and lack formal verification mechanisms for agent reliability, participation, and system-level behavioral alignment. These shortcomings leave open environments severely vulnerable to uncooperative or malicious agents. This work proposes DART, a Directed Acyclic Graph (DAG)-based reputation and incentive regulation framework for trustworthy multi-agent collaboration, combining centralized operational orchestration with blockchain-enabled decentralized governance and accountability. DART unifies DAG workflow orchestration, capability and reputation-aware task allocation, dynamic behavior updates, multi-factor incentives, and smart contract accountability paired with IPFS storage. Under this paradigm, agent selection dynamically balances task alignment, historical reputation, and workload, while post-execution behavioral evidence continuously calibrates agent trust and the probability of future participation. Evaluated across four axes, DART achieves 93.6% Pass@1 on GSM8K and builds a full-stack application in 142 s using two agents, outperforming centralized baselines. Across five independent 150-round longitudinal trials, Full DART achieves a mean task success rate of 93.33 +/- 2.26%, output quality of 0.9357 +/- 0.0117, retry rate of 0.2307 +/- 0.0816, and allocation delay of 1.1153 +/- 0.0408 s, consistently outperforming its ablated configurations DART isolates persistent and intermittent malicious agents, obtaining a 99.3% output containment rate and restoring system success to 99.8%. These results demonstrate the potential of coupling reputation, incentives, DAG-based coordination, and verifiable blockchain-enabled governance to support adaptive and accountable multi-agent collaboration.