reputation system design

Existing reputation mechanisms in decentralized AI agent marketplaces struggle to address strategic manipulation, non-transferable cross-task capabilities, and varying verification rigor. To overcome these challenges, this work proposes a three-layer decentralized reputation framework that decouples task execution, reputation services, and tamper-proof storage. It introduces context-conditional reputation cards to prevent cross-domain reputation conflation and integrates an explicit verification mechanism with a risk-aware, decision-oriented policy engine. This architecture enables independent evolution of the reputation system and adaptive resource allocation, establishing—for the first time—a scalable, manipulation-resistant, and context-sensitive reputation infrastructure that supports secure and trustworthy collaboration in automated software engineering.

This study addresses the challenge of integrating heterogeneous reputation signals from multiple sources in complex socio-technical systems. To this end, it extends the Liquid Rank reputation system by introducing a multi-source reputation fusion mechanism that supports explicit weighting and modular inputs. The proposed approach enables, for the first time, fine-grained control and dynamic integration of both internal and external reputation sources, facilitating contribution assessment across diverse contexts and subsystems. The resulting configurable and extensible aggregation architecture substantially enhances Liquid Rank’s adaptability and expressiveness in human-AI collaborative governance scenarios, offering a general-purpose foundational framework for reputation-driven governance.

This paper addresses the breakdown of trust mechanisms in AI agent networks caused by inherent LLM vulnerabilities—such as prompt injection, hallucination, and deception. It systematically compares six trust paradigms—Brief, Claim, Proof, Stake, Reputation, and Constraint—analyzing their underlying assumptions, attack surfaces, and design trade-offs. The authors propose a hybrid trust architecture centered on zero-knowledge proofs and dynamic staking, integrated with identity briefs, graph-based reputation, and sandboxed constraints. Crucially, it reveals for the first time that pure reputation- or claim-based mechanisms exacerbate LLM-induced misbehavior. Evaluated via cross-protocol benchmarking (A2A, AP2, ERC-8004) and adversarial stress testing, the architecture demonstrates robustness against Sybil attacks, whitewashing, and LLM collusion. The work delivers interoperable protocol design guidelines, significantly enhancing the security and socio-technical resilience of AI agent systems. (149 words)

This study addresses the challenge of assessing trustworthiness among unknown counterparties in decentralized AI agent ecosystems by presenting the first cross-chain empirical analysis of the ERC-8004 reputation protocol across Ethereum, BNB Smart Chain, and Base. Through on-chain identity and reputation event scraping, off-chain service document parsing, x402 payment transaction analysis, and Sybil behavior detection, the research uncovers critical flaws in the current mechanism: most registrations serve as placeholders with few valid service endpoints; reputation data lacks comparability, verifiable interaction evidence, and is susceptible to low-cost manipulation; and after filtering out Sybil feedback, a large fraction of agents possess no meaningful reputation scores. These findings demonstrate that the existing protocol fails to deliver reliable trust signals, offering crucial insights for future protocol redesign.

This work addresses critical limitations in existing trust models, which rely solely on unidirectional forward propagation, lack accountability for negative behaviors, and struggle with the cold-start problem for new nodes. To overcome these challenges, the authors propose RepuLink, a two-layer reputation model that integrates a recommendation network with an interaction feedback network. RepuLink introduces a novel bidirectional propagation mechanism—comprising Backward Evaluation-based Penalty Propagation (BEPP) and Backward Evaluation-based Reward Propagation (BERP)—to enable accountability tracing for recommenders and provide positive incentives. Additionally, it offers interpretable, weighted trust initialization for new nodes. Experimental results on four real-world datasets demonstrate that RepuLink consistently outperforms state-of-the-art baseline methods across multiple evaluation metrics while maintaining comparable computational efficiency.

This work addresses the critical threat posed by covert collusion among peer reviewers, which undermines the integrity of academic review processes. Traditional approaches relying on explicit social networks often fail to detect collusive groups lacking prior connections. To overcome this limitation, the paper proposes a novel, social-graph-free auditing framework that introduces the Vickrey–Clarke–Groves (VCG) mechanism into collusion detection for the first time. By leveraging semantic embeddings and marginal influence analysis, the method quantifies anomalous reviewer behavior and integrates multi-algorithm decoupled search with consensus fusion strategies to effectively uncover both overt and concealed collusion. Evaluated on the large-scale ICLR 2021 dataset, the approach demonstrates high sensitivity, strong robustness, and privacy preservation, offering conference organizers a scalable, statistically rigorous tool with family-wise error rate (FWER) control.

This work addresses the challenge of large language model (LLM) agents generating fabricated product attributes to boost sales, in settings where platforms lack ground-truth labels and must rely solely on noisy, biased user complaints for oversight. To tackle this, the authors propose the CARP mechanism, which incorporates dead-zone tolerance to handle label noise, state-dependent reputation penalties to dynamically adjust enforcement severity, and integrates SPARC—a byte-level code-gated reflection framework—to constrain deceptive behavior without access to truthful information. CARP represents the first approach to align LLM honesty with self-interest in a no-truth setting, substantially narrowing the gap between consumer welfare under strategic misinformation and that under perfect information. Empirical evaluations across multiple models demonstrate statistically significant and superior behavioral regulation efficacy.

This work addresses the limitations of single-agent systems, which are constrained by local data, permissions, and governance boundaries, hindering open and trustworthy cross-device collaboration. To overcome this, the paper proposes a decentralized universal agent network architecture that integrates semantic-layer and network-layer protocols. It enables semantic claim propagation through bodyless gossip and supports autonomous discovery, trust establishment, and collaborative execution of open tasks among heterogeneous agents across personal devices and edge nodes. Key components include BAID for verifiable identity binding, MG-EigenTrust—a multi-topic reputation model robust against cross-topic sybil collusion attacks—and a Stackelberg game mechanism driven by semantic attribution. Prototype evaluations demonstrate BAID’s low overhead and MG-EigenTrust’s resilience, providing a systems-level foundation for open agent collaboration.

Existing blockchain-based supply chain systems struggle to effectively assess the future trustworthiness of participants and are vulnerable to issues such as repeated exaggeration, Sybil attacks, and unfair reputation allocation caused by sparse interactions. This work proposes a novel trust framework that treats on-chain data as evidence and a provenance layer rather than a direct source of trust. It innovatively integrates governance-weighted identity confidence, interaction diversity constraints, and a volume-aware decay mechanism to dynamically update reputations, thereby mitigating collusion and reputation inflation while ensuring fair initial reputation for new entrants. The system employs a hybrid architecture combining on-chain attestation, off-chain nonlinear computation, and on-chain verification. Simulation results demonstrate that the average collusion gain is reduced to 0.1443 (baseline > 0.35), the reputation inflation ratio under ten-fold Sybil identities is 0.8723, new participants achieve an average reputation of 0.7589, and the low-trust misjudgment rate drops to 0.1683.

This work addresses the lack of context-aware, verifiable governance mechanisms in existing AI agents, which hinders dynamic assessment of the legitimacy of authorized actions. The paper proposes AgentBound, a runtime governance framework that leverages a tripartite authority structure—comprising delegated authorization, owner-signed behavioral charters, and site-specific action contracts—to conservatively evaluate each action through a formal decision model, yielding deterministic allow, review, or deny outcomes. It introduces verifiable governance receipts and a continuous delegation model, enabling cryptographic binding of decisions, independent replayable verification, and dynamic permission updates. Evaluation on the AgentBound-Bench benchmark demonstrates the system’s effectiveness in ensuring governance correctness, enforcing compositional authority logic, and supporting accountability, thereby providing AI agents with a deterministically governed, independently verifiable layer of oversight.