Ilmu Komputer & AI editorial
Guixu: Valuation-Driven Data Discovery for Autonomous AI Agents with On-Chain Attestation
The core problem
Autonomous agents increasingly depend on external data to complete downstream tasks such as model training and decision support. Yet the data discovery systems they rely on remain largely retrieval-oriented: they surface candidate datasets from heterogeneous sources but offer limited support for estimating task-specific utility, selecting cost-effective datasets under budget constraints, or incorporating trustworthy feedback from prior usage. This gap matters because an agent that can find data but cannot judge its value for a specific task, or cannot pay for it within a budget, is only half-equipped for autonomous operation.
The paper introduces **Guixu**, a valuation-driven data discovery system for autonomous agents. Guixu is positioned as a demonstration system that moves an agent beyond keyword-based dataset retrieval toward task- and budget-aware, trustworthy data discovery and procurement. Its design rests on three pillars: a three-phase valuation pipeline with proxy-label propagation and multi-round knapsack optimization for task-aware data valuation; integration of an agentic payment protocol to enable budget-constrained data procurement workflows; and use of an on-chain
Innovation
The paper is presented as a demonstration system, so its reported results center on the workflow that attendees can exercise interactively rather than on benchmark tables. The demonstration highlights how Guixu enables an agent to move beyond keyword-based dataset retrieval toward task- and budget-aware, trustworthy data discovery and procurement. Attendees can explore the full pipeline: entering a natural-language task specification, watching multi-source search surface candidate datasets, observing the three-phase valuation pipeline assign task-specific values, and following the procurement step through the agentic payment protocol to verifiable transaction feedback on the on-chain data market.
The concrete artifacts of the demonstration are the components themselves: proxy-label propagation for label-scarce valuation, multi-round knapsack optimization for budget-constrained selection, an agentic payment protocol for autonomous procurement, and on-chain attestation signals for verifiability. Together these produce a discovery experience in which the agent's output is not merely a list of datasets but a cost-effective, task-aware selection accompanied by trustworthy feedback. The
Why it matters
Guixu's contribution is best understood as a reframing of the data discovery problem. Retrieval-oriented systems optimize for surfacing candidates; Guixu optimizes for utility under constraints. That shift has three consequences. First, valuation becomes the central primitive: without task-aware value estimates, budget-constrained selection degenerates into arbitrary choice. Proxy-label propagation is the enabling technique that makes valuation possible when target-task labels are absent. Second, the knapsack formulation makes the budget constraint explicit and optimizable, and the multi-round variant acknowledges that value estimates are uncertain and improve with feedback. Third, trust is treated as an infrastructure problem: on-chain attestation and an on-chain data market provide verifiable signals about datasets and prior usage, while the agentic payment protocol closes the loop from discovery to procurement.
The design also raises open questions. Proxy labels are, by construction, approximations; their fidelity bounds the quality of downstream valuation, and the paper does not quantify that bound. Multi-round knapsack optimization introduces a trade-off between exploration (spending budget to learn better value estimates) and exploitation (spending budget on currently high-value datasets) that is not resolved in the abstract. On-chain attestation offers verifiability but ties discovery to the economics and latency of the underlying chain. Finally, the demonstration format means the system's behavior is shown rather than benchmarked, leaving comparative evaluation against retrieval-oriented baselines as future work.
Within the taxonomy candidates of Architecture, Cybersecurity, Network, and Cryptography, Guixu sits at the intersection of system architecture and cryptographic verifiability: its on-chain attestation and payment components are cryptographic mechanisms, while its valuation pipeline and agentic workflow are architectural. The system's emphasis on trustworthy feedback from prior usage also connects to security concerns around data provenance and market integrity.
Who should read this
Opening member content…