A new research framework aims to make data valuation faster, cheaper, and privacy-safe for decentralized AI training pipelines.
The paper introduces DMVM, short for Decentralized Multi-task Valuation via Model Merging. Instead of retraining models from scratch or running Shapley-value calculations — both of which are computationally expensive — DMVM infers how much each dataset contributes by examining how models trained on different data combine at the parameter level. The researchers call this technique task arithmetic. A secure aggregation protocol lets multiple parties collaborate on the valuation without any single party seeing another's raw data or model weights. The team validated the approach on computer vision and natural language processing benchmarks and provided theoretical error bounds on the approximation.
Data marketplaces are a real and growing problem in AI development. As more organizations try to buy and sell training data, fair pricing is hard to establish — especially when a dataset's value shifts depending on which tasks a buyer cares about. Most existing methods assume a single task and a central coordinator, neither of which reflects how multi-party AI pipelines actually work. DMVM's model-merging approach sidesteps both constraints simultaneously.
The catch, as with most academic frameworks, is the gap between a controlled benchmark and a messy production data marketplace — task arithmetic is still a relatively young technique, and how well it holds up when datasets vary wildly in quality and domain remains an open question.