New mechanism ensures truthful worker feedback for LLM fine-tuning in mobile crowdsourcing

Mobile crowdsourcing platforms, such as those for navigation and traffic prediction, increasingly rely on LLM-generated content that must be aligned with human preferences. However, workers (mobile users) may strategically misreport their preferences to maximize influence or payment. Existing pipelines, like EM-based weight estimation, fail to identify the most accurate worker in this online setting, leading to linear regret O(T) over T time slots.

To solve this, Hao and Duan formulate a dynamic Bayesian game to model the multi-agent online learning interaction between the platform and strategic workers. They propose a novel online weighted aggregation mechanism that dynamically adjusts each worker's weight according to their feedback accuracy. The mechanism provably ensures truthful feedback from workers and achieves sublinear regret O(√T). An extension to scenarios with limited per-slot feedback also guarantees O(√T) regret. Experiments on LLM fine-tuning with real-world data demonstrate significant gains over benchmark methods.