New Algorithm Learns Rewards and Worker Reliability from Comparisons.

Learning item rewards from pairwise comparisons is a fundamental problem across various fields, including recommendation systems and fine-tuning large language models. Often, these comparisons are sourced from crowdworkers, whose reliability can be compromised by limited knowledge or malicious "spamming" behavior. A key challenge is to simultaneously determine both the item rewards and the competency of the workers providing the comparisons. Researchers have developed an EM-based algorithm to address this, building upon the Boltzmann-rational model which incorporates worker competencies into the standard Bradley-Terry-Luce model. The algorithm uses Polya-Gamma latent variables to transform the logistic likelihood into a conditionally Gaussian form, simplifying optimization and enabling a tractable E-step. This technique effectively reduces the problem to a matrix sensing task, for which theoretical convergence guarantees have been established. Extensive experiments on both real-world and synthetic datasets confirm the algorithm's advantages over existing baselines. It exhibits remarkable robustness against both spammers and adversarial workers, proving its practical utility in realistic crowdsourcing and reward learning environments. The code and data are publicly available.