HPO Models Beat Classical Ranking on LLM Agent Traces

Classical ranking models like Mallows and Plackett-Luce assume total-order consensus rankings, limiting their ability to handle incomparabilities in preferences. Recent work relaxes this using partial orders (posets), but real-world data often has group structure—e.g., preferences from different user segments or AI agent behaviors. A new paper introduces Hierarchical Partial-Order (HPO) models, which extend poset-based models to grouped data via a hierarchy of latent posets. This framework parallels mixture model extensions of Mallows and Plackett-Luce, enabling principled information sharing across groups while preserving partial-order structure. The authors show that Plackett-Luce and its hierarchical variants are special cases of HPO models.

Experimental validation on synthetic data and two real-world datasets—pairwise acoustic preference data and LLM agent traces—demonstrates that HPO and its unsupervised clustering extension (HCPO) consistently outperform existing approaches in predictive performance and structural interpretability. Bayesian inference for the latent poset hierarchy uses Markov chain Monte Carlo methods, detailed across 34 pages with 13 figures and 8 tables. The work has immediate applications for preference aggregation in AI alignment, recommendation systems, and any domain where ranked preferences exhibit group-level patterns.