Back to Blackwell: Closing the Loop on Intransitivity in Multi-Objective Preference Fine-Tuning

arXiv:2602.19041v2 Announce Type: replace Abstract: A recurring challenge in preference fine-tuning (PFT) is handling extit{intransitive} (i.e., cyclic) preferences. Intransitive preferences often stem from either extit{(i)} inconsistent rankings along a single objective or extit{(ii)} scalarizing multiple objectives into a single metric. Regardless of their source, the downstream implication of intransitive preferences is the same: there is no well-defined optimal policy, breaking a core assumption of the standard PFT pipeline. In response, we propose a novel, game-theoretic solution concept, the extit{Maximum Entropy Blackwell Winner} (extit{MaxEntBW}), that is well-defined under multi-objective intransitive preferences. To enable computing MaxEntBWs at scale, we derive exttt{PROSPER}: a provably efficient PFT algorithm. Unlike prior self-play techniques, exttt{PROSPER} directly handles multiple objectives without requiring scalarization. We then apply exttt{PROSPER} to the problem of fine-tuning large language models (LLMs) from multi-objective LLM-as-a-Judge feedback (e.g., rubric-based judges), a setting where both sources of intransitivity arise. We find that exttt{PROSPER} outperforms all baselines considered across both instruction following and general chat benchmarks, releasing trained model checkpoints at the 7B and 3B parameter scales.

Source: arXiv cs.LG | 2026-05-07