Stage 4 of a four-stage run on HelpSteer2 with four objectives (helpfulness, correctness, coherence, conciseness). Preferences come from a prompted pairwise oracle (Qwen3-32B, one attribute rubric per comparison, swap-averaged over presentation order); no reward model is fitted. Base and reference are meta-llama/Llama-3.1-8B-Instruct.

All rules in this panel share prompts, samples and judged comparisons, and differ only in the per-objective weight applied inside the same update. Every rule was carried to the same four stages.

Win rate against the reference on 500 held-out prompts, averaged over two evaluation judges (Phi-4 and Llama-3.3-70B), neither of which supplied the training preferences: worst objective 0.471, average 0.582. Single seed.