LMArena open-sources web preference datasets and Interaction-as-a-Judge paradigm

LMArena released a paper on Interaction-as-a-Judge and open-sourced two web app preference datasets, showing that fine-tuning Qwen-3.5-9B with interaction-only rewards improves win rate by 6.7pp.

2026-11-13 ~ 2026-11-13 · 2 related posts

• Interaction-as-a-Judge beats code rewards: +6.7pp win rate, 35.7% less code bloat — arena · 2026-11-13
• LMArena open-sources 6,390-pair web app preference dataset, shows interaction-based judging beats code review — arena · 2026-11-13