Meta's RL-XAR Uses Expert-Aligned Rubrics to Fix AI Slop
Meta AI's Jason Weston team introduced RL-XAR, which replaces LLM judges with learned expert-aligned rubrics as RL rewards to combat low-quality 'AI slop' writing. Training Qwen3.5-27B on scientific papers, story continuation, and Wikipedia writing, the model achieved 89% and 95% human-eval win rates on papers and stories.
2026-09-28 ~ 2026-09-28 · 4 related posts
- Meta claims it can fix AI slop with RL-XAR: RL on expert-aligned rubrics for writing — jaseweston · 2026-09-28
- RL-XAR results: 89% and 95% human win rates on papers and stories, judged across model families — jaseweston · 2026-09-28
- RL-XAR recipe: replace LLM judges with expert-aligned rubrics for RL — ivan_bezdomny · 2026-09-28
1 near-duplicate retellings: jaseweston