RL-XAR results: 89% and 95% human win rates on papers and stories, judged across model families
jaseweston · x · 2026-09-28
- Meta's RL-XAR thread (4/5) details training: Qwen3.5-27B trained with RL on learned rubrics scored by a cross-family Qwen3.8-2.4T-A95B judge, with final results verified by GPT5-6.
- Strong gains across scientific paper sections, story continuations and Wikipedia pages, largest on the first two.
- Small human evals show 89% (papers) and 95% (stories) win rates for RL-XAR; training fixes over-scoping and section focus in papers, removes clichéd writing in stories.
Related event: Meta's RL-XAR Uses Expert-Aligned Rubrics to Fix AI Slop(4 posts)→
More from Research
- New blog surveys world models: definitions, SOTA, and eval axes — mervenoyann · 2026-09-29
- Researchers Question Pedagogical RL: Student-Likelihood May Be a Flawed Learnability Proxy — novasarc01 · 2026-09-29
- Researcher Flags Risk That Pedagogical RL's Student-Likelihood Optimization Filters Rare Reasoning — novasarc01 · 2026-09-29
- Goodfire grants geometric_intel lab funding for AI interpretability research — ninamiolane · 2026-09-29
- Bespoke Labs Launches AutoResearchExam Benchmark, Again, With a Demo Video — gregd_nlp · 2026-09-28
- Agentick benchmark accepted at NeurIPS: LLM vs RL agents on same tasks, no single winner — pcastr · 2026-09-28