Indexed from the September 29 newsletter. Linked-source claims have not been independently verified.
Development
Meta trains a model for expert writing with RL-XAR
The author frames training for writing judgment as another attempt to automate the work of model builders.
Claim
The issue reports Meta's rubric-based adversarial RL-XAR training making Qwen3.5-27B outperform frontier models on papers and stories in the cited comparisons.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to September 29, 2026 (theinnermostloop.substack.com)
- Source linked by the September 29 issue (facebookresearch.github.io)
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.