10.
Training Common Crawl chronologically reduces recency failure
Models trained on web data ordered from 2018 to 2025 performed much better on recent facts than models seeing the same data in a nonsequential order
@a_karvonen on Backlist
3 appearances on the backlist front page in the last 30 days.
84.
Some cool work that I co-mentored with (x.com)
Some cool work that I co-mentored with @NeelNanda5 I recommend the appendix section on practical AO evaluation details. In particular, consensus sampling significantly reduces hallucinations, and eval performance majorly improves with
30.
I am truly blown away by Qwen-3.5-27B.
I am truly blown away by Qwen-3.5-27B. It's doing better than Haiku 4.5 on my OOD interp task that involves 50k context in an agentic setting. Such a great cheap model for research tasks.