2B Model Hits FunSearch Target Once After 1.7 Days

A developer reported that a 2B model took about 1.677 days and 22 steps to hit a correct solution just once on the FunSearch Dudeney 1917 task, suggesting that distilling a better starting policy from a 3.8 27B LoRA could improve results.

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