Share your thoughts, 1 month free Claude Pro on usSee more

Click-in-target on WinDOM in-distribution 2,000-example subsample (held-out)

82.1Click-in-target Accuracy

SFT

Updated 1mo ago

Evaluation Results

Method	Links
SFT 2026.06		82.1
SFT +RL Late 2026.06		80.8
SFD-4B +RL Late 2026.06		79.7
SFD-EMA +RL Late 2026.06		79
SFT +RL Early 2026.06		77.2
SFD-EMA 2026.06		76.8
SFD-4B 2026.06		76.5
SFD-4B +RL Early 2026.06		76.5
SFD-EMA +RL Early 2026.06		75.4
Qwen3.5-4B (base) 2026.06		71.2
Qwen3.5-2B (base) 2026.06		70