Interesting new TIM masking scheme. Directly use the absolute difference rather than the relative difference
Cool stuff
TIM masking scheme uses absolute difference instead of relative difference
AISummary
@nrehiew_ proposes using the absolute difference rather than the relative difference in a new TIM masking scheme. Prime Intellect says IcePop collapsed around step 1.2k when trainer-inference KL rose 20×, while IPO trained through, masking about 0.25% of tokens in agentic SWE RL with GLM-4.5-Air.
Post on XView on X
whVerified on X
@nrehiew_
We tested it in a deliberately hard setting: agentic SWE RL with GLM-4.5-Air • IcePop collapsed around step 1.2k when trainer-inference KL blew up 20× • IPO trained straight through, masking only ~0.25% of tokens
Source: wh · x.comPublished