Announcement
@zhenzhangzz says a 256-token soft prompt matches full-parameter RL
Separately trained prompts stack without joint training and beat individual experts, according to @zhenzhangzz.
TLDR
Personalization without the weight-swapping bill is the appeal for @GenAI_is_real, who highlights reported throughput within 1.5% of the bare model. @zhenzhangzz says separately trained prompts can stack without joint training, with almost no inference overhead. @GenAI_is_real still wants scaling evidence: longer prefixes consume context, and more prompts could interfere.
Combined views
890.1K
4 Sources, first seen ago

Comments