๊ฐ€์ž… ํ›„ ์ดˆ๋Œ€ ๋งํฌ๋ฅผ ๊ณต์œ ํ•˜๋ฉด ๋™์˜์ƒ ์žฌ์ƒ ๋ฐ ์ดˆ๋Œ€ ๋ณด์ƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Zhen Zhang
@zhenzhangzz
PhD @UCSB, BS @TsinghuaNLP. Post training, Efficiency. Prev intern @ByteDanceSeed_ RL scaling team
๊ฐ€์ž… June 2016
570 ํŒ”๋กœ์ž‰ ์ค‘    594 ํŒฌ
We think weโ€™ve found a better path to continual post-training: merging models in input space. ๐Ÿ”น A 256-token soft prompt (~0.01% of params) matches full-parameter RL ๐Ÿ”น Plug-and-play composition: prompts trained separately just stack, with no joint training, and the combo beats the experts. No more MOPD! ๐Ÿ”น Better OOD generalization, almost no forgetting ๐Ÿ”น ~0 inference overhead, even with a different prompt per request. Built for personalization.
๋” ๋ณด๊ธฐ