Announcement
Gacha Decoding is claimed to diversify language-model responses with over 11x the sample efficiency of prior work
The author says the method uses instruction following and external random-number generation, rather than token entropy, to vary responses.
TLDR
The author introducing Gacha Decoding says it produces more varied language-model responses and beats prior work by more than 11x in sample efficiency. The post credits instruction following and an external random-number generator, rather than token entropy, for the diversity, and suggests using it to generate ideas, data and reinforcement-learning environments.
Combined views
5.8K
3 Sources, first seen ago
