Announcement
Gacha Decoding is claimed to beat prior work by over 11x in sample efficiency for diverse AI responses
A post introducing the method says it uses instruction following and external random-number generation, not token entropy.
TLDR
The Gacha Decoding announcement claims the method produces varied language-model responses with over 11x the sample efficiency of prior work. It says the approach draws on instruction following and external random-number generation rather than token entropy, and suggests using it to generate varied ideas, data and reinforcement-learning environments.
Combined views
1.8K
4 Sources, first seen ago
