Gacha Decoding Increases Generation Diversity via Instruction Following

Gacha Decoding treats diversity as an instruction-following task rather than relying on token entropy. This inference-time method achieves up to 2.4x Vendi diversity and requires 11x fewer samples than prior approaches to reach the same number of high-quality modes in domains like protein design.