We introduce Gacha Decoding, an inference-time method for eliciting diverse language model generations that scales with model capability. The method treats diversity as an instruction-following problem, combining a language model with randomness from an external RNG to identify and realize distinct response modes.
@article{geng2026gacha,title={Gacha Decoding: Eliciting Diverse Generations Through Instruction Following},author={Geng, Scott and Zhang, Yufei and Lee, Joseph and Li, Jerry and Ghazvininejad, Marjan and Koh, Pang Wei},journal={arXiv preprint arXiv:2610.01382},year={2026},eprint={2610.01382},archiveprefix={arXiv},primaryclass={cs.AI},}