The authors designed PRISM with two components: (i) submodular curation to construct training datasets where each prompt is paired with multiple correct but semantically diverse solutions from various teacher models, and (ii) stochastic embedding modulation, which injects Gaussian noise into a small set of special-token embeddings during training and inference.
Training-time framework teaches LLMs to sample diverse outputs via learned noise signals
PRISM combines submodular data curation with stochastic embedding modulation to make diversity a controllable model capability
Big Tech
Qixin Xu · Zihan Zhu · Xiaojiang Liu · Bohan Lyu
Tsinghua University · Apple
Research Digest··2 min read
The authors propose PRISM, a training-time framework that internalizes output diversity in LLMs.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.