BayesPO: Bayesian Prompt Optimization via Parallel-Tempered Gradient-Guided Discrete MCMC
The paper treats prompt tuning as posterior sampling over discrete tokens, not heuristic prompt search.
BayesPO combines a task likelihood with a language-model prior, then samples prompts from the resulting energy-based posterior. Its sampler uses Metropolis-Hastings corrected Gibbs-with-Langevin proposals plus parallel tempering to move through difficult LLM energy landscapes. In Qwen2.5 experiments, post-optimizing APE prompts across 24 instruction-induction subtasks raised average accuracy from 60.04% to 63.23%. The authors also report overfitting on small optimization sets and high compute cost as current limits. ArXiv · AI/CL/LG's note
BayesPO combines a task likelihood with a language-model prior, then samples prompts from the resulting energy-based posterior. Its sampler uses Metropolis-Hastings corrected Gibbs-with-Langevin proposals plus parallel tempering to move through difficult LLM energy landscapes. In Qwen2.5 experiments, post-optimizing APE prompts across 24 instruction-induction subtasks raised average accuracy from 60.04% to 63.23%. The authors also report overfitting on small optimization sets and high compute cost as current limits. ArXiv · AI/CL/LG's note
score 5