1: \begin{abstract}

2: % Energy-Based Models (EBMs) are a class of models that produce unnormalized non-negative scores for every element in their domain.

3: % While these models are more expressive than normalized models, this expressivity comes at the cost of giving up the possiblity of direct sampling.

4: % In this work, we introduce Quasi Rejection Sampling (QRS), a Monte-Carlo (MC) approach that makes it possible to obtain high-quality samples from an EBM as long as a reasonable proposal distribution is available.

5: % In contrast to well-known Markov Chain Monte-Carlo (MCMC) techniques, this method provides explicit convergence bounds and diagnostics allowing to judge the quality of the approximation at any given level of efficiency.

6: % Using Controlled Natural Language Generation as an experimental domain, we apply QRS to obtain samples from an EBM that imposes a set of distributional constraints on the generations of a pre-trained language model.

7: % We experiment both with training a proposal distribution and with using pre-trained language models with hand-crafted prompts as a proposal distribution.

8: % Our results show that we can approximate the target EBM to an arbitrary precision at the cost of some sampling efficiency, leading to a quality/efficiency trade-off that can be tuned according to the requirements of the desired application.

9: % \end{abstract}

10: