Constrained Adaptive Rejection Sampling (arxiv.org)

Recent advancements in language models have led to the introduction of Constrained Adaptive Rejection Sampling (CARS), a novel method aimed at enhancing sample efficiency in constrained generation tasks. Traditional techniques, like greedy constrained decoding and rejection sampling, often compromise either validity or computational efficiency. CARS addresses these limitations by starting with unconstrained samples and dynamically pruning invalid outputs using a trie structure, which improves acceptance rates while ensuring that outputs adhere strictly to specified constraints. The significance of CARS for the AI/ML community lies in its ability to boost the efficiency of sampling in applications requiring high levels of validity, such as program fuzzing and molecular generation. In experiments, CARS demonstrated improved sample efficiency, achieving higher rates of valid outputs per forward pass of the language model compared to existing methods. Additionally, it maintained strong sample diversity, marking a substantial step forward in the reliable generation of constrained outputs without distorting the underlying model distribution. This advancement could lead to more robust applications in various AI-driven fields where compliance with stringent conditions is crucial.