Researchers introduce Constrained Adaptive Rejection Sampling (CARS), a novel technique that improves the efficiency of generating constrained outputs from language models while maintaining distributional fidelity. The method adaptively prunes invalid continuations using a trie data structure, achieving higher sample validity rates without sacrificing output diversity.
CARS addresses a fundamental tension in constrained language model generation: existing methods either enforce validity at the cost of distributing output probability incorrectly, or preserve the model's true distribution while wasting computation on invalid samples. This research demonstrates a middle path through adaptive pruning, where the system learns which constraint-violating paths to avoid and progressively improves sampling efficiency across iterations.
The technical innovation lies in using a trie structure to record constraint violations and subtract their probability mass from subsequent sampling attempts. This ensures previously-explored invalid prefixes are never revisited, improving acceptance rates monotonically while maintaining exact fidelity to the constrained distribution. The approach emerged from practical demands in domains like program fuzzing and molecular generation, where both correctness and sample diversity directly impact downstream applications.
For AI developers and researchers, CARS reduces computational overhead in constrained generation tasks, enabling faster experimentation and deployment of language models in safety-critical applications. Program synthesis tools, code generation systems, and scientific workflows benefit from fewer wasted forward passes while maintaining output quality. The consistency across experimental domains suggests broad applicability rather than narrow optimization.
Looking ahead, this work may influence how language models integrate constraints at scale. As constrained generation becomes more prevalent in production systems, methods that improve efficiency without compromising distribution quality become increasingly valuable. The research opens questions about whether adaptive pruning generalizes to hierarchical or context-dependent constraints, and whether similar principles apply to other sampling-based inference techniques.
- →CARS improves rejection sampling efficiency by adaptively removing constraint-violating continuations using a trie data structure
- →The method maintains exact fidelity to the constrained distribution unlike greedy decoding approaches that distort probabilities
- →Acceptance rates improve monotonically as the trie learns invalid paths, reducing wasted computation over time
- →Experiments show CARS achieves higher efficiency and stronger sample diversity than existing constrained generation methods
- →The technique addresses critical needs in program fuzzing and molecular generation where validity and diversity are both essential