This research develops a novel importance sampling algorithm for constrained decoding that eliminates biases while improving computational efficiency.

• Enables controlled text generation for product selection, safety compliance, and specific formatting
• Achieves asymptotic unbiasedness while maintaining computational efficiency
• Delivers superior performance compared to existing prefix-tree methods
• Provides practical solutions for safety-compliant content generation

Why it matters: Security teams can implement more efficient constraints to ensure LLMs generate content that complies with safety standards without sacrificing performance or introducing unintended biases.

Efficient and Asymptotically Unbiased Constrained Decoding for Large Language Models