Author(s): “The AI Engineer”

Originally published on Towards AI.

Subtitle

You asked the model for JSON. You wrote “return valid JSON only” in capital letters. You added an example. You added a second example. For 99 calls out of 100, it worked.

After the lead, the article explains the difference between prompting and constrained decoding: prompts shift probabilities via context but cannot force invalid tokens to zero, while constrained decoding applies a token mask to eliminate choices that would break a formal rule (e.g., a grammar or JSON schema). It walks through a simple masking mechanism and discusses real-world implementation challenges—especially that building the allowed-token mask is the hard, sometimes expensive part—highlighting research on faster preprocessing and GPU-friendly mask generation. The piece warns that “valid” doesn’t mean “likely” or “true,” since greedy local masking can distort the model’s distribution and may reduce output quality or reasoning performance under strict constraints; it mentions proposed fixes like ASAP and evidence from studies comparing constraint methods. Finally, it recommends practical patterns (e.g., think freely then constrain extraction, adjust strictness to the task), cautions against misleading metrics that become trivially green, and closes with guidance on when constrained decoding is worth using and how to interpret its guarantees.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI