Constrained decoding refers to techniques that restrict a language model’s token generation to outputs matching a predefined structure, such as a grammar, finite state machine, or schema. In this dataset it covers filings that combine a decoding-constraint mechanism, like logit masking or automata-based restrictions, with a validation or vocabulary-alignment step. Patent claims in this space typically focus on how the constraint is represented and enforced during generation, rather than on the underlying language model itself.

The ranked assignee list contains 12 companies, and one leader holds 9 of the 22 records in scope — more than triple the count of the next closest filers. The top 5 assignees combined account for 68.2% of all records, so the field is concentrated at the top with a long tail of single-filing entrants making up the rest of the ranking.

Filing activity rose from zero in 2017 to a peak of 8 records in 2022, then appears to taper in the years since. That apparent slowdown should be treated cautiously: publication typically lags filing by around 18 months, so 2025 and 2026 counts are understated and a genuine growth rate cannot yet be calculated from this data.