relm (regular expression language modeling)

**RELM (Regular Expression Language Modeling)** is a structured generation technique that constrains LLM output to match specified **regular expression patterns**. It bridges the gap between the flexibility of free-form language generation and the precision of formal pattern specifications. **How RELM Works** - **Pattern Specification**: The user provides a **regex pattern** that the output must conform to (e.g., `\\d{3}-\\d{4}` for a phone number format, or `(yes|no|maybe)` for constrained choices). - **Token-Level Masking**: At each generation step, RELM computes which tokens are **valid continuations** according to the regex and masks out all others before sampling. - **Finite Automaton**: Internally converts the regex to a **deterministic finite automaton (DFA)** and tracks the current state during generation, only allowing tokens that lead to valid transitions. **Key Benefits** - **Guaranteed Format Compliance**: Output is mathematically guaranteed to match the pattern — no post-processing or retries needed. - **Flexible Patterns**: Regular expressions can specify everything from simple enumerations to complex structured formats. - **Composability**: Can combine multiple regex constraints for different parts of the output. **Limitations** - **Regex Expressiveness**: Regular expressions cannot capture all useful formats — they can't express recursive structures like nested JSON. For those, **context-free grammar (CFG)** constraints are needed. - **Quality Trade-Off**: Heavy constraints can force the model into unnatural text that, while format-compliant, may lack coherence. - **Token-Boundary Issues**: Regex patterns operate on characters, but LLMs generate **tokens** (which may span multiple characters), requiring careful handling of partial matches. **Relation to Broader Structured Generation** RELM is part of a larger family of **constrained decoding** techniques including **grammar-based sampling** (using CFGs), **JSON mode**, and **type-constrained decoding**. Libraries like **Outlines** and **Guidance** implement RELM-style regex constraints alongside more powerful grammar-based approaches.