Schema Key Wording as Implicit Instruction in Structured LLM Generation

A recent study published on arXiv (2604.14862) thoroughly examines the role of schema keys in constrained decoding as an implicit channel for instructions within large language models. The researchers present structured generation as a multi-channel instruction challenge, indicating that task signals may be found in prompts, schema keys, or a combination of both. They propose a projection-aware analysis revealing that a chain-of-thought-style key is beneficial only when its semantic advantage surpasses the distortion caused by grammar-constrained projection, clarifying the model-dependent effects of keys. Tests conducted on mathematical reasoning benchmarks indicate that simply changing the wording of schema keys can greatly influence accuracy.