Structured output (JSON schema) for LLM generation #1528

Problem description

LLM generation cannot be forced to match a JSON schema, so apps put the schema in the prompt. With Llama 3.2 1B that gave 2/20 valid outputs (review).

Proposed solution

Grammar-constrained decoding (mask logits per step, e.g. via xgrammar), exposed as responseFormat: { type: 'json_schema', schema } on generate / sendMessage.

Alternative solutions

Prompt + validate + retry: what users do today, unreliable on small models.

Benefits to React Native ExecuTorch

Reliable structured output for extraction, classification and tool calling.

Additional context

Blocked on the LLM runner rewrite; the current runner uses upstream's sampler.