Traditional structured generation treats the task as a deterministic mapping from input to a single structured output. In natural language, however, many structurally distinct yet semantically valid interpretations often exist. Single‑reference supervision collapses this ambiguity into one prediction, causing unstable and inconsistent outputs when the input is underspecified.
We reformulate structured generation as an \textit{ambiguity‑aware inference} problem: the goal is to reason over a set of plausible structured outputs rather than to predict a single solution. Standard training unintentionally suppresses valid alternatives, so we adopt a multi‑valid learning framework that preserves ambiguity during optimization.
To operationalize this, we build a candidate space of structured outputs and represent each candidate as a graph encoding entities and their relationships. Selection is performed via graph‑based reasoning that evaluates semantic alignment and relational consistency, enabling principled comparison of competing interpretations beyond token‑level likelihoods.
Experiments on several structured generation benchmarks show consistent gains in accuracy, stability, and robustness across varying ambiguity levels. These findings highlight the importance of explicitly modeling ambiguity and suggest a shift from single‑output prediction to set‑level reasoning in structured generation.
Review