The Point: JSON schema constraints compile tool-call tokens into unreachable regions of token space, causing models to suppress function calls despite both functions working in isolation.
Multiple open-weight Large Language Models suppress tool calls when JSON schema constraints for structured outputs are simultaneously enabled. This phenomenon reveals a critical gap in the reliability assessment of agent systems.
Researchers have documented a reproducible phenomenon in production agent systems: when tool calling and JSON schema constraints are simultaneously activated, multiple open-weight models fail to recognize the need for function calls — although both capabilities remain functional when tested separately. The working group termed this behavior “tool suppression”.
The cause lies in the technical implementation of schema constraints: JSON schema restrictions are encoded via grammatical token masks. These masks make tool-call tokens unreachable during the decoding phase by setting their probability space to zero. Controlled experiments across multiple model families reproduced the behavior consistently, while the models demonstrated both correct schema compliance and tool execution when evaluated independently.
The researchers formulated the “Constraint Priority Inversion” (CPI) hypothesis, proposing that under multiple simultaneous constraints, schema management dominates action-selection behavior. They presented this as a behavioral hypothesis rather than a verified internal mechanism. To remediate, they propose “Transparent Two-Pass Execution”: an inference strategy that decouples tool execution from schema-anchored response generation. Tests show function restoration while maintaining structured output guarantees — without requiring model retraining.
The study emphasizes that isolated tests of tool use and structured output can overlook significant reliability issues in production agent systems. Code, data, and documentation will be provided on GitHub.
Source: arxiv.org · Published June 23, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.