Search papers, labs, and topics across Lattice.
This paper introduces Physics-Anchored Certification (PHACT), a novel framework that separates the proposal of physical designs from their certification, ensuring reliability by using a deterministic engine for validation. By deriving certified quantities from fixed inputs, the method eliminates the risk of forgery, achieving zero false certifications across eighty adversarial trials with varying models and conditions. This approach not only enhances the trustworthiness of designs generated by language models but also establishes a robust mechanism for certifying physical outputs across multiple scientific domains.
A groundbreaking certification method ensures that language models can generate reliable physical designs without the risk of forgery, achieving zero false certifications in adversarial conditions.
An unreliable language model can be made to produce reliable physical designs if the authority to assert is moved out of the model: the model proposes, and a deterministic engine alone certifies, returning certified, impossible, or unknown. We introduce Physics-Anchored Certification (PHACT), a propose-certify loop spanning five scientific domains, and identify what makes such a certificate trustworthy. A checker that accepts a model-supplied value can be forged; deriving the certified quantity from fixed inputs instead makes forgery impossible by construction. Across eighty adversarial trials spanning two models, two decoding temperatures, and a deliberately faulted engine, this contract produced zero false certifications.
