Search papers, labs, and topics across Lattice.
The authors introduce CIPHER, an expert-validated benchmark spanning consumer-finance, clinical, and law-enforcement records to evaluate cross-record reasoning over privacy-hardened hybrid data (structured tables combined with unstructured text) using executable SQL supervision. Evaluating modern retrieval, prompting, table-specialist, and neuro-symbolic systems under native redaction and surrogate-based restoration reveals severe performance degradation across all model families, even when supporting records are provided. Analysis shows that these failures stem primarily from errors in record selection and predicate interpretation rather than arithmetic computation, with privacy transformations exerting non-uniform effects that alternately obscure critical evidence or inadvertently filter distractors.
Standard privacy redactions induce unpredictable failures in LLM reasoning across enterprise and clinical data, causing models to break down on record selection and predicate interpretation rather than computational execution.
Reasoning over privacy-constrained records requires combining structured attributes with evidence from free-text narratives. We introduce CIPHER (Cross-record Inference over Privacy-Hardened Evidence Records), a benchmark of expert-validated questions from consumer-finance, clinical, and law-enforcement records. The questions cover common tabular operations and include executable SQL supervision. We evaluate retrieval, prompting, table-specialist, and hybrid symbolic-neural systems under native redaction and surrogate-based evidence restoration. All system families exhibit substantial failures even when supporting records are provided. Most errors arise from incorrect record selection and predicate interpretation rather than arithmetic execution. Privacy transformations have non-uniform effects, sometimes obscuring necessary evidence and sometimes reducing distraction. CIPHER provides a reproducible testbed for diagnosing these failures and assessing how transformations of sensitive text affect reasoning over hybrid records.
