On Decision-Valued Maps and Representational Dependence
Your AI's output might depend entirely on how you format the input data.
Deep Dive
A new paper introduces "decision-valued maps," showing that an AI model can produce different discrete outcomes based solely on how the same underlying data is represented or formatted. The research formalizes this representational dependence and describes DecisionDB, an infrastructure that logs and audits these relationships to ensure deterministic replay. This reveals that representation space has "persistence regions" where outputs are stable and "boundaries" where small formatting changes flip the result.
Why It Matters
This exposes a critical, often hidden source of unreliability in AI systems, challenging assumptions about deterministic outputs.