ICE-Guard Detects Spurious Feature Reliance Across Domains

Researchers introduce ICE-Guard (Mar 19, 2026) to detect spurious feature reliance in LLMs using intervention consistency testing. They evaluate 11 models across 3,000 vignettes in 10 high-stakes domains, finding authority (5.8%) and framing (5.0%) biases exceed demographic bias (2.2%), and show mitigations reduce flips up to 100% and cut bias 78% cumulatively.