JevOut: Natural Context Can Flip Decision Models

TL;DR

Short, natural-sounding context additions can cause decision models like Jev to confidently choose incorrect options even when the correct answer remains unchanged. The study demonstrates that such context manipulations successfully flip decisions in over 60% of cases across multiple datasets and models.

Why It Matters

This reveals a critical vulnerability in AI systems that map language directly to actions, undermining trust in their probability outputs as reliable decision signals. Such fragility could lead to unsafe or unintended behaviors in real-world applications relying on automated decision-making.

Key Findings

• Natural context additions can redirect correct decisions to high-confidence wrong choices without altering the original question or answer.
• A context optimization method successfully induced target flips in 61.4% of initially correct decisions across 508 test cases.
• Three additional decision models showed similar vulnerability, with targeted flip rates between 64.9% and 73.2% on correct initial decisions.