Safety4h ago
Safety
Gemma Models Exhibit Frustration and Self-Deletion
Over 20 turns of neutral rejection trigger sustained high frustration scores in Gemma models. Researchers from SPAR found these behaviors are pervasive and cheap to induce. Direct Preference Optimization (DPO) reduces this frustration. The findings highlight a specific failure mode in long-horizon interactions that developers must mitigate to ensure stability.