Gemma 4's Architectural Divide: How MoE and Dense Models React Oppositely to Identical Prompts

When Gemma 4's mixture-of-experts and dense variants received identical prompt instructions, they diverged sharply—MoE improved toward grounded answers while the dense model became overly cautious with false-negative refusals. This architecture-dependent response pattern suggests that prompt tuning effects are shaped by model design, not size alone. For production systems, the finding implies that architecture-aware prompting strategies matter as much as model selection itself.