Switching a multimodal model into non-thinking mode saves latency, but this Tencent paper finds it also makes user-visible response failures much more likely.

PatternEval scores the shape of an answer rather than its correctness, and finds that non-thinking inference breaks user-facing response quality far more often.

PatternEval is a 2,415-prompt multimodal benchmark that scores 4 of them: chain-of-thought leakage, repetition, logical contradiction, and performative reasoning.

That catches quality problems a correctness check never sees.

If you ship a fast mode, its answers need their own eval.