You are a careful technical writer. Your task is to summarize the following research note for an audience of senior software engineers who are short on time. Produce a summary that meets every constraint listed below.

Constraints:
1. The summary must be no longer than 180 words.
2. Open with a single-sentence "headline" that states the main result.
3. Follow the headline with three to five bullet points covering: the problem the work addresses, the approach taken, the experimental setup, the headline numerical result, and one notable caveat or limitation.
4. Use plain prose. Avoid marketing language, hedging adverbs, and rhetorical questions.
5. Do not introduce facts that are not present in the source note.

Research note to summarize:

We investigate whether speculative decoding still pays off when the draft model and the verifier model share weights via Multi-Token Prediction (MTP) heads, as opposed to using a separately trained draft model. On three open-weight verifier models in the 1B to 9B parameter range, we measure end-to-end decoding throughput across four prompt distributions: synthetic random tokens, English Wikipedia paragraphs, Python source code, and multi-turn conversational data. We find that token acceptance rates depend heavily on prompt distribution: random-token prompts produce acceptance rates near 0.28 even at speculation depth 2, while coherent English text yields acceptance rates between 0.55 and 0.75 at the same depth. The throughput gap is large enough that a benchmark run on random tokens can incorrectly conclude that MTP slows down generation. The main caveat is that our measurements are taken at batch size 1 on a single GPU; the picture changes at higher batch sizes where verifier compute becomes the bottleneck.

Begin your summary now.
