TreeSpark builds on DSpark, a semi-autoregressive drafter that combines parallel block prediction with lightweight sequential correction.
Tree-trained draft branches make speculative decoding accept longer continuations
TreeSpark trains a parallel drafter on the same candidate-tree construction used at inference, improving accepted length and throughput on two Qwen3 models.
Chinese Tech
Jihao Zhao · Qianyu Jiang · Shuaishuai Zu · Pengfei Li · Guanghuan Fang · Ben Feng · +3 more
Renmin University of China · National Key Lab for Novel Software Technology, Nanjing University · Huawei Technologies Ltd.
Research Digest··3 min read
Zhao and colleagues extend speculative decoding, where a small drafter proposes tokens that a larger language model verifies, from a single continuation to a budgeted tree of alternatives.
Why this paper
From Huawei Technologies Ltd. and 2 others
In one line
TreeSpark improves speculative decoding throughput by 15-25% over DSpark by training drafters on model-generated parent states with tree-structured candidates.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§