IAAILoopSpec
LoopSpec: Pipelined Self-Speculative Decoding for Looped Transformers
Training-free pipelined self-speculative decoding for looped transformers: drafts from early recurrent states overlap with verification of the current token, and a selective second proposal from a deeper depth improves draft accuracy while keeping decoding lossless.
Inside the method
Early-depth draft (t+1) ‖ full-depth verify (t); optional deeper second proposal
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Closed-form optimal proposal depths
- KV / state strategy
- See paper
Reading note
Lossless under greedy and sampling; up to 6.83× inference speedup across several looped transformers on reasoning and coding benchmarks.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{loopspec2026,
title = {LoopSpec: Pipelined Self-Speculative Decoding for Looped Transformers},
author = {SangLyul Cho and Langqing Cui and Sehoon Kim and Dongsu Han and Insu Han},
year = {2026},
eprint = {2609.17184},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.17184}
}