Pipelined speculative decoders for looped Transformers (LoopSpec, CopyLoop, LoopSkim) run positions at different loop depths in one batched call. Whether they r
Pipelined speculative decoders for looped Transformers batch one request's positions across loop depths, so reproducing autoregressive (AR) output is a batch-in
Pipelined decoding overlaps recurrent computation across token positions, but predictions made behind incorrect speculative tokens are usually absent from a ref
Exact pipelined decoders for looped Transformers, such as LoopSkim, commit a token only at full depth. Oracle drafts (the reference output) are often simulated
Which draft source can speed up exact pipelined decoding of looped Transformers on math, and what limits it? We test this on Ouro-1.4B and Raven-TinyLlama, at b