LoopSpec: Pipelined Self-Speculative Decoding for Looped Transformers

简体中文

L1 mechanism diagnostic. Reference depth scheduling, not paper GPU throughput.

Paper information

Field	Value
Paper	LoopSpec: Pipelined Self-Speculative Decoding for Looped Transformers
Affiliations	Seoul National University / KAIST
First public release	2026-09-15 (arXiv v1)
Authors' code	No public author repository found as of 2026-09-16.
Adapter / method	loopspec
Local code	src/auto_research/foundation_latest_20260916.py

Original paper

Background and contribution

Use intermediate predictions across recurrent depth for self-speculation. Pipeline residual proposals to hide verification latency.

flowchart LR
  I[Looped model state] --> P[Intermediate draft] --> R[Pipelined residual proposal] --> V[Verification]

Key figure from the paper

Original Figure 1: self-speculation workflow. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

The reference preserves defining gate, weighting, and state transitions, exposing auditable scheduling/depth intermediates. Fixed-budget three-seed tests check invariants, not paper-scale training.

Paper results

Paper capability, cost, and speed are not local evidence or production gains.

Local reproduction

Three-seed results are diagnostic_only=true and not treated as full capability gains.

Reproduction limits

Only the reference depth scheduler executes; no reproduction of paper GPU throughput.