ECHO: Early-layer Collaborative Hierarchical Orchestration with Bonus Logits in Speculative Decoding¶
L1 mechanism diagnostic. Candidate merging and lossless correction run locally; CUDA verifies the tensor path, not throughput.
Paper information¶
| Field | Value |
|---|---|
| Paper | ECHO: Early-layer Collaborative Hierarchical Orchestration with Bonus Logits in Speculative Decoding |
| Affiliation | First author: Wuhan University |
| First public release | 2026-09-15 (arXiv v1) |
| Authors' code | ECHO |
| Adapter / method | echo |
| Local code | foundation_latest_20260916_followup.py |
Original paper¶
Background and contribution¶
A fast early-layer loop explores while a lower-frequency final-layer loop verifies. Bonus logits on both sides fill out the tree while retaining exact rejection-sampling correction.
flowchart LR
I[Public input and state] --> M[ECHO mechanism] --> A[Auditable intermediates] --> O[Output or state update]Key figure from the paper¶
Original architecture from the paper; copyright belongs to the authors.
Mechanism and evidence¶
The reference preserves the defining selection/gating and state changes, emitting audit statistics. Three-seed fixtures test invariants and boundaries, not paper capability. Paper accuracy, speed, training loss, or benchmark conclusions are not local or online evidence.
Local reproduction¶
See three-seed mechanism results, marked diagnostic_only=true and excluded from capability rankings. NVIDIA A100 executed CUDA; the sanitized receipt omits hostnames, SSH aliases, and driver/build strings.
Reproduction limits¶
Candidate merging and lossless correction, not a replication of the paper's throughput.