Efficient decoding · Lossless verification
Decoupled Contrastive Decoding via Expert-Aligned Drafting
DCD removes the amateur model from the serial proposal path, then restores the exact contrastive-decoding target during parallel verification.
Research question
Can contrastive decoding keep its target distribution while eliminating the expensive serial amateur proposal?
Correctness
Verification must reproduce the original contrastive distribution—not an approximation to it.
Latency
Proposal latency sits on the critical path and directly limits wall-clock throughput.
Systems
The runtime must coordinate expert, amateur, and drafter states without cache divergence.
Core method
rt(x) ∝ pt(x)1+α / qt(x)αThe proposer changes efficiency and acceptance—not the target distribution.
Three-model state
Independent KV caches, token pools, batch lifecycle, and draft/verify synchronization.
Proposal error vs. CD signal
Cross-alpha, dual-drafter, and stronger-proposer controls isolate why contrastive-aware drafting does not reliably help.
Latency decomposition
Separates proposal, expert verification, and amateur verification instead of relying on accepted length alone.
Key results
Average greedy speedup over vanilla contrastive decoding.
Per-step proposal latency on MMLU.
Average speedup at α = 0.1.
DCD can achieve higher wall-clock throughput even when accepted length is not the largest, because it removes the expensive serial proposal pass.
Paper and code
Public arXiv v1 and code available · 13 August 2026 · 28 pages · 11 figures · 20 tables · DOI