Search papers, labs, and topics across Lattice.
This paper introduces Decoupled Contrastive Decoding (DCD), a method that optimizes the Contrastive Decoding (CD) process by utilizing an expert-aligned lightweight proposer for drafting while relegating the amateur model to verification tasks. The study reveals that contrastive-aware drafting does not consistently outperform expert-aligned drafting due to the weaker contrastive signal compared to drafter errors, which can be exacerbated during reconstruction. DCD achieves significant speed improvements, with average greedy speedups of 1.65 to 1.95 times over vanilla CD and a reduction in MMLU proposal-path latency by 5 to 12 times compared to traditional amateur-coupled approaches.
Expert-aligned drafting outperforms contrastive-aware methods, leading to up to 12x faster proposal paths in decoding.
Contrastive Decoding (CD) improves generation quality, but its amateur-model pass makes decoding expensive. Accelerating CD with speculative decoding raises a proposal-alignment question: should the contrastive signal shape the drafter, or should it remain only in verification? We study this question in the lightweight feature-level drafter regime. Two controlled diagnostics, matched Cross-alpha training and an Approximate Dual-Drafter decomposition, give the same diagnosis: contrastive-aware drafting does not consistently improve over expert-aligned drafting because the contrastive correction is usually weaker than drafter error, and reconstruction can amplify that error. We introduce Decoupled Contrastive Decoding (DCD), which drafts with an expert-aligned lightweight proposer and applies the amateur only in unchanged CD verification. Standard speculative verification preserves the vanilla-CD output distribution. Across the main 8B settings, EAGLE3-based DCD achieves average greedy speedups of 1.65 to 1.95x over vanilla CD and reduces MMLU proposal-path latency by about 5 to 12x relative to amateur-coupled proposal paths.
