Deer Diffusion Speculative Decoding

Skill Summary

DEER (Draft with diffusion, vErify with autoRegressive) introduces speculative decoding using discrete diffusion language models for efficient parallel draft generation. The approach trains dLLMs through two-stage alignment: AR-style distillation enabling prefix-conditioned continuation and scribe refinement sharpening predictions near verification boundaries. Unlike AR drafters suffering left-to-right uncertainty accumulation, DEER's parallel generation makes "the proposal at position i independent of previously drafted tokens," enabling acceptance lengths up to 32 tokens vs. ~10 for competing methods. Results show 5.54× speedup on HumanEval.

When To Use

• Building efficient LLM inference systems requiring high acceptance rates
• Scenarios where speculative decoding speedups are critical for latency
• Projects where parallel draft generation prevents error accumulation
• Research on efficient alternatives to autoregressive drafting

When NOT To Use

• Applications already meeting latency goals with simpler methods
• Scenarios where training discrete diffusion models is computationally infeasible
• Domains where AR drafters' lower acceptance rates aren't performance bottlenecks
• Models with strict parameter budgets for additional draft model

Core Technique

Two-stage alignment pipeline trains diffusion drafters:

1. AR-Style Distillation (Stage I) Train discrete diffusion language model (dLLM) for prefix-conditioned continuation. Learn from truncated teacher answers marked with SEP token, enabling model to generate coherent suffixes given fixed prefix. This bridges AR teacher and diffusion student paradigms.

2. Scribe Refinement (Stage II) Enhance accuracy near verification boundary through "weighted suffix masking with exponentially decaying loss," focusing training on tokens most critical for acceptance. Improve predictions most likely to affect speculative decoding outcome.

3. Parallel Generation Advantage Unlike AR drafters with left-to-right uncertainty accumulation, dLLM generates entire token blocks in single denoising step. "The proposal at position i is independent of previously drafted tokens," preventing error propagation and enabling acceptance lengths up to 32 tokens versus ~10 for competing methods.

Speedup Results

• 5.54× speedup on HumanEval vs. EAGLE-3's 2.41×
• Substantially better than AR-based speculative decoding approaches
• Maintains competitive verification overhead

Implementation Notes

Start with pretrained teacher AR model. Train discrete diffusion LM for parallel generation conditioned on prefixes. Implement AR-style distillation with SEP token marking. Add scribe refinement focusing on boundary tokens. Integrate into verifier framework for speculative decoding. Measure acceptance rates and overall speedup.

References

• Original paper: DEER (Dec 2025)
• Discrete diffusion language models
• Speculative decoding frameworks