Researchers propose DeLS-Spec, a method that treats the fixed DFlash model as a long-context expert while introducing a lightweight local head as a short-context expert. This approach allows the local head to be trained independently with a standard next-token prediction objective, avoiding the need for joint training with the target model or the DFlash backbone.
- DeLS-Spec combines long-context and short-context logits at inference time without tying the local head to a specific DFlash checkpoint.
- The method achieves extremely low training costs by decoupling the drafting process.
- Experiments on Qwen3 models demonstrate consistent improvements in speedup and average acceptance length across math, code, and dialogue benchmarks compared to DFlash.
This modular design offers greater flexibility and efficiency for speculative decoding without requiring expensive retraining of draft models.