Adaptive Multi-Step Lookahead Decoding for Diffusion Language Models

Short summary

AdaLook is an adaptive lookahead decoding framework for masked diffusion language models that dynamically decides whether to continue rollout based on candidate-score variance and enables branch expansion at informative intermediate states. Unlike fixed-depth lookahead methods, it avoids unnecessary computation while improving accuracy across longer decoding trajectories. Experiments show it achieves a better accuracy-to-decoding-steps trade-off than existing one-step lookahead methods.

• •Proposes AdaLook, adaptive multi-step lookahead for diffusion language model decoding
• •Uses candidate-score variance to dynamically decide rollout depth and branch expansion
• •Outperforms one-step lookahead methods on accuracy-efficiency trade-off across benchmarks

Generated with AI, which can make mistakes.