Planned Diffusion
This addresses the problem of slow text generation for users of large language models by offering a practical method to improve speed with minimal quality loss, though it is incremental as it combines existing paradigms.
The paper tackles the trade-off between generation speed and output quality in large language model inference by proposing planned diffusion, a hybrid method that uses an autoregressive plan to break output into spans generated in parallel via diffusion, achieving Pareto-optimal trade-offs with speedups of 1.27x to 1.81x and win rate drops of 0.87% to 5.4% on AlpacaEval.
A central challenge in large language model inference is the trade-off between generation speed and output quality. Autoregressive models produce high-quality text but generate tokens sequentially. Diffusion models can generate tokens in parallel but often need many iterations to match the same quality. We propose planned diffusion, a hybrid method that combines the strengths of both paradigms. Planned diffusion works in two stages: first, the model creates a short autoregressive plan that breaks the output into smaller, independent spans. Second, the model generates these spans simultaneously using diffusion. This approach expands the speed-quality Pareto frontier and provides a practical path to faster, high-quality text generation. On AlpacaEval, a suite of 805 instruction-following prompts, planned diffusion achieves Pareto-optimal trade-off between quality and latency, achieving 1.27x to 1.81x speedup over autoregressive generation with only 0.87\% to 5.4\% drop in win rate, respectively. Our sensitivity analysis shows that the planning mechanism of planned diffusion is minimal and reliable, and simple runtime knobs exist to provide flexible control of the quality-latency trade-off.