Researchers have introduced On-Policy Residual Learning (ORL), a new method designed to significantly accelerate diffusion language models (dLLMs) without compromising generation quality. ORL achieves this by enabling dLLMs to refine multiple tokens in parallel, carrying forward and progressively refining previous predictions as residual context. This approach has demonstrated impressive speedups, reaching 2,788 tokens/s on a single RTX 5090 and 7,708 tokens/s on 8x B200 GPUs, while also improving benchmark scores like HumanEval pass@1. AI
IMPACT Accelerates diffusion LLM inference, potentially enabling faster and more efficient generation of content across various applications.
RANK_REASON The cluster describes a new research method for accelerating diffusion language models, including technical details and benchmark results. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
