MARS: Enabling Autoregressive Models Multi-Token Generation

Paper: MARS: Enabling Autoregressive Models Multi-Token Generation (2604.07023)
Published: 8 Apr 2026.

Learn more on Emergent Mind: https://www.emergentmind.com/papers/2604.07023
arXiv: https://arxiv.org/abs/2604.07023
Sign up for our free trending papers email digest: https://www.emergentmind.com/subscribe
Follow us on X: https://x.com/EmergentMind
Join our Discord: https://discord.gg/BhfTC4mTXq

This presentation explores MARS, a breakthrough fine-tuning approach that enables standard autoregressive language models to generate multiple tokens per decoding step without architectural changes or extra parameters. We examine how MARS preserves strict autoregressive compatibility while achieving up to 1.7× speedup, the critical role of dual-stream training with SFT loss in maintaining quality at scale, and the practical implications of dynamically tunable speed-quality tradeoffs for production deployment.

Видео MARS: Enabling Autoregressive Models Multi-Token Generation канала Emergent Mind