itaigat
Abstract
Set Block Decoding accelerates language model generation by integrating next token prediction and masked token prediction, enabling parallel sampling of future tokens and reducing computational cost without sacrificing accuracy.
Autoregressive next token prediction language models offer powerful capabilities but face significant challenges in practical deployment due to the high computational and memory costs of inference, particularly during the decoding stage. We introduce Set Block Decoding (SBD), a simple and flexible paradigm that accelerates generation by integrating standard next token prediction (NTP) and masked token prediction (MATP) within a single architecture. SBD allows the model to sample multiple, not necessarily consecutive, future tokens in parallel, a key distinction from previous acceleration methods. This flexibility allows the use of advanced solvers from the discrete diffusion literature, offering significant speedups without sacrificing accuracy. SBD requires no architectural changes or extra training hyperparameters, maintains compatibility with exact KV-caching, and can be implemented by fine-tuning existing next token prediction models. By fine-tuning Llama-3.1 8B and Qwen-3 8B, we demonstrate that SBD enables a 3-5x reduction in the number of forward passes required for generation while achieving same performance as equivalent NTP training.
Community
We present a new paradigm called Set Block Decoding, combining next-token-prediction and masked (or discrete diffusion) models, allowing parallel decoding (x3-5 speedup) without any architectural changes and with exact KV cache. Matches NTP performance!
Absolutely phenomenal results, folks - congratulations, and thank you both for this amazing work and sharing it with the world!
One quip: Figure 1 is a data vizualization chart crime. The horizontal axis starting at 1 means that you can't actually visually compare SBD against NTP — should start at 0 instead. Additionally, "Speedup" is a rather poor choice for the horizontal axis label, as the word can be interpreted as delta improvement — a better choice being something like "Speed Factor" or "Speed wrpt/NTP", etc.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Your LLM Knows the Future: Uncovering Its Multi-Token Prediction Potential (2025)
- OverFill: Two-Stage Models for Efficient Language Model Decoding (2025)
- Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction (2025)
- Wide-In, Narrow-Out: Revokable Decoding for Efficient and Effective DLLMs (2025)
- Efficient Speculative Decoding for Llama at Scale: Challenges and Solutions (2025)
- DPad: Efficient Diffusion Language Models with Suffix Dropout (2025)
- CARD: Cache-Assisted Parallel Speculative Decoding for Efficient Large Language Model Inference (2025)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Great work
Great work! Will the training and inference code be open-sourced?
Is there a SBD fine-tuned llama-3b available to download? As presented in the paper?
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper