Diffusion Forcing: Unifying Autoregressive and Diffusion Models #genai #llm #ai

5 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Jul 9, 2026

Diffusion Forcing: what happens when next-token prediction meets full-sequence diffusion.

There are two ways to generate a sequence, and each has a weak spot. Next-token models (LLMs) are causal and can run to any length, but over long rollouts small errors pile up and the output drifts. Full-sequence diffusion models (think Sora) denoise a whole fixed-length clip at once and can be guided toward good trajectories, but they cannot grow past that length or generate causally.

Diffusion Forcing, from MIT (NeurIPS 2024), unifies both with one simple move: train a diffusion model to denoise tokens that each carry their own independent noise level. Treat noise as a soft, fractional mask - zero noise means a token is fully observed, maximum noise means fully hidden, and everything in between is partial. One causal model now spans pure next-token prediction, full-sequence diffusion, and every mixture.

Why it matters:
- Training is a standard per-token denoising loss - easy to add.
- At sampling you pick any per-token schedule on a 2D grid (tokens by steps), keeping the future noisier while you commit the present, so long rollouts stay stable.
- It beat six baselines on 2D maze planning and carries latent memory for robot imitation. Its successor, the Diffusion Forcing Transformer (DFoT, ICML 2025), rolls out extremely long, consistent videos.

Reach for it on long-horizon video, planning, and control. For plain text, ordinary next-token models are still simpler and stronger.

Did this make Diffusion Forcing click? Follow for plain-English GenAI breakdowns.

#GenAI #DiffusionModels #MachineLearning #AIresearch #DeepLearning