EMNLP paper: diffusion models form stable attractors for never-seen test samples
LucaAmb · x · 2026-09-08
An EMNLP 2026 main-track paper by Dima Krotov and collaborators studies how memory and novelty (creativity) coexist in language diffusion models (LDMs):
- Beyond memorizing training data, the diffusion dynamics can error-correct perturbations of test samples it never saw during training
- As training set size grows, unseen test samples also become stable attractors of the generative dynamics; token recovery improves on unseen examples while degrading on training ones
- The behavior closely mirrors Dense Associative Memory models as memory count increases, suggesting a unified dynamical account of memorization vs. generalization
- Collaboration between IBM and RPI; project page is public
Related event: EMNLP Paper Shows Language Diffusion Models Are Associative Memories(2 posts)→
More from Multimodal
- Beating Img2Video ghosting: 3 chained 4-second clips beat one 12-second render, twice as fast — Key_Education4018 · 2026-09-11
- Face Capture System v3 Shows Signs of Life — andrew_n_carr · 2026-09-11
- Single-Word Midjourney Prompt Series Uses Scots Word 'Smeddum' for Striking Imagery — tisch_eins · 2026-09-11
- One-shot animation with GPT-6 Astra wows users as model demos keep impressing — paw_lean · 2026-09-11
- The prompt behind that 75M-view AI volcano video is now out — charis_ai · 2026-09-11
- GPT-6 Astra + Magnific MCP: directing motion-graphics videos via conversation — charis_ai · 2026-09-11