Nature Comms paper links diffusion model generativity to associative memory's spurious states
DimaKrotov · x · 2026-11-05
A new Nature Communications paper by Dima Krotov's team explores how generative behavior in diffusion models emerges from associative memory, identifying three phases of generated samples:
- Memorization: reproducing training examples verbatim
- Spurious states: discovering new patterns between stored memories while keeping a well-defined basin of attraction — these appear near the onset of generalization as data grows
- Generalization: generating genuinely novel samples with a flat energy landscape around them
The surprising twist: spurious states, traditionally seen as a flaw of associative memory, may actually be the first sign of generative capability.
More from Research
- Open-source segmenter tracks a breathing tumor across 10 CT phases on one GPU — MaziyarPanahi · 2026-11-05
- Mathematician: AI can grind lemmas, but 100-page arXiv proofs must ship with full Lean formalization — LucaAmb · 2026-11-05
- Tumor moves 11.3mm per breath; he tracked it across 10 CT phases with one open model on one GPU — MaziyarPanahi · 2026-11-05
- Bregman vantage point trees: VP-trees using Bregman balls for fast nearest-neighbor queries — FrnkNlsn · 2026-11-05
- Pre-print: iterated creation, grading, and revision of data science projects by language models — cjmaddison · 2026-11-05
- Latent Sparse Attention on Apple Silicon: 31% faster TTFT at 128k with a hybrid prefill/decode split — GabGarrett · 2026-11-05