Distilled influence embeddings make diffusion training data attribution a nearest-neighbor lookup
serrjoa · x · 2026-10-09
Shixuan Liu and colleagues improved a gradient-based training data attribution method, made it amenable to online calculation, and distilled it with a ranking loss into "influence embeddings"—so accurate TDA becomes a simple nearest-neighbor lookup in a relatively low-dimensional space.
Related event: Sony's TIDE Speeds Up Training Data Attribution(2 posts)→
More from Research
- DeepMind essay: what 15M Gemini interactions reveal about AI-accelerated science — JMateosGarcia · 2026-10-09
- ENPIRE turns coding agents into autonomous robotics researchers, hitting 99% success on dexterous tasks — chris_j_paxton · 2026-10-09
- ICLR 2027 asks ACs to catch reviewer conflicts they have no way to see — shaohua0116 · 2026-10-09
- Dietterich: treat OpenAI's math models like a deceased mathematician's notebook — tdietterich · 2026-10-09
- Paper: Naming a Fine Makes AI Agents Less Compliant, 46-Point Spread Across Models — niloofar_mire · 2026-10-09
- Odyssey-3 Tested: Frozen Backbone Trained on 20 Hours of Driving Data Transfers — testingcatalog · 2026-10-09