Particle-collision pretraining cuts cosmology training data to 100 simulations
bravo_abad · x · 2026-10-26
Mikuni and colleagues adapt a point-cloud transformer pretrained on a billion particle jets to cosmology, exploiting the shared representation of collision events and cosmological catalogs as unordered point clouds. With only 100 cosmological simulations, the model predicts matter clustering at R²≈0.87 versus 0.40 for training from scratch on the CAMELS-SAM benchmark — showing cross-domain pretraining can bypass expensive universe simulations.
More from Research
- First-order world-model gradients give minor training speedups over zero-order PG — breadli428 · 2026-10-26
- TACL editor Ehud Reiter: don't submit LLM-written papers, human review will reject them — EhudReiter · 2026-10-26
- DeepGHS previews AnimeCSIP-v3, an artist-style recognition model built on AnimeDINOv3 — bdsqlsz · 2026-10-26
- LeCun co-authored Temporal Straightening: straightening latent paths boosts gradient-based planning success rates — ylecun · 2026-10-26
- Alibaba's four-generation TTS evolution: CosyVoice to Qwen-Audio-3.0-TTS — solyarisoftware · 2026-10-26
- LoopUno: 52-Minute LoRA Training Makes 3B LLM Draft 2 Tokens per Step Losslessly — LucaAmb · 2026-10-26