TempoPFN: synthetic pre-training of linear RNNs beats real-data models on GIFT-Eval
Yossarian_1234 · reddit · 2026-10-19
The TempoPFN paper introduces a library of synthetic time-series generators and augmentations used to pre-train a time-series foundation model (a linear RNN) for zero-shot forecasting.
- Method: fully synthetic pre-training via diverse generators and an augmentation pipeline (mixing, noise, quantization, scaling, NaN handling), no real data needed
- Results: on GIFT-Eval, synthetic-only training achieves 0.533 MASE, beating TabPFN-TS (0.544) and Chronos Bolt (0.574)
- Adoption: the synthetic data has already been used in Toto-2 and TS-ICL
- Venue: accepted at NeurIPS 2026
More from Research
- Researcher Aran Nayebi: consciousness paper shows sustained dominance, not 'flickering' — aran_nayebi · 2026-10-19
- H-CDLM brings DINO-style hierarchical semantics to continuous diffusion LMs for cheaper generation — sedielem · 2026-10-19
- Stanford professor argues AI should aid peer review, blind spots remain for human experts — anshulkundaje · 2026-10-19
- Doris Tsao's team invites open peer review on bioRxiv, bypassing journal timelines — anshulkundaje · 2026-10-19
- Amazon's Whole-Body Dexterity hits 90% success teaching Unitree G1 to haul boxes like humans — CyberRobooo · 2026-10-19
- Researcher: Openness uncorrelated with benchmark saturation, distinct from contamination — evijit · 2026-10-19