KAN author argues training-curve prediction fits tabular foundation models like TabPFN, not LLMs
ZimingLiu11 · x · 2026-10-31
Ziming Liu (KAN author) critiques LLM-based training-curve prediction in a thread. He notes that NCPL (neural configuration-to-performance scaling law) by Zhang et al. is clever but requires LLM fine-tuning and fails to generalize to OOD scenarios — e.g., it struggles when architectures like MoE are out of vocabulary.
He argues training-curve prediction is fundamentally more of a scientific problem (like predicting forces and energy) than a language prediction problem. Follow-up finding: a pre-trained TabPFN, with zero extra training, already serves as a strong meta model, outperforming NCPL and generalizing to OOD settings.
More from Models
- Rumor: OpenAI delayed new model over alignment issues, more than any other lab — haider1 · 2026-10-31
- ARC-AGI Kaggle competitions hit major slowdowns as ARC Prize refuses deadline extension — JFPuget · 2026-10-31
- Speech model taxonomy: native S2S and full-duplex are independent dimensions, not exclusive categories — rdesh26 · 2026-10-31
- Pre-registered training-curve forecast hits: TabPFN covers 14/14 later checkpoints — ZimingLiu11 · 2026-10-31
- Code-switched questions hurt LLM accuracy; CANVAS steering fixes internal representations — burkov · 2026-10-31
- OpenAI reportedly to train on users' private sessions with no opt-out — evilsocket · 2026-10-31