KAN author's 'meta model': predict training curves to pick hyperparams, then design architectures
ZimingLiu11 · x · 2026-10-31
Ziming Liu (KAN creator) shares part 9 of his 'meta model' research thread: the idea is a model that predicts other models' training curves, letting you select optimal hyperparameters without endless trial-and-error. An attached StepLaw chart (537M params, 50B tokens) shows TabPFN, NCPL, and power-law predictions landing close to the measured-best learning rate/batch size configuration. He stresses meta models aren't tied to any architecture — LLMs, TabPFN, or KAN all work — and the end goal goes beyond hyperparameter selection to designing entirely new architectures.
More from Research
- Evaluating decider models: randomize everything, warns researcher as IR concepts get rediscovered — antoine_chaffin · 2026-10-31
- As IR research regains popularity, the field is re-deriving old methods step by step — antoine_chaffin · 2026-10-31
- DeepMind's Gemini agent pipeline cracks rules for 11 unresolved BB(6) Turing machines — rvp · 2026-10-31
- Christophe Giraud's Introduction to High-Dimensional Statistics, 2nd edition — FrnkNlsn · 2026-10-31
- Researcher: randomize everything when evaluating decider models—offline signals don't hold on real traffic — din0s_ · 2026-10-31
- Beating the compiler at signed 32-bit division: multiply-shift tricks cut latency — lemire · 2026-10-31