Ziming Liu proposes "meta model" to predict training results before running experiments
ZimingLiu11 · x · 2026-10-31
Ziming Liu (MIT, KAN author) introduces his team's proposed "meta model" — a model that models the learning of models — aiming to predict training results before literally running the experiment.
He clarifies this isn't entirely new: Zhongqi He, Kaiyue Wen and Tengyu Ma earlier proposed NCPL (neural configuration-to-performance scaling law), fine-tuning an LLM to predict training curves from configuration files. Per the shared figure, the new setup takes a training config (data/model size, layers, lr schedule, etc.) as input, embeds numbers with a two-layer MLP, and predicts the residual loss relative to a Chinchilla baseline.
More from Research
- 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
- Beating the compiler at signed 32-bit division: multiply-shift tricks cut latency — lemire · 2026-10-31
- NetHack port challenge: AI transpilers hit 99.5% correctness but tank readability — davidbau · 2026-10-31
- AIGENIE: Open-Source R Package Uses LLMs to Auto-Generate and Validate Psychometric Scales — GolinoHudson · 2026-10-31
- LLM judges are strongly biased toward position, researcher warns to randomize everything — antoine_chaffin · 2026-10-31