KAN first author Ziming Liu proposes "meta model" to predict training outcomes before experiments run

Ziming Liu, an MIT PhD student and first author of the KAN paper, introduced the concept of a "meta model" in a long thread on October 31: a model that models the learning process of other models, with the goal of predicting training outcomes before experiments actually run. He candidly described a daily reality of research: even top researchers usually have little idea what will happen before running an AI experiment, and often only find out after it finishes.

Confirmed

• The core positioning is "using models to predict other models' performance, and ultimately designing models in return"; Liu emphasized that this is not tied to any specific architecture—LLMs, TabPFN, and KANs can all serve as meta models, with the positioning itself being what matters.
• On implementation paths, he commented on using LLMs to predict training curves: the NCPL (neural configuration-to-performance scaling law) previously proposed by Han Zhang et al. is a clever idea but requires further conditions (post truncated); his comparative conclusion is that fine-tuned LLMs fall short of the tabular foundation model TabPFN.
• He noted the idea is not entirely new: Shiyi He, Kaiyue Wen, and Yumou Teng already have related prior work (post details truncated).
• The content was released on Day 3 of MetaCircleAI Research Exposition, with a paper link attached.

Why it matters

• If training curves and final performance can be predicted fairly accurately before experiments run, it could dramatically cut compute and time costs, changing the "run it to know it" paradigm of AI experimentation.
• Explicitly decoupling the meta model from specific architectures leaves room for using different tools such as tabular models and KANs for model performance prediction and automated model design.

2026-10-31 ~ 2026-10-31 · 6 related posts

• [source] KAN author Ziming Liu: even top researchers can't predict AI experiment outcomes before running them — ZimingLiu11 · 2026-10-31
• MIT researcher proposes a "meta model" that predicts training results before experiments run — ZimingLiu11 · 2026-10-31
• Ziming Liu proposes "meta model" to predict training results before running experiments — ZimingLiu11 · 2026-10-31
• KAN author argues training-curve prediction fits tabular foundation models like TabPFN, not LLMs — ZimingLiu11 · 2026-10-31
• [source] KAN author's 'meta model' framing: models that predict and design models — ZimingLiu11 · 2026-10-31
• [source] KAN author's 'meta model': predict training curves to pick hyperparams, then design architectures — ZimingLiu11 · 2026-10-31