Zhang, Wen & Ma's CPL method reports 20–40% lower pretraining-loss prediction error than the Chinchilla law, generalizing to runs using up to 10x more compute than its training set

neural-scaling-lawschinchilla-lawtengyu-mascaling-lawssingle-source-cap

Zhang, Wen & Ma's "Configuration-to-Performance Scaling Law with Neural Ansatz" (CPL, arXiv:2602.10300, Tsinghua IIIS / Stanford, 10 Feb 2026) reports a neural (LLM-ansatz) scaling-law predictor — fine-tuning Qwen3-1.7B over 3,000+ pretraining logs — that "accurately predicts how training configurations influence the final pretraining loss, achieving 20-40% lower prediction error than the configuration-agnostic Chinchilla law and generalizing to runs using up to 10x more compute than any run in the training set." The comparison target is Hoffmann et al.'s 2022 Chinchilla law, a parametric scaling-law family that does not take the specific training configuration (beyond parameter/data split) as input; CPL's claimed advantage is replacing that configuration-agnostic parametric form with a learned, configuration-aware one.

This is CPL's own headline empirical result, read directly from a full-text fetch of the paper rather than a search summary. It is a methodological cousin of Hu et al.'s NeuNeu (arXiv:2601.19831) — both teams independently reached "use an LLM as the non-parametric scaling-law ansatz" within months of each other — but CPL cites NeuNeu only in passing and does not test or corroborate it; this note records CPL's own number as a standalone claim about CPL, not as evidence toward NeuNeu's claim either way.

As of this writing CPL is itself an unrefereed arXiv preprint (also listed on ICLR 2026's virtual site, which is not yet a confirmation of acceptance read directly from ICLR's own decision record). Its 20–40% figure has not been independently corroborated by this vault.