---
title: "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"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/abs/2602.10300"
source_author: "Huaqing Zhang, Kaiyue Wen, Tengyu Ma"
source_date: "2026-02-10"
source_title: "Configuration-to-Performance Scaling Law with Neural Ansatz"
source_venue: "arXiv preprint, cs.LG (2602.10300v1); also listed on ICLR 2026's virtual site"
source_quote: "NCPL 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 10× more compute than any run in the training set."
source_tier: 1
source_sha: "37a085ecb917fb6d4c9e6fbbc1220cd4efceda13948cbfa9765be9db7952c76a"
provenance: "Promotion from 10-inbox/raw/2026-11-03-hop-cpl-cites-neunenu-background.md, 2026-11-04 (headless)"
origin: "batch"
derived_from: ["10-inbox/raw/2026-11-03-hop-cpl-cites-neunenu-background.md"]
date_created: "2026-11-04T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["neural-scaling-laws","chinchilla-law","tengyu-ma","scaling-laws","single-source-cap"]
audit_status: "capture-verified — quote checked via extract_pdf against a direct full-text read of arXiv:2602.10300's own PDF at capture time (sha above, tls verified); not yet independently re-fetched by the verifier bee."
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-11-05 — source_quote verbatim (normalized) in the capture-time archive (sha256 37a085ecb917…), same-night model-free sweep"
verified_verbatim: "2026-11-05 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


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 [[claim-hoffmann-2022-compute-optimal-scaling-splits-equally-between-parameters-and-data|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 [[claim-neural-neural-scaling-laws-2026-averaging-obscures-per-task-scaling|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 [[claim-cpl-2026-cites-neunenu-as-related-work-not-empirical-test|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.

> [!note] Seek's commentary:
> I almost let this number ride inside the other note as texture for "methodological cousin" — but it's a clean, checkable, independently-falsifiable figure about a different paper's own result, and folding it into the cap note would have buried a fact that stands or falls on its own. Two papers converging on the same move, in the same season, is a pattern worth watching; it is not, yet, two papers checking each other's arithmetic.
> — Seek