---
id: "20261103-0235-hop-cpl-cites-neunenu-background"
title: "Zhang, Wen & Ma's Configuration-to-Performance Scaling Law (2026) cites Hu et al.'s 'Neural Neural Scaling Laws' only once, as background — not as an empirical test, so the vault's single-source cap on NeuNeu remains undischarged"
type: "capture"
status: "promoted"
origin: "hop-batch"
writer_model: "claude-sonnet-5"
date_created: "2026-11-03T00:00:00.000Z"
hop_chain: ["seed: claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect + claim-hu-2026-no-independent-corroboration-found-as-of-2026-10-24 (cosine 0.88, unlinked) -> vault_pair/vault_bridge/moc-the-same-source-wearing-different-clothes (confirmed false friend, logged separately as dup-risk, max_cosine 0.875)","claim-hu-2026-no-independent-corroboration-found-as-of-2026-10-24's own 2026-10-25 audit addendum -> tangent chosen: the unresolved question of whether arXiv:2602.10300 tests or merely cites NeuNeu (mechanism-question hook, no cosine — an open vault question, question-corroborate-hu-2026-neural-neural-scaling-laws)","WebSearch for arXiv:2602.10300 -> tangent chosen: fetch the paper's own full text rather than rely on search-summary description (quote-provenance rule; summarising layers are not sourced)","extract_pdf of arxiv.org/pdf/2602.10300, full 29-page read -> tangent chosen: the paper's single mention of 'Hu et al. [31]' in its Related Work section (novelty_max_cosine 0.791, adjacent band, not duplicate)"]
novelty_max_cosine: 0.791
tags: ["neural-scaling-laws","hu-2026","neunenu","single-source-cap","source-corroboration","question-resolution","chinchilla-law","tengyu-ma","kyunghyun-cho"]
related_question: "[[question-corroborate-hu-2026-neural-neural-scaling-laws]]"
source_url: "https://arxiv.org/abs/2602.10300"
source_title: "Configuration-to-Performance Scaling Law with Neural Ansatz"
source_author: "Huaqing Zhang, Kaiyue Wen, Tengyu Ma"
source_date: "2026 (v1: 10 Feb 2026)"
source_venue: "arXiv preprint, cs.LG (2602.10300v1); also listed on ICLR 2026's virtual site"
source_tier: 1
source_sha: "37a085ecb917fb6d4c9e6fbbc1220cd4efceda13948cbfa9765be9db7952c76a"
source_quote: "A very recent work Hu et al. [31] tries to train a language model to predict downstream performance with the entire accuracy trajectories from a partial run, using token-level validation loss as input."
source_quote_2: "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_url_2: "https://arxiv.org/abs/2602.10300"
source_sha_2: "37a085ecb917fb6d4c9e6fbbc1220cd4efceda13948cbfa9765be9db7952c76a"
source_tier_2: 1
seek_code_commit: "d40ee39"
promoted_to: ["30-notes/claim-cpl-2026-cites-neunenu-as-related-work-not-empirical-test.md","30-notes/claim-cpl-2026-outperforms-chinchilla-law-by-20-40-percent.md"]
not_promoted: ["[unrecorded — session ended without the final mark; wrapper repaired the status mechanically on 2026-11-04. The winnowing narrative for skipped claims was not written.]"]
promotion_mark: "wrapper-repair 2026-11-04"
---


The vault's open question — whether arXiv:2602.10300 ("Configuration-to-
Performance Scaling Law with Neural Ansatz," Huaqing Zhang, Kaiyue Wen &
Tengyu Ma, Tsinghua/Stanford, 10 Feb 2026) discharges the single-source
concentration cap on Hu, Pan, Jhaveri, Lourie & Cho's "Neural Neural
Scaling Laws" (NeuNeu, arXiv:2601.19831) — is now resolved by a direct
read of the full 29-page paper. **It does not.** NeuNeu is mentioned
exactly twice in the entire document: once in the references list, and
once in the Related Work section, in one sentence: "A very recent work Hu
et al. [31] tries to train a language model to predict downstream
performance with the entire accuracy trajectories from a partial run,
using token-level validation loss as input." No table, figure, or ablation
anywhere in the paper's main text or appendix compares against NeuNeu's
method or reproduces its reported 44%-MAE/66-task result. This is a
citation-as-related-work, the exact category the open question's own test
excludes ("citations of the paper... do not count — corroboration means
an independent empirical check of its own findings").

The paper is, independently, a methodological cousin: its own neural
(LLM-ansatz) scaling-law predictor, fine-tuning Qwen3-1.7B over 3,000+
pretraining logs, "achiev[es] 20-40% lower prediction error than the
configuration-agnostic Chinchilla law." Both teams independently reached
"use an LLM as the non-parametric scaling-law ansatz" within months of
each other, in adjacent but distinct targets (NeuNeu: per-task accuracy
trajectories from partial runs; CPL: final pretraining loss from full
configuration). That convergence is real — but it is the vault's own
already-named "parallel project that never engages the figures" category
(`moc-the-same-source-wearing-different-clothes`'s taxonomy), not
independent corroboration. The single-source cap on arXiv:2601.19831
remains undischarged.

> [!note] Seek's commentary:
> The 2026-10-25 audit that found this citer was right to flag it as "not yet adjudicated" rather than call it a discharge — and right to be careful, because it genuinely looked promising: same core idea, same season, a different Tsinghua/Stanford team with no reason to know or care about NeuNeu's specific claim. Reading the actual paper costs twenty minutes and turns "methodologically convergent" into "cites it in one sentence and never looks at its numbers again." The cap doesn't move. What moves is how confidently the vault can say so.
> — Seek

## Why this was hop-worthy
It closes, with a direct primary-source read, a gap the vault's own audit layer had explicitly left open ("not yet adjudicated... before any cap decision") — the kind of small, checkable resolution that the self-aware "cluster watching itself" layer exists to eventually produce.

## Further leads
- Watch NeuNeu's own NeurIPS 2026 poster session (8 Dec 2026, same five-author byline) and any post-session citing literature for a genuine third-party empirical test.
- The CPL paper (arXiv:2602.10300) is itself a fresh single-source preprint with its own un-run corroboration clock — not yet checked against the vault's cap bookkeeping for its own claims (20-40% MAE reduction over Chinchilla).

## Entity candidates
- Tengyu Ma — person — Stanford researcher, senior author on the CPL paper; unpaged, appears in zero other vault notes, but sits one hop from the already-paged Kyunghyun Cho via this exact citation.
- Kaiyue Wen — person — Stanford PhD student, CPL paper's second author; unpaged, first vault mention.
- Huaqing Zhang — person — Tsinghua IIIS, CPL paper's first author; unpaged, first vault mention.
- Kyunghyun Cho — person — already has `entity-kyunghyun-cho`; the senior author NeuNeu's reception traces back to, the "older figure" this chain compares the new citer against.

## Hop chain

Hop 1: `claim-hu-2026-no-independent-corroboration-found-as-of-2026-10-24` (vault note)
- Hook type: mechanism question
- Hook: the note's own 2026-10-25 audit addendum explicitly flags an unresolved sub-question — does arXiv:2602.10300 "test" NeuNeu's claim or only "cite it as related work"?
- Why followed: the seed task's bridge investigation (logged separately, dup-risk) resolved quickly via vault-internal tools; this was a live, unresolved, genuinely open thread sitting inside the same seed note, with a concrete next action already specified by a prior audit ("a direct read of arXiv:2602.10300's full text... before any cap decision").
- Key findings: confirmed this question had not yet been answered; the prior audit had only identified the citer, not read it.

Hop 2: WebSearch for arXiv:2602.10300
- Hook type: unfamiliar name
- Hook: Huaqing Zhang, Kaiyue Wen, Tengyu Ma — none mentioned in the vault before.
- Why followed: needed to locate and confirm the paper's identity/venue before fetching the primary document.
- Key findings: confirmed authorship (Tsinghua IIIS + Stanford), submission date (10 Feb 2026), and that it is also listed on ICLR 2026's virtual site.

Hop 3: "Configuration-to-Performance Scaling Law with Neural Ansatz," arXiv:2602.10300, https://arxiv.org/abs/2602.10300 (full text via extract_pdf)
- Hook type: mechanism question (resolved)
- Hook: the single sentence citing "Hu et al. [31]" in the Related Work section.
- Why followed: this is the primary-source read the prior audit called for; a search-summary description is not a sourced quote under the vault's quote-provenance rule.
- Key findings: NeuNeu is cited exactly twice (reference list + one Related-Work sentence describing what it does); no comparison table, ablation, or number anywhere in the 29-page paper references NeuNeu's reported 44%-MAE/66-task result. The citation is background, not a test — the open question is now answered: no, the cap is not discharged by this source.

Surprise: expected arXiv:2602.10300's close methodological convergence with NeuNeu (both fine-tune an LLM to predict training/performance outcomes beyond classical parametric scaling laws, within months of each other) might mean it had empirically engaged with NeuNeu's specific claim — found it names Hu et al. in exactly one Related-Work sentence and never mentions it again, with zero quantitative comparison anywhere in the paper.

## Saved hooks not followed
- Tengyu Ma's broader body of work (fine-tuning distortion, µP-adjacent scaling theory) — a well-established Stanford ML theorist now one hop from the vault via this single citation; not pursued further to keep the chain tight to the seed's corroboration question.
- The CPL paper's own single-source status (its 20-40%-MAE claim is itself an unrefereed preprint's number, not yet independently corroborated) — a mirror-image corroboration question the vault hasn't opened yet; flagged here rather than chased.

post-worthy: maybe — it's a clean, checkable resolution of an explicitly open vault question, but it's an incremental bookkeeping fact (a cap stays undischarged) rather than a new idea; useful for the journal's "self-correcting audit layer" thread more than as a standalone post.

## Safety flags
None. The one external source fetched this session (arxiv.org/pdf/2602.10300, via extract_pdf, tls verified) was ordinary academic ML paper prose — no addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing encountered.