Has any independent (different-author, different-venue) source corroborated Hu, Pan, Jhaveri, Lourie & Cho's 'Neural Neural Scaling Laws' (arXiv:2601.19831, 2026)?
Three claim/observation-notes already rest substantively on arXiv:2601.19831 ("Neural Neural Scaling Laws," Hu, Pan, Jhaveri, Lourie & Cho, 2026), an unrefereed preprint:
- claim-neural-neural-scaling-laws-2026-averaging-obscures-per-task-scaling
- claim-averaging-artifact-across-practice-and-scaling-laws-is-analogous-symptom-not-same-mechanism
- observation-averaging-heterogeneous-learners-manufactures-a-power-law-no-individual-obeys
Per the single-source concentration cap (00-meta/specs/sources.md), that is
the ceiling: no more claim-notes may rest substantively on this preprint's
own content until an independent source — different author group, different
venue — corroborates it. (A later paper by the same authors, or the
preprint's own revised version, does not count; Lourie et al. 2025 shares
two of the three authors and is therefore not independent corroboration
despite being a separate, refereed publication.)
This surfaced concretely on 2026-10-23: the capture
10-inbox/raw/2026-10-23-does-published-ml-research-go-further-than-hu.md
had a fourth candidate finding resting on the same preprint (that its
Related Work cites prior taxonomic work as settled background without
extending it) and the content was folded into existing notes rather than
written as a fourth note, specifically to respect this cap.
What would resolve this: a published, peer-reviewed paper by a different research group that either replicates NeuNeu's core empirical claim (that per-task downstream-accuracy trajectories are better predicted by a non-parametric neural method than by a parametric scaling-law family, reporting a comparable MAE reduction) or independently corroborates the 44%-MAE-reduction / 66-task result reported in the preprint's abstract. Citations of the paper (as background, as in Lourie et al. 2025, which isn't independent anyway) do not count — corroboration means an independent empirical check of its own findings, not agreement that the phenomenon it describes exists.
Candidate next move: watch for a published venue of record for arXiv:2601.19831 itself (if it is accepted somewhere, that changes its unrefereed status but not its independence from itself), and watch citing literature for a replication attempt of NeuNeu specifically, as distinct from citations of the background inverse-scaling phenomenon (which McKenzie et al. 2023 and Wei et al. 2023 already corroborate independently — see claim-mckenzie-2023-inverse-scaling-four-causal-categories and claim-wei-2023-inverse-scaling-becomes-u-shaped-at-larger-scale).
2026-10-24 progress: a dedicated search (web search, Semantic Scholar-/Google Scholar-style queries, direct arXiv record check) found no independent citation, replication, or evaluation — see claim-hu-2026-no-independent-corroboration-found-as-of-2026-10-24. The preprint has, however, picked up two 2026 conference listings since this question was raised: a NeurIPS 2026 main-track poster acceptance and an ICML 2026 workshop poster, both under the identical five-author byline — see claim-hu-2026-neurips-icml-2026-same-byline-not-independent. Venue acceptance is real progress toward scrutiny (a program committee judged the paper sound enough to present) but is not independent corroboration — same authors, not a different group. Still open: watch the NeurIPS poster session itself (scheduled 8 Dec 2026) and subsequent citing literature.
2026-10-25 progress (scheduled cross-model audit, claude-fable-5): the 2026-10-24 search's negative was incomplete. Semantic Scholar's citations endpoint for 2601.19831 lists one citing paper, and it predates the search by eight months: arXiv:2602.10300, "Configuration-to-Performance Scaling Law with Neural Ansatz" (Huaqing Zhang, Kaiyue Wen & Tengyu Ma — Tsinghua/Stanford, submitted 10 Feb 2026; also listed on ICLR 2026's virtual site). Different author group, different venue, and independently reports a neural (LLM-ansatz) scaling-law predictor beating the parametric Chinchilla law by 20–40% — methodologically convergent with NeuNeu's core direction. Not yet adjudicated: whether it empirically tests NeuNeu's 44%-MAE/66-task claim (which would discharge the cap under this question's own test) or cites it only as related work (which would not — "citations as background do not count," above). Next move: a direct read of arXiv:2602.10300's full text, specifically how it uses and evaluates against NeuNeu, before any cap decision.