| # Closest existing code-scaling work |
|
|
| ## Scaling Laws for Code: Every Programming Language Matters |
|
|
| ACL Findings paper: https://aclanthology.org/2026.findings-acl.487/ |
|
|
| Earlier arXiv version: https://arxiv.org/abs/2512.13472 (v1, 2025-12-15). |
|
|
| This is directly aligned with the original Code-ATLAS idea: it fits |
| language-specific code scaling laws, measures a bilingual synergy matrix, and fits |
| a proportion-dependent multilingual allocation law. It reports more than 1,000 |
| runs and studies seven languages: Python, Java, JavaScript, TypeScript, C#, Go, and |
| Rust. |
|
|
| Important differences that preserve a substantial Code-ATLAS contribution: |
|
|
| - It covers seven high-resource languages rather than roughly 20--24 languages |
| spanning COBOL, Fortran, Zig, OCaml, Haskell, Julia, and other resource bands. |
| - It does not use ATLAS's directed time-to-same-loss Bilingual Transfer Score or |
| repetition-aware effective-data saturation law. |
| - Its bilingual experiment compares 64B target + 64B auxiliary tokens against a |
| 128B repeated-target baseline; Code-ATLAS will retain ATLAS's 50/50 |
| time-to-same-loss formulation at a fixed reference horizon. |
| - Its reported 900B-token parallel code corpus is Python-pivoted and its public |
| paper does not identify the corpus sources, per-language availability, |
| licenses, revisions, deduplication procedure, or a downloadable corpus and |
| training pipeline. It therefore cannot be used as Code-ATLAS training data. |
| Code-ATLAS makes provenance, licenses, dedup clusters, split IDs, and |
| measured-versus-estimated transfer cells public. |
| - The arXiv v1 limitations explicitly list low-resource and domain-specific |
| languages as unstudied. The ACL final replaced that substantive limitations |
| text with a summary, so claims about the limitation must be version-labelled. |
| - The published equations, experimental contrasts, and aggregate results are |
| useful hypotheses and baselines, but the reported transfer coefficients and |
| mixture should not be imported as data: several equations, table values, and |
| prose claims are internally inconsistent, and the underlying corpus is not |
| auditable. |
|
|
| Accordingly, Code-ATLAS should position itself as a broader, provenance-first, |
| ATLAS-faithful replication and extension, not as the first multilingual code |
| scaling-law paper. |
|
|
