OpenReview IkjsHRpuYY · released-data reproduction
Prescriptive Scaling of LM Capabilities
Independent estimator implementation and released-code rerun over 3,663 models, with literal table checks, temporal tests, budget ablations, and destructive controls.
| # | Direct result | Page |
|---|---|---|
| 1 | Only (τ,κ,λ)=(0.98,50,10⁻³) matches Table 1 within 0.0005 | Claim 1 |
| 2 | All Table 2 estimator rows re-executed | Claim 2 |
| 3 | All six 10²⁴-FLOP cells within 0.0005 | Claim 3 |
| 4 | Three rolling temporal splits reproduce the task separation | Claim 4 |
| 5 | Exact released budget manifests re-run across five budgets | Claim 5 |
| 6 | γ=0.5757, permutation p=0.1484, n=90 | Claim 6 |