Retrospective: Hero scaling forecast fixed without using Hero losses, code open-sourced
dlwh · x · 2026-10-27
dlwh notes the fix is retrospective but required no Hero losses—just redoing the Paloma-style fit; code is on GitHub (marin-community/marin-logbooks, issue #8). Key numbers: a shared floor fitted on the original TPU ladder and 27 Delphi endpoints gives 1.4242 and scores 0.00859 RMSE on the separate Hero trajectory; combining Paloma-style compute with a ladder-fitted floor cuts RMSE to 0.00647 with no architecture-specific offset needed. The revised conditional 4K endpoint is 1.497 at the original 390,251-step budget.
More from Research
- Exponential launches RepoInject and BrowserInject benchmarks for prompt injection robustness — maksym_andr · 2026-10-27
- Redwood: CoT controllability evals are under-elicited; better prompts boost scores 2-3x — dfrsrchtwts · 2026-10-27
- Inverse-designed optical system solves MNIST using BeamZ — S_Conradi · 2026-10-27
- DeCert supervises generative reward models' rationales without reference explanations — Guiquan Sun · 2026-10-27
- VLM harnesses act as compositional generalizers, new blog post shows — a1zhang · 2026-10-27
- Article proposes a Human Genome Project for AI interpretability — ninamiolane · 2026-10-27