Noumena Research https://www.noumena.com/research Research publications from Noumena http://www.rssboard.org/rss-specification python-feedgen en Sun, 05 Apr 2026 03:18:03 +0000 Let the Speedrun Search Itself https://www.noumena.com/research/0011-let-the-speedrun-search-itself/ Eval-gated config-only autoresearch on the canonical super fp8 lane https://www.noumena.com/research/0011-let-the-speedrun-search-itself/ Sat, 14 Mar 2026 00:00:00 +0000 Reproducing Canon, mHC, and Engram https://www.noumena.com/research/0010-physics-architecture-repros/ A research narrative: wrong starts, PhysicsLM4 alignment, and one real polysemy failure https://www.noumena.com/research/0010-physics-architecture-repros/ Sat, 14 Mar 2026 00:00:00 +0000 RDEP https://www.noumena.com/research/0009-rdep/ keeping sparse expert compute hot across a whole NVLink fabric https://www.noumena.com/research/0009-rdep/ Sat, 14 Mar 2026 00:00:00 +0000 Do MoE Experts Need Different Learning Rates? https://www.noumena.com/research/0008-expert-learning-rate/ Why Moonlet's old 15x expert-LR rule overshoots in bf16 AdamW https://www.noumena.com/research/0008-expert-learning-rate/ Sat, 14 Mar 2026 00:00:00 +0000 The Atlas Hypothesis https://www.noumena.com/research/0007-the-geometry-hypothesis/ why output-only dashboards cannot name what pretraining built, and what a real receipt would have to measure https://www.noumena.com/research/0007-the-geometry-hypothesis/ Sat, 14 Mar 2026 00:00:00 +0000 Super-4096 https://www.noumena.com/research/0006-super-4096/ Loss keeps improving while routing collapses under extreme sparsity https://www.noumena.com/research/0006-super-4096/ Sat, 14 Mar 2026 00:00:00 +0000 NVFP4 Dynamics https://www.noumena.com/research/0005-nvfp4-dynamics/ Why our NVFP4 recipe lagged BF16, and what actually closed almost all of the gap https://www.noumena.com/research/0005-nvfp4-dynamics/ Sat, 14 Mar 2026 00:00:00 +0000 What Are We Holding Fixed? https://www.noumena.com/research/0004-420-moe-edition/ Dense-vs-MoE comparison depends on the fairness contract; a failed `#420` transfer exposed the real problem https://www.noumena.com/research/0004-420-moe-edition/ Sat, 14 Mar 2026 00:00:00 +0000 The Speedrun Loop https://www.noumena.com/research/0003-speedrun-calibration/ A small-model speedrun is our fastest honest instrument for architecture research https://www.noumena.com/research/0003-speedrun-calibration/ Sat, 14 Mar 2026 00:00:00 +0000 Make It Measurable https://www.noumena.com/research/0002-make-it-measurable/ What to track when loss isn't enough https://www.noumena.com/research/0002-make-it-measurable/ Sat, 14 Mar 2026 00:00:00 +0000 What We Built https://www.noumena.com/research/0001-what-we-built/ A production-grade MoE training system, because reproducibility is the experiment https://www.noumena.com/research/0001-what-we-built/ Sat, 14 Mar 2026 00:00:00 +0000 Why Training MoEs is So Hard https://www.noumena.com/research/0000-why-training-moes-is-hard/ Three failure modes that make frontier MoE training qualitatively different https://www.noumena.com/research/0000-why-training-moes-is-hard/ Sat, 14 Mar 2026 00:00:00 +0000