ISO: Fixed-Spectrum Optimization Merges RL Experts on 375B MoE in Minutes
zdhnarsil · x · 2026-11-22
Researchers from UT Austin, UIUC, Together AI and others released ISO, an RLVR-native optimization stack built on one observation: during RLVR training, a model's singular-value spectra stay close to the base model's, while capabilities are acquired mostly in the singular frames (U, V) — restoring the base spectrum preserves most gains.
This regularity becomes an inductive bias in two tools:
- ISO-Merger: merges shared-base RL experts with no extra data, rollouts, gradients, or on-policy distillation — just minutes of merging. It has been used for RL expert merging on K2-Horizon at 375B MoE scale.
- ISO-Optimizer: runs AdamW/Muon on the frames, matching RLVR accuracy on Qwen3-4B-Base with 1.5× fewer actor updates.
The takeaway: a simple insight into how RL changes model weights leads to methods that scale practically to very large MoE models.
More from Research
- MIT study: coding agents boost commits 240% but releases only 30% — apps ship into the void — GaryMarcus · 2026-11-22
- Redditor trains Krea2 to skip VAE Decode and output pixels directly using DWD paper — Remote_Net374 · 2026-11-22
- 'AI proves the Riemann hypothesis in 2026' — versus Riemann's 20-watt brain in 1859 — doodlestein · 2026-11-22
- Scientific American's 'End of Math' cover resurfaces after Riemann hypothesis reportedly solved — EricTopol · 2026-11-22
- 'Don't train on model outputs' is unenforceable when the data is all over the open web — anpaure · 2026-11-22
- Researchers debate: after AI cracked the RH, should AI stop hill-climbing math benchmarks? — JJitsev · 2026-11-22