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:

The takeaway: a simple insight into how RL changes model weights leads to methods that scale practically to very large MoE models.

Original post →

More from Research

Research channel →