🛰️ Daily AI Frontier
‹ back to 2026-07-22

ISO: An RLVR-Native Optimization Stack

Research LLMs & Foundation Models

Merged summary

TL;DR - ISO is an RLVR-specific optimization framework that freezes a base language model’s weight spectra while updating its singular frames. It improves data-free specialist merging and reportedly reaches matched reasoning and coding accuracy with fewer training steps.

  • Identifies “spectral inheritance”: RLVR can preserve base-model singular values while learning behavior through changes to input/output singular frames.
  • ISO-Merger combines shared-base specialists without data, rollouts, gradient updates, or on-policy distillation.
  • ISO-Optimizer supports base optimizers such as AdamW and Muon while keeping spectra fixed.
  • On Qwen3-8B-Base, ISO-AdamW matched AdamW’s 0.495 aggregate accuracy in 100 versus 270 steps, then reached 0.509 after 210 steps.

Sources (1)

ISO: An RLVR-Native Optimization Stack

arXiv cs.LG Hanqing Zhu, Wenyan Cong, Zhizhou Sha, Sagnik Mukherjee, Xinyuan Song, David González-Martínez, Xiaoxia Wu, Yuandong Tian, Shiwei Liu, David Z. Pan, Zhangyang "Atlas" Wang 2026-07-21 arXiv:2607.19331

TL;DR - ISO is an RLVR-specific optimization framework that freezes a base language model’s weight spectra while updating its singular frames. It improves data-free specialist merging and reportedly reaches matched reasoning and coding accuracy with fewer training steps.

  • Identifies “spectral inheritance”: RLVR can preserve base-model singular values while learning behavior through changes to input/output singular frames.
  • ISO-Merger combines shared-base specialists without data, rollouts, gradient updates, or on-policy distillation.
  • ISO-Optimizer supports base optimizers such as AdamW and Muon while keeping spectra fixed.
  • On Qwen3-8B-Base, ISO-AdamW matched AdamW’s 0.495 aggregate accuracy in 100 versus 270 steps, then reached 0.509 after 210 steps.
item →