ISO: An RLVR-Native Optimization Stack
The paper argues RLVR post-training can hold weight spectra fixed and update singular frames instead.
ISO turns that claim into two methods: ISO-Merger for data-free merging of shared-base specialists, and ISO-Optimizer for online training with optimizers such as AdamW or Muon. The authors report that ISO-Merger recovers complementary specialist abilities without rollouts, gradients, post-merge data, or on-policy distillation. In reported reasoning and coding runs from 1.5B to 8B parameters, ISO-Optimizer improves accuracy or reaches matched scores in fewer steps. On Qwen3-8B-Base, ISO-AdamW matches AdamW’s 270-step aggregate accuracy after 100 steps, then rises from 0.495 to 0.509 by 210 steps. HF Daily Papers' note
ISO turns that claim into two methods: ISO-Merger for data-free merging of shared-base specialists, and ISO-Optimizer for online training with optimizers such as AdamW or Muon. The authors report that ISO-Merger recovers complementary specialist abilities without rollouts, gradients, post-merge data, or on-policy distillation. In reported reasoning and coding runs from 1.5B to 8B parameters, ISO-Optimizer improves accuracy or reaches matched scores in fewer steps. On Qwen3-8B-Base, ISO-AdamW matches AdamW’s 270-step aggregate accuracy after 100 steps, then rises from 0.495 to 0.509 by 210 steps. HF Daily Papers' note
score 5