---
title: "ISO: An RLVR-Native Optimization Stack — Inherit the Spectrum, Optimize the Frames"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.19331v1"
permalink: "https://feed.astralor.com/items/itm_7c3927dcc50f7cbc834dd3ad"
published: "2026-07-21T17:51:36.000Z"
score: 79
channel: "ai"
tags: ["AdamW","Muon","Qwen3","RLVR","Isospectral Optimization"]
---

# ISO: An RLVR-Native Optimization Stack — Inherit the Spectrum, Optimize the Frames

## AI commentary

> This paper hits the core pain point of current LLM post-training—the blindness of RLVR optimization. ISO's idea is elegant: fix the spectrum, only tune frames. The experiments are convincing, especially compressing training steps to under one-third. For teams working on RLHF/RLVR research and deployment, this is a directly applicable solution. Recommended for researchers and engineers focusing on reasoning improvement, model merging, and efficient fine-tuning.

## At a glance

This paper addresses the missing optimization layer in reinforcement learning with verifiable rewards (RLVR) by identifying spectral inheritance in weight singular structure. The authors propose Isospectral Optimization (ISO) with offline and online instantiations. Offline ISO-Merger combines frame changes of shared-base specialists into a single fixed-spectrum model without post-merge data, achieving the best data-free merging results. Online ISO-Optimizer optimizes frame variables while keeping base spectra fixed, significantly accelerating convergence (e.g., ISO-AdamW reaches same accuracy as AdamW in 100 steps vs. 270 on Qwen3-8B-Base). The work advocates designing post-training around reward-driven adaptation: inherit the spectrum, optimize the frames.

## Key points

- Spectral inheritance: RLVR reuses base model's weight spectra while acquiring new behavior through changes in singular frames.
- ISO framework: offline ISO-Merger enables data-free merging of multiple specialists; online ISO-Optimizer trains frames with fixed spectra for faster convergence.
- On Qwen3-8B-Base, ISO-AdamW matches AdamW's 0.495 accuracy at 100 steps vs 270, and reaches 0.509 at 210 steps.

## Full text

Reinforcement learning with verifiable rewards (RLVR) is rapidly advancing the reasoning capabilities of language models, yet the optimization layer that converts reward feedback into weight-space updates remains poorly understood. Building on our prior analysis (Zhu et al., 2025), we study this missing layer through the singular structure of model weights and identify spectral inheritance: RLVR can reuse the base model's weight spectra while acquiring new behavior through changes in the associated input and output singular frames. We operationalize spectral inheritance as Isospectral Optimization (ISO), an RLVR-native, fixed-spectrum optimization framework with complementary offline and online instantiations. Offline, ISO-Merger combines the frame changes of shared-base specialists into a single fixed-spectrum model, requiring no post-merge data, rollouts, gradient updates, or on-policy distillation (OPD). It recovers complementary specialist capabilities and achieves the strongest aggregate performance among the compared data-free merging methods. Online, ISO-Optimizer applies a chosen base optimizer, including AdamW and Muon, to the frame variables while keeping the base spectra fixed. Across reasoning and coding tasks ranging from 1.5B to 8B parameters, ISO-Optimizer improves accuracy in the reported runs and reaches matched scores with substantially fewer training steps. On Qwen3-8B-Base, AdamW reaches an aggregate accuracy of 0.495 after 270 training steps. ISO-AdamW reaches the same accuracy after only 100 training steps and improves further to 0.509 after 210 training steps. Together, ISO offers a concrete answer to RLVR's missing optimization layer: rather than inheriting pre-training optimization wholesale, design post-training around the structure of reward-driven adaptation: inherit the spectrum, optimize the frames.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_7c3927dcc50f7cbc834dd3ad