ISO: An RLVR-Native Optimization Stack

• URL: http://arxiv.org/abs/2607.19331v1
• Date: Tue, 21 Jul 2026 17:51:36 GMT
• Title: ISO: An RLVR-Native Optimization Stack
• Abstract summary: We operationalize spectral inheritance as Isospectral Optimization (ISO)<n>Offline, ISO-Merger combines the frame changes of shared-base specialists into a single fixed-spectrum model.<n>Online, ISO-Merger applies a chosen base, including AdamW and Muon, to the frame variables while keeping the base spectra fixed.
• Score: 50.40395312897848
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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.

Related papers

• Neural Network Optimization Reimagined: Decoupled Techniques for Scratch and Fine-Tuning [49.751529745537546]
  We propose DualOpt, a novel approach that decouples optimization techniques specifically tailored for training from scratch.<n>For training from scratch, we introduce real-time layer-wise weight decay, designed to enhance both convergence and generalization.<n>We extend the layer-wise weight decay to dynamically adjust the rollback levels across layers, adapting to the varying demands of different downstream tasks.
  arXiv  Detail & Related papers  (2026-04-21T06:27:18Z)
• A Self-Attentive Meta-Optimizer with Group-Adaptive Learning Rates and Weight Decay [0.42970700836450487]
  We propose a new modulation that integrates a self-attention mechanism to dynamically modulate per-group learning rates and weight decay.<n>To train the attention module, we introduce a meta-learning objective that combines gradient alignment, loss decrease, and gap.<n>A key novel contribution is the extension of homoscedastic uncertainty weighting (HUW) with task-specific priorities that directly scale the regularization terms.
  arXiv  Detail & Related papers  (2026-04-10T02:56:01Z)
• Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models [58.3184497327891]
  Distilled autoregressive (AR) video models enable efficient streaming generation but frequently misalign with human visual preferences.<n>We present Astrolabe, an efficient online reinforcement learning framework tailored for distilled AR models.
  arXiv  Detail & Related papers  (2026-03-17T18:32:18Z)
• NOVAK: Unified adaptive optimizer for deep neural networks [0.0]
  NOVAK is a gradient-based optimization algorithm that integrates adaptive moment estimation, rectified learning-rate scheduling, decoupled weight regularization, multiple variants of Nesterov momentum, and lookahead synchronization into a unified, performance-oriented framework.
  arXiv  Detail & Related papers  (2026-01-11T13:03:57Z)
• Machine Learning Algorithms for Improving Black Box Optimization Solvers [4.994218549120224]
  Black-box optimization (BBO) addresses problems where objectives are accessible only through costly queries.<n>Recent advances use machine learning (ML) and reinforcement learning (RL) to enhance BBO.
  arXiv  Detail & Related papers  (2025-09-29T23:28:42Z)
• Beyond Outliers: A Study of Optimizers Under Quantization [82.75879062804955]
  We study impact of choice on model robustness under quantization.<n>We evaluate how model performance degrades when trained with different baselines.<n>We derive scaling laws for quantization-aware training under different parameters.
  arXiv  Detail & Related papers  (2025-09-27T21:15:22Z)
• ALoRE: Efficient Visual Adaptation via Aggregating Low Rank Experts [71.91042186338163]
  ALoRE is a novel PETL method that reuses the hypercomplex parameterized space constructed by Kronecker product to Aggregate Low Rank Experts.<n>Thanks to the artful design, ALoRE maintains negligible extra parameters and can be effortlessly merged into the frozen backbone.
  arXiv  Detail & Related papers  (2024-12-11T12:31:30Z)
• BiERL: A Meta Evolutionary Reinforcement Learning Framework via Bilevel Optimization [34.24884427152513]
  We propose a general meta ERL framework via bilevel optimization (BiERL) We design an elegant meta-level architecture that embeds the inner-level's evolving experience into an informative population representation. We perform extensive experiments in MuJoCo and Box2D tasks to verify that as a general framework, BiERL outperforms various baselines and consistently improves the learning performance for a diversity of ERL algorithms.
  arXiv  Detail & Related papers  (2023-08-01T09:31:51Z)
• Optimization-Derived Learning with Essential Convergence Analysis of Training and Hyper-training [52.39882976848064]
  We design a Generalized Krasnoselskii-Mann (GKM) scheme based on fixed-point iterations as our fundamental ODL module. Under the GKM scheme, a Bilevel Meta Optimization (BMO) algorithmic framework is constructed to solve the optimal training and hyper-training variables together.
  arXiv  Detail & Related papers  (2022-06-16T01:50:25Z)
• HERO: Hessian-Enhanced Robust Optimization for Unifying and Improving Generalization and Quantization Performance [43.478851400266926]
  We propose HERO, a Hessian-enhanced robust optimization method, to minimize the Hessian eigenvalues through a gradient-based training process. HERO enables up to a 3.8% gain on test accuracy, up to 30% higher accuracy under 80% training label perturbation, and the best post-training quantization accuracy across a wide range of precision.
  arXiv  Detail & Related papers  (2021-11-23T16:32:58Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.