ISO says RLVR can keep the weight spectrum and learn 2.7× faster
UTEXAS · hf · 2026-07-22
ISO proposes fixed-spectrum optimization for RLVR and cuts training steps by 2.7×
This paper studies the optimization layer behind reinforcement learning with verifiable rewards (RLVR).
- The authors argue that RLVR can preserve a base model's weight spectra while changing behavior through the associated input and output singular frames, a phenomenon they call spectral inheritance.
- They formalize this as Isospectral Optimization (ISO), with both offline and online versions.
- ISO-Merger combines specialists trained from the same base model into a single fixed-spectrum model, without post-merge data, rollouts, gradient updates, or on-policy distillation.
- The merged model recovers complementary specialist abilities and performs best among the compared data-free merging methods.
- ISO-Optimizer applies optimizers such as AdamW and Muon to the frame variables while keeping the base spectra fixed.
- Across reasoning and coding tasks from 1.5B to 8B parameters, ISO-Optimizer improves accuracy in the reported runs and reaches the same score with far fewer steps.
- On Qwen3-8B-Base, standard AdamW reaches 0.495 aggregate accuracy after 270 steps, while ISO-AdamW hits the same score after 100 steps and climbs to 0.509 after 210 steps.
- The headline claim: RLVR post-training should inherit the spectrum but optimize the frames.
Related event: ISO Framework Optimizes RLVR Training by 2.7x(3 posts)→
More from coding & agent
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- BELL: Open-Source Stack Combines Browserbase, Eve, iMessage and Stripe Link for Personal Agents — jeff_weinstein · 2026-10-27
- Legora runs 2,000 parallel agent evals on Modal Sandboxes for legal AI — blelbach · 2026-10-27
- Two engineers rebuilt a FedEx supplier's billing core in 3.5 months vs an estimated 7-8 — alex_verem · 2026-10-27
- GPT-6.1 Terra scores 48 on AA Coding Agent Index at one-third the cost of GPT-5.6 Terra — ArtificialAnlys · 2026-10-27
- Shift-1 router cuts coding agent spend up to 55% while keeping SWE-bench accuracy — ypatil125 · 2026-10-27