ISO claims 2.7x fewer steps for RLVR by reusing the base spectrum
xiuyu_l · x · 2026-07-23
ISO proposes an isospectral optimization stack for RLVR
A new paper argues that reinforcement-learning-based verifiable reasoning (RLVR) can reuse the base model’s spectrum and learn new behavior through the singular frames instead of relearning everything from scratch.
What it introduces
- ISO-Merger: offline composition of RL experts into one model with no data, no rollouts, and no OPD.
- ISO-Optimizer: a drop-in wrapper on AdamW or Muon.
Reported result
- On Qwen3-8B-Base, ISO-Optimizer matches AdamW accuracy with about 2.7× fewer training steps.
- The image also shows spectral inheritance experiments and specialist-score recovery across coding, tool use, memory, and math.
Overall, the paper frames RLVR optimization as a spectral reuse problem rather than a standard fine-tuning problem.
Related event: ISO Framework Optimizes RLVR Training by 2.7x(3 posts)→
More from Research
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27