ISO: An RLVR-Native Optimization Stack¶
Candidate-policy objective, not paper-scale RLVR.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | The University of Texas at Austin; first author Hanqing Zhu. |
| First public release | 2026-07-21 (arXiv v1) |
| Author code | zhuhanqing/ISO. |
| Local adapter or method key | iso-rlvr |
| Local reproduction code | src/auto_research/post_training/historical_b08_b09.py |
Paper method and evidence¶
Fix pretrained singular values and optimize only input/output singular frames. The paper also provides a data-free specialist merger.
flowchart LR
A[Pretrained factorization] --> B[Freeze singular values]
B[Freeze singular values] --> C[Optimize singular frames]Original paper figure¶
Figure shows isometric optimization. From the original paper; copyright remains with the authors.
Core equation¶
\[
W=U\Sigma_0V^\top,\quad \Sigma_0\ \text{fixed}
\]
Reported results¶
Qwen3-8B aggregate 0.495 at 270 AdamW steps versus 100 ISO-AdamW steps; ISO reaches 0.509 at 210 steps. No production inference.
Local reproduction¶
auto-research post-train --algorithm iso-rlvr --dataset arithmetic-smoke --steps 120 --seed 42 --no-network
Same arithmetic-smoke candidate policy, 120 updates: 0.1953→0.5391 (+0.3438), not a fair cross-method advantage. Metrics, Batch index. No original checkpoints, full data/compute, multi-seed or benchmark. Raw runs go to runs/post-training/; only metrics committed. Experiment artifacts.