RL Post-Training Updates Can Be Sparsely Decomposed
menhguin · x · 2026-07-15
This repost covers new research on RL post-training, addressing the core issue: while RL updates are effective, the parameter changes themselves act as a "black box."
The paper proposes treating the genuinely effective "reasoning component" of RL as a compact reconnection matrix within the base model's spectral space. Based on this, they introduce SAR: a retraining-free post-processing method that projects raw RL updates onto this "reasoning core" to better understand, purify, and merge RL-trained models.
Key conclusions from the text include:
- The reasoning core is highly sparse: Less than 1% of the spectral parameters are needed to recover or even improve upon the full RL gains.
- Capable of denoising: Removing irrelevant/noisy directions maintains or enhances model performance.
- Useful for model merging: Beyond analyzing RL updates, it allows post-training modifications to be integrated back into models more cleanly.
More from Research
- Deep Learning Course Adds Mechanistic Interpretability Lecture This Winter — CSProfKGD · 2026-10-26
- INCLUDE-2.0 benchmark evaluates regional knowledge across 91 languages and 86 countries — ABosselut · 2026-10-26
- CoRL paper: flow-matching VLAs can run 1-step Euler inference with no retraining — kaixhin · 2026-10-26
- CoVer at EMNLP 2026: Training Code LLMs to Generate and Self-Verify Simultaneously — algo_diver · 2026-10-26
- Robotic dexterity should be judged on speed, force control and tool use, not copying humans — DJiafei · 2026-10-26
- Turing-complete malware can run in PyTorch's backward pass via a zero-valued loss — retr0jirachi · 2026-10-26