Spectral Rewiring for Exploration, Purification, and Model Merging
The paper argues that useful RL post-training updates can be isolated in a tiny spectral subspace of the base model.
Its SAR method edits trained models after the fact, keeping the “reasoning-effective” part of the update and stripping directions the authors say hurt performance or create interference. In their tests, that core can be as small as about 0.58% of parameters while preserving more than 99% of post-training performance. The authors report gains in math exploration, improvements on six of seven coding benchmarks for an in-house agentic model, and stronger expert merging than prior baselines. Source: HF Daily Papers' note.
Its SAR method edits trained models after the fact, keeping the “reasoning-effective” part of the update and stripping directions the authors say hurt performance or create interference. In their tests, that core can be as small as about 0.58% of parameters while preserving more than 99% of post-training performance. The authors report gains in math exploration, improvements on six of seven coding benchmarks for an in-house agentic model, and stronger expert merging than prior baselines. Source: HF Daily Papers' note.
score 5