Local Conflict, Targeted Repair: A Perturbation Analysis of Multi-Domain RL
Reinforcement learning post-training improves LLMs in different areas but causes interference between them. The study introduces a local perturbation model showing that this interference happens through sparse, shared computation routes rather than global conflicts. A short 'refresh' on the affected domain effectively contracts this harmful conflict component, proving targeted repair is an effective way to mitigate cross-domain degradation.