Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion Paradigms
The paper finds the three RLVR fusion methods are close on average, but can split sharply by benchmark.
The authors compare Merge, Mix RL, and multi-teacher on-policy distillation using shared experts and data across model scales. Average performance differs by no more than 1.4 points, while one benchmark shows an 8.6-point gap. They tie those domain swings to cross-domain relationships visible in task-vector geometry. Their guideline is practical: Merge for cheap reuse of existing experts, Mix RL for unified training with tuned domain proportions, and MOPD when preserving teacher gains matters most. ArXiv · AI/CL/LG's note
The authors compare Merge, Mix RL, and multi-teacher on-policy distillation using shared experts and data across model scales. Average performance differs by no more than 1.4 points, while one benchmark shows an 8.6-point gap. They tie those domain swings to cross-domain relationships visible in task-vector geometry. Their guideline is practical: Merge for cheap reuse of existing experts, Mix RL for unified training with tuned domain proportions, and MOPD when preserving teacher gains matters most. ArXiv · AI/CL/LG's note
score 4