Disagreement-Guided Weak-to-Strong Routing
Model disagreement can guide when to seek a stronger model’s help without ground-truth training labels.
The authors derive a routing rule from comparisons between model-generated responses, allowing multiple responses from the weaker model. Their approximation improves as the stronger model better represents the ground-truth distribution. Across agentic and single-turn tasks, they report a favorable balance between routing cost and response accuracy. They also report that disagreement-based routing can outperform methods trained to predict correctness, even when correctness labels are available.
ArXiv · AI/CL/LG's note
The authors derive a routing rule from comparisons between model-generated responses, allowing multiple responses from the weaker model. Their approximation improves as the stronger model better represents the ground-truth distribution. Across agentic and single-turn tasks, they report a favorable balance between routing cost and response accuracy. They also report that disagreement-based routing can outperform methods trained to predict correctness, even when correctness labels are available.
ArXiv · AI/CL/LG's note
score 4