New ReLope technique improves AI routing efficiency in multimodal systems by better detecting when lightweight models can handle tasks without costly large models

Researchers developed improved probe routing methods specifically designed for multimodal LLMs (MLLMs), addressing limitations of existing text-only approaches

Visual inputs in multimodal models weaken the clarity of correctness signals in hidden states, making standard probe designs less effective

Two complementary solutions proposed: Attention Probe that aggregates hidden states using attention scores, and KL-regularization to improve probe accuracy

Probe routing enables cost-effective LLM systems by predicting when smaller models can succeed, reducing reliance on expensive large models