From Sampled Outcomes to Capability Distributions: Rethinking Supervision for LLM Routing
Published in EMNLP 2026, 2026
Recommended citation: Guannan Lai, Haoran Hu, Long Chen, Zhenguo Li, and Han-Jia Ye. (2026). From Sampled Outcomes to Capability Distributions: Rethinking Supervision for LLM Routing. EMNLP 2026. https://arxiv.org/abs/2606.06924
Abstract
This paper rethinks the supervision paradigm for LLM routing: rather than relying on discrete sampled outcomes (win/loss), we propose modeling the underlying capability distributions of different models. This shift enables more robust and fine-grained routing decisions.
Role: Second author.
Status: Accepted at EMNLP 2026