AI Paper Atlas

Research map / Large language models / Alignment and preferences

Meta-Router: Bridging Gold-standard and Preference-based Evaluations in LLM Routing

Yichi Zhang, Fangzheng Xie, Shu Yang, Chong Wu

ICLR 2026 · Large language models › Alignment and preferences · approach: LLM-as-judge and feedback data

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use R- or DR-learner meta-router to debias preference data instead of pooling GS and PB data directlyrouter trained on GS plus PB data · HealthBenchefficiency gain over random routingImproved
replace random forest with XGBoost as the regressor (ablation)meta-router · HealthBenchefficiency gainMixed
add linear mean-shift debiasing of PB data instead of CATE-based debiasing (ablation)pooled router · HealthBenchefficiency gainNo clear effect
normalize GS outcomes to match PB variance (ablation)meta-router · HealthBenchefficiency gainImproved
PCA dimension 100 instead of 50 (ablation)meta-router · HealthBenchefficiency gainNo clear effect
use Grok 4 Fast as the PB judge instead of GPT-5-mini (ablation)meta-router · HealthBenchefficiency gainNo clear effect
increase GS sample size n (ablation)routers using GS data · HealthBenchefficiency gain: n = 100, 500, 1000Improved
use DR-learner rather than R-learner under small samples (ablation)meta-router · PRBench (676 single-turn questions)efficiency gainImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy