AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Optimized Deferral for Imbalanced Settings

Corinna Cortes, Anqi Mao, Mehryar Mohri, Yutao Zhong

ICML 2026 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology · 0 citations (OpenAlex)

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 margin-based MILD loss instead of TDEFtwo-stage deferral · SVHN synthetic experts (Setup I, error)deferral loss: 0.0468 -> 0.0254Improved
use MILD instead of TDEFtwo-stage deferral · CIFAR-10 synthetic experts (Setup I, error)deferral loss: 0.0520 -> 0.0403Improved
use MILD instead of TDEFtwo-stage deferral · real experts (CIFAR-100 Setup I, error)deferral loss: 0.4368 -> 0.4265Improved
use MILD instead of TDEFtwo-stage deferral · severe expert imbalance (CIFAR-10)deferral loss: 0.0810 -> 0.0125Improved
use MILD instead of TDEFLLM router · MMLU routing, error onlydeferral loss: 0.438 -> 0.425Improved
use MILD instead of TDEFLLM router · MMLU routing, error + costdeferral loss: 0.928 -> 0.813Improved
use MILD instead of CE, CWCE, LDAMLLM router · MMLU routing, error + costdeferral loss: 0.985 / 0.942 / 0.915 -> 0.813Improved
use theory-derived rho instead of uniform rho (ablation)MILD · MMLU routing, error + costdeferral loss: 0.884 -> 0.822Improved
tune rho by grid search around theory values (ablation)MILD · MMLU routing, error + costdeferral loss: 0.822 -> 0.813No clear effect
use MILD under 20% noise in cost matrixMILD vs TDEF · MMLU routing, error + costdeferral loss: TDEF 0.928 -> 1.054; MILD 0.813 -> 0.826Improved
use MILD under extreme 95%/5% imbalanceMILD vs TDEF · MMLU routing extreme imbalancedeferral loss: 0.950 -> 0.885Improved
use MILD instead of TDEFtwo-stage deferral · Tiny ImageNet synthetic experts (Setup III, error + cost)deferral loss: 0.8481 -> 0.8167Improved

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