Research map / Label-efficient and robust learning / Recognition and long-tail

Model Cascades with Provable Per-Class Quality

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a model cascade routing method that provides provable per-class quality guarantees, rather than only aggregate quality control.

Problem

Existing model cascade exit rules (confidence thresholds, learned scorers, agreement signals) optimize aggregate quality and do not control quality at the class level, so a cheap model that is systematically unreliable on rare, fine-grained, or minority classes can silently pass errors through.

Approach

Adds per-class quality control to cascade exit decisions so that the cascade's accuracy is guaranteed for each class, with provable guarantees rather than aggregate-only tuning.

Experimental setup

Baselines: confidence thresholding learned scorers agreement-based exit signals

Closest papers by meaning

• Gatekeeper: Improving Model Cascades Through Confidence Tuning (NeurIPS 2025 · similarity 0.87)
• Few-Class Arena: A Benchmark for Efficient Selection of Vision Models and Dataset Difficulty Measurement (ICLR 2025 · similarity 0.85)
• Online Cascade Learning for Efficient Inference over Streams (ICML 2024 · similarity 0.84)
• CIFAR-10-Warehouse: Broad and More Realistic Testbeds in Model Generalization Analysis (ICLR 2024 · similarity 0.84)
• LR0.FM: Low-Res Benchmark and Improving robustness for Zero-Shot Classification in Foundation Models (ICLR 2025 · similarity 0.84)
• Concept-Aware Batch Sampling Improves Language-Image Pretraining (CVPR 2026 · similarity 0.84)