Test-Time Coverage: Test-Conditioned Data Curation for Deployment-Aware Learning

2026/07/18 by Nadine Chang, Maying Shen, Shizhe Diao +6
#cs.AI #cs.CV #cs.LG

paper · pdf

Abstract

Deployed AI systems are often trained from broad candidate data pools, necessitating data curation towards the deployment test distribution. However, standard data curation methods score training-side criteria rather than directly optimizing deployment match. We introduce TTCov (Test-Time Coverage), a data-level test-conditioned curation method that uses test-side information before training instead of updating model weights at inference. TTCov decomposes deployment-conditioned curation into coverage and distribution. To represent coverage, it builds a task Atlas, a collection of LLM-based atomic propositions (APs) describing deployment-relevant concepts, seeded from open task knowledge and expanded with unmatched APs extracted from unlabeled deployment samples. To represent distribution, it instantiates the matched deployment APs with their frequencies, yielding a Knowledge Atlas (K-Atlas) that operationalizes the deployment distribution as a curation target. TTCov then selects a budgeted training set whose deployment APs distribution approximates this target. We apply TTCov towards autonomous driving (AD), keeping adaptation off the inference path while selecting data with greater deployment-relevant coverage, closer K-Atlas matching, and stronger downstream end-to-end driving performance than data-curation baselines, including seamless adaptability to novel domains via city-to-city expansion.

Citations

• Position: Stop Reactively Patching Your Model Every Time and Start Proactive Test-Driven AI Development
• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• AdaDeDup: Adaptive Hybrid Data Pruning for Efficient Large-Scale Object Detection Training
• Pseudo-Simulation for Autonomous Driving
• Qwen3 Technical Report
• Centaur: Robust End-to-End Autonomous Driving with Test-Time Training
• KGGen: Extracting Knowledge Graphs from Plain Text with Language Models
• Survey of Large Multimodal Model Datasets, Application Categories and Taxonomy
• SSE: Multimodal Semantic Data Selection and Enrichment for Industrial-scale Data Assimilation
• NAVSIM: Data-Driven Non-Reactive Autonomous Vehicle Simulation and Benchmarking
• Hydra-MDP: End-to-end Multimodal Planning with Multi-target Hydra-Distillation
• FairDeDup: Detecting and Mitigating Vision-Language Fairness Disparities in Semantic Dataset Deduplication
• Scaling Laws for Data Filtering -- Data Curation cannot be Compute Agnostic
• VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning
• The Faiss library
• A Survey on Autonomous Driving Datasets: Statistics, Annotation Quality, and a Future Outlook
• Is Ego Status All You Need for Open-Loop End-to-End Autonomous Driving?
• Parting with Misconceptions about Learning-based Vehicle Motion Planning
• Test-Time Training on Nearest Neighbors for Large Language Models
• Prediction-Oriented Bayesian Active Learning
• VAD: Vectorized Scene Representation for Efficient Autonomous Driving
• SemDeDup: Data-efficient learning at web-scale through semantic deduplication
• GPT-4 Technical Report
• Towards Stable Test-Time Adaptation in Dynamic Wild World
• LAION-5B: An open large-scale dataset for training next generation image-text models
• Beyond neural scaling laws: beating power law scaling via data pruning
• TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Driving
• Efficient Test-Time Model Adaptation without Forgetting
• Continual Test-Time Domain Adaptation
• MEMO: Test Time Robustness via Adaptation and Augmentation
• KITTI-360: A Novel Dataset and Benchmarks for Urban Scene Understanding in 2D and 3D
• Learning Transferable Visual Models From Natural Language Supervision
• A Survey of Deep Active Learning
• Scalability in Perception for Autonomous Driving: Waymo Open Dataset
• Selection via Proxy: Efficient Data Selection for Deep Learning
• LVIS: A Dataset for Large Vocabulary Instance Segmentation
• nuScenes: A multimodal dataset for autonomous driving
• An Empirical Study of Example Forgetting during Deep Neural Network Learning
• End-to-end Driving via Conditional Imitation Learning

Related

• AR-TTA: A Simple Method for Real-World Continual Test-Time Adaptation
• Test-time Adaptation in the Dynamic World with Compound Domain Knowledge Management
• Revisiting Realistic Test-Time Training: Sequential Inference and Adaptation by Anchored Clustering Regularized Self-Training
• Revisiting Realistic Test-Time Training: Sequential Inference and Adaptation by Anchored Clustering
• IT³: Idempotent Test-Time Training